Multitask Principal-Agent Analyses: Incentive Contracts, Asset Ownership, and Job Design

Author(s): Bengt Holmstrom and Paul Milgrom

Source: Journal of Law, Economics, & Organization, Vol. 7, Special Issue: [Papers from the
Conference on the New Science of Organization, January 1991] (1991), pp. 24-52
Published by: Oxford University Press
Stable URL: http://www.jstor.org/stable/764957
Accessed: 13/08/2009 11:15

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=oup.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with the
scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that
promotes the discovery and use of these resources. For more information about JSTOR, please contact [email protected].

Oxford University Press is collaborating with JSTOR to digitize, preserve and extend access to Journal of Law,
Economics, & Organization.

http://www.jstor.org
 24 JLEO.V7 Sp

MultitaskPrincipal-Agent Analyses:Incentive
Contracts, Asset Ownership, and Job Design
Bengt Holmstrom
YaleUniversity

Paul Milgrom
StanfordUniversity

1. Introduction
In the standardeconomic treatmentof the principal-agentproblem, compen-
sation systems serve the dual function of allocatingrisks and rewardingpro-
ductive work. A tension between these two functionsarises when the agent is
risk averse, for providingthe agent with effective work incentives often forces
him to bear unwanted risk. Existing formal models that have analyzed this
tension, however, have producedonly limited results. It remainsa puzzle for
this theory that employmentcontractsso often specify fixed wages and more
generally that incentives within finns appearto be so muted, especially com-
paredto those of the market.Also, the models have remainedtoo intractable
to effectively address broaderorganizationalissues such as asset ownership,
job design, and allocation of authority.
In this article, we will analyze a principal-agentmodel that (i) can account
for paying fixed wages even when good, objective outputmeasuresare avail-
able and agents are highly responsive to incentive pay; (ii) can make recom-
mendationsand predictionsaboutownershippatternseven when contractscan
take full account of all observablevariablesand courtenforcementis perfect;
(iii) can explain why employmentis sometimes superiorto independentcon-
We are gratefulto the National Science Foundationfor financialsupportand to Gary Becker,
James Brickley, MurrayBrown, Joel Demski, Joseph Farrell, Oliver Hart, David Kreps, Kevin
Murphy,Eric Rasmussen, Steve Ross, Steve Stem, and especially AvnerGreif, JaneHannaway,
and Hal Varianfor their many insightful comments, examples, and suggestions. We also wish to
thank Froystein Gjesdal for pointing out two errorsin an earlier draft.
1. Some of the predictiveweaknesses of standardagency models are discussed in the surveys
by MacDonald, Hart and Holmstrom, and Baker, Jensen, and Murphy.
? 1991 by Oxford University Press. All rights reserved. ISSN 8756-6222
 Multitask Principal-Agent Analyses 25

tractingeven when there are no productiveadvantagesto specific physical or

human capital and no financial market imperfections to limit the agent's
borrowings;(iv) can explain bureaucraticconstraints;and (v) can shed light
on how tasks get allocated to differentjobs.
The distinguishingmarkof our model is thatthe principaleitherhas several
differenttasks for the agent or agents to perform,or the agent's single task has
several dimensions to it. Some of the issues raised by this modeling are well
illustratedby the currentcontroversyover the use of incentivepay for teachers
based on their students' test scores.2 Proponentsof the system, guided by a
conception very like the standardone-dimensionalincentivemodel, arguethat
these incentives will lead teachers to work harder at teaching and to take
greaterinterestin theirstudents'success. Opponentscounterthatthe principal
effect of the proposed reform would be that teachers would sacrifice such
activities as promotingcuriosity and creative thinking and refining students'
oral and writtencommunicationskills in orderto teach the narrowlydefined
basic skills that are tested on standardizedexams. It would be better, these
critics argue, to pay a fixed wage withoutany incentivescheme than to base
teachers' compensationonly on the limiteddimensionsof studentachievement
that can be effectively measured.3
Multidimensionaltasks are ubiquitousin the world of business. As simple
examples, productionworkers may be responsible for producinga high vol-
ume of good quality output, or they may be requiredboth to produceoutput
and to care for the machines they use. In the firstcase, if volume of outputis
easy to measurebut the quality is not, then a system of piece rates for output
may lead agents to increasethe volume of outputat the expense of quality.Or,
if quality can be assured by a system of monitoringor by a robust product
design, then piece rates may lead agents to abuse sharedequipmentor to take
inadequatecare of it. In general, when there are multipletasks, incentive pay
serves not only to allocate risks and to motivate hard work, it also serves to
direct the allocation of the agents' attentionamong their various duties. This
representsthe first fundamentaldifferencebetween the multidimensionalthe-
ory and the more common one-dimensionalprincipal-agent models.
There is a second fundamental difference as well, and it, too, can be
illustratedby reference to the problem of teaching basic skills: If the task of
teaching basic skills could be separatedfrom that of teaching higher-level
thinking, then these tasks could be carriedout by differentteachersat different
times during the day. Similarly, in the example of the production worker,
when the care and maintenanceof a productiveasset can be separatedfrom
the use of that asset in producingoutput, the problemthat a piece rate system
would lead to inadequatecare can be mitigatedor even eliminated.In general,
in multitaskprincipal-agent problems,job design is an importantinstrument

2. See Hannaway for a discussion of these issues.

3. As a concrete illustrationof the distortionsthat testing can cause, in 1989 a ninth-grade
teacher in Greenville, South Carolina was caught having passed answers to questions on the
statewide tests of basic skills to students in her geography classes in order to improve her
performancerating (Wall Street Journal, November 2, 1989).
26 The Journalof Law.Economics,&Organization,
V7 Sp

for the control of incentives. In the standardmodel, when each agent can
engage in only one task, the grouping of tasks into jobs is not a relevant
issue.4
Our formal modeling of these issues utilizes our linear principal-agent
model (Holmstromand Milgrom, 1987), mainly specialized to the case where
the agent's costs depend only on the total effortor attentionthe agent devotes
to all of his tasks. This modeling assures that an increase in an agent's
compensationin any one task will cause some reallocationof attentionaway
from other tasks. First, we show that an optimal incentive contractcan be to
pay a fixed wage independentof measuredperformance,just as the opponents
of incentives based on educationaltesting have argued. More generally, the
desirability of providing incentives for any one activity decreases with the
difficultyof measuringperformancein any other activities that make compet-
ing demands on the agent's time and attention. This result may explain a
substantial part of the puzzle of why incentive clauses are so much less
common than one-dimensionaltheories would predict.
Second, we specialize our model to the case where the unmeasurableaspect
of performanceis how the value of a productiveasset changes over time. The
difficulties of valuing assets are well recognized, and the vast majority of
accountingsystems value assets using fixed depreciationschedules based on
historical costs, deviating from this procedureonly in exceptional circum-
stances. Under these conditions, when the principalowns the returnsfrom the
asset, the optimal incentive contract will provide only muted incentives for
the agent to produceoutput, in orderto mitigateany abuse of the asset or any
substitutionof effort away from asset maintenance.However, when the agent
owns the asset returns, the optimal incentive contract will provide more
intensive incentives to engage in production,in orderto alleviate the reverse
problem that the agent may use the asset too cautiously or devote too much
attention to its care and improvement. This analysis supportsWilliamson's
observation that "high-powered" incentives are more common in market
arrangementsthan within firms, without relying on any assumptionsabout
specific investments. Moreover, it provides a rudimentarytheory of owner-
ship, accordingto which the conditions thatfavorthe agent owning the assets
are (i) thatthe agent is not too risk averse, (ii) thatthe varianceof asset returns
is low, and (iii) thatthe varianceof measurementerrorin other aspects of the
agent's performance is low. Thus, it emphasizes measurementcost as an
importantdeterminantof integrationin contrastto the leading approaches,
which stress asset specificity.5
4. Riordan and Sappington also analyze an incentive model in which job assignment is
central, but for a very different reason. They ask when the principal should do one of two
sequentialproductionstages herself in orderto reduce the agent's informationadvantage.In our
model, job assignments do not affect the principal'sinformation.
5. Alchian and Demsetz argued that monitoring difficulties account for the formation of
firms, but their theorywas subsequentlyrejectedin favorof the view that asset specificity and ex
post bargainingproblemsdrive integration(Grossmanand Hart, Williamson). We are reintroduc-
ing measurementcost as a key factor, but in a way thatdiffersfrom the originalAlchian-Demsetz
theory. In particular, we do not argue that owners can better monitor the work force. Our
approachis more closely related to Barzel's work.
 Multilask Prncipal-Agent Analyses 27

Our predictionfits well with the empiricalevidence reportedby Anderson

and Andersonand Schmittlein.They found thatfirmsin the electronicsindus-
try tend to employ their own sales forces ratherthan independentmanufactur-
er's representativeswhen some aspects of the representative'sperformance
are hard to measure. Our result can also help explain why franchisees face
steep performanceincentives, while managersof identical company-owned
stores receive no incentive pay at all (Krueger,Brickley and Dark),and why a
free-lance writermight be paid for articlesby the word, while a staff reporter
for the same publicationreceives a fixed wage.
Third, we explore how a firmmight optimallyset policies limiting personal
business activities on companytime. Again, it is notjust the characteristicsof
the "outside activities" themselves that determine whether these activities
should be permitted. We find that outside activities should be most severely
restrictedwhen performancein the tasks that benefit the firm-the "inside
activities"-are hardto measureand reward.Thus, a salespersonwhose pay
is mostly in the form of commissions will optimallybe permittedto engage in
more personalactivities duringbusiness hoursthan a bureaucratwho is paid a
fixed wage, because the commissions direct the salesperson toward inside
activities in a way that cannot be duplicatedfor the bureaucrat.Our theory
also predicts that home office work should be accompaniedby a stronger
reliance on performance-basedpay incentives, a predictionthat seems to fit
casual observation.
Our analysis of restrictionson outside activities underscoresthe fact that
incentives for a task can be providedin two ways: eitherthe task itself can be
rewardedor the marginal opportunitycost for the task can be lowered by
removing or reducing the incentives on competing tasks. Constraintsare
substitutesfor performanceincentives and areextensively used when it is hard
to assess the performanceof the agent. We believe this opens a new avenuefor
understandinglarge-scale organization.It also offers an alternativeinterpreta-
tion of the Anderson-Schmittlein evidence. It is inefficientto let a salesper-
son, whose performanceis poorly measured,divert his time into commission
selling of competing products.If the employer has an advantagein restricting
the employees' other activities, as both Simon and Coase have argued, then
problems with measuring sales performancewill lead to employing an in-
house sales force.
Finally, we obtain a series of results in the theory of job design, using a
model in which the employer can divide responsibilityfor many small tasks
between two agents and can determinehow performancein each task will be
compensated. The resulting optimization problem is a fundamentallynon-
convex one, and we have had to make some extra assumptionsto keep the
analysis tractable. Nevertheless, the results we obtain seem intriguing and
suggestive. First, we find that each task should be made the responsibilityof
just one agent. To our knowledge, this is the first formal derivation in the
incentive literatureof the principleof unity of responsibility,which underlies
the theoryof hierarchy.Second, we findthattasks should be groupedintojobs
in such a way that the tasks in which performanceis most easily measuredare
assigned to one worker and the remaining tasks are assigned to the other
28 The Journalof Law,Economics,& Organization,
V7 Sp

worker. This conclusion squares nicely with the intuition that it is the
differencesbetweenthe measurabilityof quantityandqualityin production,or
of the so-called "basicskills" and "higher-orderthinkingskills" in education,
that make those incentive problems difficult. The theory indicates that even
when the agents have identical ex ante characteristics,the principal should
still design theirjobs to have measurementcharacteristicsthatdifferas widely
as possible. The principalshould then provide more intensive incentives and
require more work effort from the jobholder whose performanceis more
easily measured.
Our results are variations on the general theme of second best, which
stresses that when prices cannotallocate inputsefficiently,then optimalincen-
tives will typically be provided by subsidizing or taxing all inputs. For in-
stance, Greenwaldand Stiglitz, in a vivid metaphor,point out the value of a
governmentsubsidy for home fire extinguishers,since homeownerswith fire
insurancehave too little incentive to invest in all forms of fire preventionand
to fight fires once they have started. This mechanismhas been most exten-
sively analyzed in the theory of optimal taxation and in welfare theory.
However, the study of interdependenciesamong incentives and the use of
instrumentsother than compensation to alleviate incentive problems have
enteredagency analyses more recently.Lazeararguesthat where cooperation
among workers is important,we should expect to see less wage differentia-
tion, that is, "lower-powered"incentives. Holmstromand Ricarti Costa have
observed how a firm's capital budgetingpolicy, including the hurdlerate and
the way the firmassesses idiosyncraticrisks, can affectthe willingness of risk-
averse managersto propose risky investmentprojects. Milgrom and Milgrom
andRobertshave studiedhow organizationaldecisionprocessesaffectthe allo-
cation of effort between politicking and directly productivework. Farrelland
Shapiroshow thata priceclause may be worse thanno contractat all, becauseit
reduces incentives to supply quality;this is similarto our resultthat it may be
optimal to provide no quantity incentives when quality is poorly measured.
Some articles containing related ideas have been developed contempo-
raneously.Itoh (1991), in an analysis complementaryto ours, studies condi-
tions under which an employer might induce workersto work separatelyon
their tasks, and those in which it is best for them to spend some efforthelping
one another.LaffontandTirole show thatconcernsfor qualityhelp explainthe
use of cost-plus contractingin procurement.Baker investigates a model in
which observable proxies of marginal product are imperfect in a way that
causes the agent to misallocateeffort across contingenciesand thereforeleads
to incentives that are not as powerful as standardtheory would suggest.
Minahanreportsa result on task separationthat suggests a job design similar
to ours but based on a differentargument,as we will later explain.
The remainderof this article is organized as follows. In Section 2, we
recapitulateourbasic principal-agenttheory,upon which the entireanalysis is
based. In Section 3, we specialize the analysis to the case where the agent's
costs depend only on the total attentionsuppliedand prove the various propo-
sitions aboutthe optimalityof fixed wages, the factorsdeterminingthe assign-
ment of ownership, and the optimal limits on outside business activities. In
 Multitask Principal-Agent Analyses 29

Section 4, we consider restrictionson privatetasks. In Section 5, we offer a

summaryand suggest directions in which this line of researchcan be taken.

2. The Linear Principal-Agent Model

2.1 Descriptionof the Model
Considera principal-agent relationshipin which the agent makes a one-time
choice of a vector of efforts t = (t,...,tn) at personalcost C(t). The efforts t
lead to expected gross benefits of B(t), which accruedirectly to the principal.
We assume that the function C is strictly convex and that the function B is
strictly concave. The agent's efforts also generate a vector of information
signals

x = i(t) + e,

where we assume that/: 9nS -> Tk is concave and e is normallydistributed

with mean vector zero and covariancematrix2. If the compensationcontract
specifies a wage of w(x), then the agent's expected utility is assumed to take
the form

u(CE) = E{u[w(u(t) + e) - C(t)]},

where u(w) = -e-r and CE denotes the agent's "certaintyequivalent"

money payoff. The coefficient r measuresthe agent's risk aversion. The prin-
cipal is risk neutral.
If the compensationrule were linear of the form w(x) = aTx + B, then one
could utilize the exponentialform to deduce that the agent's certaintyequiv-
alent is

CE = aTu(t) + i - C(t) - lraTZa.

That is, the agent's certaintyequivalentconsists of the expected wage minus

the private cost of action and minus a risk premium. The term aTZa is the
variance of the agent's income under this linear compensationscheme.
The principal'sexpected profit is B(t) - E{wl(t) + e]} which, under the
linear compensation scheme, is B(t) - atu(t) - /. Consequently,the total
certaintyequivalent of the principaland the agent (theirjoint surplus)under
the linear compensation plan is B(t) - C(t) - araTza. Notice that this
expression is independentof the interceptterm,f; this interceptserves only to
allocate the total certaintyequivalentbetween the two parties.This last obser-
vation simplifies the principal-agent problem drastically. It implies that,
given any technological and incentive constraintson the set of feasible (a,t)
pairs, the utility possibility frontier,expressedin certaintyequivalentterms, is
a line in [R2with slope -1. Hence, the incentive-efficientlinear contractsare
precisely those that maximize the total certainty equivalent subject to the
constraints. If (t,a,f) is such a contract, then (t,a) must be a solution to

Maximize B(t) - C(t) - ?.raT2a, (1)

rt,c
 30 The Journalof Law,Economics.&Organization.
V7 Sp

subject to

t maximizes aTu(t') - C(t'). (2)

If the agent's certaintyequivalentis CE, then it follows that the interceptis /B

= CE - aTu(t) + C(t) + -.raxTa. This intercept is equal to the agent's
certaintyequivalent income, minus the expected compensationfrom the in-
centive term, plus compensation for the cost that the agent incurs, plus a
compensationfor risk.
A central featureof our model is the general way in which we may allow
observablesto enter. We can study situationsin which differentactivities can
be measuredwith varying degrees of precision, including the importantspe-
cial case in which certainactivities cannotbe measuredat all.6 We can study
cases in which performancemeasures can be influenced by activities other
than those the principal desires the agent to undertake-for instance, the
manipulationof accountingfigures. We can study cases in which the number
of observablesis much smaller than the numberof activities in t, forcing the
contractto be based on aggregate informationabout the agent's activities. A
special case of this, discussed in Holmstrom and Milgrom (1987), occurs
when the agent acts on private information(to avoid adverse selection, one
assumes that the informationis observed after contracting).We can bring in
contingent actions explicitly by specializing the model as follows. Let AT =
(A;,...,A,,) be a vector of probabilities of m possible states. Let ti be the
agent's contingent action in state i and let B,(t), Ci(ti),,ui(ti),and ei represent
state-contingentprofits, costs, signal functions, and memory errors, respec-
tively. The analysis of thatcontingent-actionmodel is equivalentto the analy-
sis of our model with the specifications:

B(t) = AiBi(ti), C(t)) = AiCi(ti),

H(t) = ,ii(ti), e = ~iei.

Anotherimportantfeatureof the model is thatB need not be partof x (i.e.,

the returnsto the principal may not be observed). This puts B and C in a
symmetricrole. Indeed, ifB = -C, the principalandthe agent sharethe same
objective and firstbest can be achieved in (1) and (2) by setting a = 0. On the
other hand, if B is different from -C, (1) and (2) may lead to a nontrivial
agency problem even without the agent being risk averse. This occurs when
the standardsolution of making the agent a residual claimant is rendered
infeasible because B is insufficientlywell observed-a point made in Baker
using a model with state-contingentactions of the type described above.
Thus, risk aversion is not essential for the analysis to follow. The cost of
measurementerror,as expressedin (1), could alternativelyarise out of a risk-
neutral formulation.7

6. Note that if an activity can be measuredwithout error, then a linear scheme allows the
principalto set this activityat any desiredlevel costlessly, assumingthatthecost functionis convex.
7. It is of interestto note that instead of a measurementerrorthe incentive problemcould be
 Multitask
Principal-AgentAnalyses 31

2.2 Optimalityof LinearPerformanceIncentives

The model describedabove involves two seemingly ad hoc assumptions.The
more obvious one is that the contract that the parties sign specifies a wage
payment that is a linear function of measuredperformance.The second as-
sumptionis more conventionaland thereforeless likely to be noticed, but it is
no less troubling. It is the assumptionthat the agent is requiredto make a
single, once-and-for-allchoice of how he will allocate his efforts during the
relationship without regard to the arrival of performanceinformationover
time. A remarkablefact, which we established in Holmstromand Milgrom
(1987), is that these tvo simplifyingassumptionsare exactly offsettingin this
model. That is, the solution to the program(1) and (2) coincides with the
solution to a principal-agent problem in which (i) the agent chooses efforts
continuously over the time interval [0,1] to control the drift vector of a
stationary stochastic process (Brownian motion) {X(z); 0 C r 1}, and
(ii) the agent can observe his accumulatedperformancebefore acting. We
show that in this continuoustime model an efficientcontractspecifies that the
agent will choose t(r) to be constant over time, regardlessof the history at
time r, and thatthe agent's wage will be of the form w = aTx + ,/, thatis, it is
a linearfunctionof the final outcomex alone, withoutregardto any intermedi-
ate outcomes. The constant t and the slope vector a are the solution to
problem (1) and (2).
In view of its underlying assumptions, the model seems especially well
suited for representingcompensationpaid over a shortperiod, like a month, a
quarter,or perhapsa year, in environmentswhere profits are the cumulative
resultof persistenteffortsover time. As such, the model seems most appropri-
ate for analyzing the use of piece rates or commission systems; however,
because the model is so tractable,we shall not avoid the temptationto stretch
its use somewhat furtherin this article.

2.3 Simple InteractionsAmong Tasks

To explore some of the propertiesof our model, let us now work with the
special case in which u(t) = t.8 Then, when t is strictly positive in all
components (t > 0), the incentive constraint(2) becomes

a, = Ci(t) for all i, (3)

where subscriptson C denote partialderivatives. Differentiating(3), we may
write
aa at =
t = [C,] and aa [C]- (4)

driven by a nonstochasticmeasurementbias. Suppose the agent can manipulatethe performance

measure. If this activity wastes resources,then incentives will optimallybe set to balancethis loss
against genuine work incentives. One can specify the cost of manipulationso that optimal
incentives come out exactly the same as in the stochastic model we are studying.
8. This is really not a special case, since we can always reformulatethe model by redefining
the agent's choice variables so that t(t) = t.
32 The Journalof Law,Economics.& Organization.
V7 Sp

by the inverse functiontheorem.The second equationin (4) characterizeshow

changes in the "prices" a affect the level of effort that will be supplied.
Using Equations(3) and (4), one can compute first-ordernecessary condi-
tions for an optimum in (1) and (2) when t > 0:

a = (I + r[Cij]Z)-IB', (5)

where B' = (B,,...,B,,) is the vector of firstderivativesof B. Condition(5) is

also sufficient when the expression C'(t)T.C'(t) is convex in t.
As a benchmarkcase, note that when the error terms are stochastically
independent(X is a diagonal matrix) and the activities are technologically
independent(all cross-partialsof the cost functionarezero), the solution in (5)
simplifies to ai = B,(1 + rC,ao2)- , for all i. In this case, commissions areset
independentlyof each other since the cost of inducing the agent to perform
any given task is independentof the othertasks. As expected, a, is decreasing
in risk aversion(r) and risk (a2). It is also decreasingin Ci. To interpretthis,
note from (4) that atilaa, = 1/Ci. Thus, the above formulasays thatai should
be higher, the more responsive the agent is to incentives.
In the general case, notice that the cross-partialsof C but not those of B
enter into (5). Complementaritiesin the agent's private cost of generating
signals can have an importantrole in determiningoptimal incentive pay. To
illustrate, consider again the case of motivatingteachers to teach both basic
skills and higher-orderthinkingskills, but assumingthathigher-orderthinking
skills cannot be measured. We model this by supposing that there are two
activities so that the agent chooses the pair (t ,t2), but that only one activity
(teaching basic skills) is observable:

x = t + E. (6)

We can apply (5) assuming that a2 is infinite and a12 is zero. Then, if the
optimal solution entails t > 0, it must satisfy

a, = (B1 - B2C12/C22) / [1 + rao(Cl - C12/C22)].9 (7)

When C 2 is negative, makingit more negative leads to a largeroptimalvalue

of a,. That is, when the activities of teachingbasic skills and higher-thinking
skills are complementary in the agent-teacher'sprivate cost function, the
desirabilityof rewardingachievementin teaching basic skills is enhanced. If

9. We are assuming that teachers are motivated to teach some higher-thinkingskills even
without explicit financialincentives to do so. In one-dimensionalagency models, it is typically
assumed that the agent will not work without incentive pay. The reason for this is not that the
agent dislikes even small amounts of work, but ratherthat the level of work the agent would
provide without explicit incentives does not affect the optimal solution. In multitask models,
however, the fact that agents supply inputseven withoutincentive pay can be quite consequential
as the teacherexample and the example in Section 3.2 show.
 Multitask Principat-Agent Analyses 33

the two dimensionsof teachingare substitutesin the agent'scost function(C12

> 0), then a, is correspondinglyreduced,because high values of aI cause the
teacher to substituteeffort away from teaching higher-thinkingskills.
In general, when inputs are substitutes, incentives for any given activity ti
can be provided either by rewardingthat activity or by reducing its oppor-
tunitycost (by reducingthe incentives for the otheractivities). Here, t2 cannot
be measuredat all, so the only way to provide incentives for t2 is to reduce a,
as (7) shows.
Notice that (7) allows the possibility that it may be optimal to set a,
negative even if B, is positive, providedB1 < B2C12/C22.If the agent can
always reduce measuredperformanceat no cost to himself, then this observa-
tion can be used to producerobustexamples in which it is optimal to provide
zero incentives for a desirable activity even when perfectly reliable perfor-
mance measures(a2 = 0) may exist. A second case in which zero incentives
can arise in this example is when effort in the two activities are perfect
substitutesin the agent's cost function and the second activity is unobserva-
ble-that is, when C(t ,t2) = c(t, + t2) and a2 = + oo. Then, if t > 0, the
incentive constraintin (3) implies that a, = a2 (intuitively, the agent must
equate the marginalreturnto effort in various tasks). If, as in our teaching
example, a2 = oo, it then follows that 0 = a2 = a,. This idea resurfaces in
several of the applicationsin Section 3.
Another importantpossibility is that (5) does not apply because it is not
optimalto set t > 0. Even in the model where t is one dimensional,the cost of
providing positive incentives for a small amount of effort is discontinuously
higher than the cost of providing no incentive for effort if C'(O) > 0. If no
effort is requiredand no incentive is provided, then the risk premiumincurred
by the agent is zero. If a small amountof effortt is required,then a = C(t) >
C'(O) > 0 and the risk premiumis thereforeat least -r[C'(0)]2Cr2.Providing
incentives for an activity involves an inherentfixed cost, and the size of that
cost can be affected by the selection and levels of the agent's other activities.
These observations will prove to be importantwhen we apply our theory to
issues of employment and job design.

3. Allocation Incentives for Effort and Attention

3.1 The Effortand AttentionAllocationModel
We now move to a groupof models in which the agent'seffortor attentionis a
homogeneous input that can be allocated among tasks however the agent
likes. We shall suppose thateffortin the varioustasks is perfectlysubstitutable
in the agent's cost function. More formally,we supposethatthe agent chooses
a vector t = (tl,...,tm) at a personal (strictly convex) cost C(Qt +..+ tm),
leading to expected profitsB(t) and generatingsignals x(t) = u(t) + e. Then,
if the agent increases the amountof time or attentiondevoted to one activity,
the marginalcost of attentionto the other activities will grow larger.
Contraryto most earlierprincipal-agent models, we shall not suppose that
all work is unpleasant(see note 9). A workeron the job may take pleasurein
working up to some limit; incentives are only requiredto encourage work
 34 The Journalof Law,Economics,&Organization,
V7 Sp

beyond that limit. Formally,we assume thatthere is some number > 0 such
that C'(t) 5 0 for t -i and C(t) = 0. This is important,because it means that
contractsthatprovidefor fixed wages may still elicit some effort, thoughmore
may be elicited by providingpositive incentives. It also means that there is a
range of effort allocationsamong which the agent is indifferentand willing to
follow the principal'spreference.

3.2 Missing IncentiveClauses in Contracts

One of the most puzzling and troublingfailures of incentive models has been
their inability to account for the paucity of explicit incentive provisions in
actual contracts. For example, it is surprisinglyuncommon in contractsfor
home remodeling to incorporateexplicit incentives for timely completion of
construction, even though construction delays arise frequently and can be
profoundlydisruptiveto the homeowner. There can be little doubt that such
clauses could be written into the contracts;similar clauses are common in
commercialconstructioncontracts.We shall arguethat these facts can best be
understoodas a result of the greaterstandardizationof commercialconstruc-
tion and the consequent ability of commercialbuyers to specify and monitor
qualitystandards.The innovationin our analysis is thatour explanationof the
presence or absence of the timely completionclause lies in an examinationof
the principal'sability to monitorother aspects of the agent's performance.l0
Thus, suppose that some desirable attributesof the contractor'sperfor-
mance (such as courtesy,attrntionto detail, or helpful advice) are unmeasura-
ble but are enhancedby attentiont, spent on that activity, while other aspects
of quality (such as timely completion) are measurable(perhaps imperfectly)
and enhancedby attentiont2 devoted to this second activity. Supposing that
the measuredqualityis one dimensional,we may write u(tI,t2) = #(t2), x -
+ e. As we have seen, the agent's efficient compensationcontract pays an
amountS = ax + t3.
Suppose thatthe overall value of thejob to the homeowneris determinedby
the functionB(t,,t2). To model the idea that the first activity is "very impor-
tant" and that both activities are valuable, we assume thatB is increasingand
- 0.
that B(0,t2) = 0, for all t2

Proposition 1. For the home contractormodel specified in the last para-

graph, the efficient linear compensationrule pays a fixed wage and contains
no incentive component (a = 0), even if the contractoris risk neutral."1

10. Anotherplausible explanationis that home constructioncontractsare frequentlychanged

to reflect design modifications, and timely completion clauses would be nullified by these
changes.
11. A related conclusion-that incorrectweightings of profit contributionsin an accounting
system lead the optimizing employer to weaken effort incentives-is derived by Baker. Baker's
analysis can be conductedwithin our model by recognizingthata state-contingentstrategyfor the
agent-that is, making different decisions in different states-is equivalent to a vector effort
strategy. Baker's assumptionthat the principalcannot distinguishthe state in which an action is
 Principal-AgentAnalyses 35
Multitask

Proof. If a = 0, then the agent can be instructedto spend total time i where
C'(t) = 0 and to choose tf E [0,t] to maximize B(tl,f - i,), which is strictly
positive because f > 0. In this case, the cost of risk-bearingby the agent is
zero, so the total wealth will be B(l ,i - i ) - C( ). If a > 0, then t, will be
set to zero and the total wealth will be 0 - C( t) - ra2 2/2 ' -C( t) < B(ft ,
- f ) -- C( ) because i is cost minimizingfor the agent. If a < 0, then t2 = 0
and t, < t [because C'(t,) < 0 = C'(f)] so the total profits are

B(t,O0) - C(t,) - ra2a2 < B(,O) - C(t) < B(t,j - tl) - C(i).
Q.E.D.

The ideas that underliethis analysis have many applications.For example,

piece rates are relatively rarein manufacturingand, where they are used, they
are frequentlyaccompaniedby careful attentionto monitoringthe quality of
the work. Our analysis indicates that if quality were poorly measured, it
would be expensive or impossible to maintain good quality while using a
piece-rate scheme. Similarly, where individualsspend partof their efforts on
individualprojectsand parton team production,and assumingthat individual
contributionsto the team effortare difficultto assess, it would be dangerousto
provide incentives for good performanceon the individualprojects.The prob-
lem, of course, is that individuals may shift their attention from the team
activity where their individualcontributionsare poorly measuredto the better
measured and well-compensated individual activity. For this reason, piece-
rate schemes may be especially dysfunctionalin large hierarchies.

3.3 "Low-PoweredIncentives"in Firms

A similar model can be used to explain Williamson's observation that the
incentives offered to employees in firms are generally "low-powered"com-
pared to the "high-powered"incentives offered to independentcontractors.
Like Williamson, we distinguishemployees from independentcontractorsby
the condition of asset ownership:Employees use and develop assets that are
owned by others while contractorsuse and develop their own assets.
Once again, the heart of our modeling is our assumption that there are
multiple activities to be undertakenand that the allocation of time and atten-
tion between them is crucial. Thus, let the expected gross profit from the
enterprise be the sum of two parts, B(t,) + V(t), where B representsthe
expected net receipts and V the expected change in the net asset value. We
assume that B and V are increasing, concave, and twice continuouslydiffer-
entiable and thatB(0) = V(0) = 0. The actual change in asset value, V + e,V
accrues to whoever owns the asset. Assets are notoriouslyhardto value (that
is why accountantsgenerally use historical cost as a valuationbasis), so we

taken is then the same as assuming that the principalcannot distinguish performancealong the
several dimensions of the vector strategy.The formal mappingfrom these "hiddeninformation"
models to our "hidden action" model is discussed in Holmstromand Milgrom (1987).
36 The Journalof Law,Economics,&Organization,V7 Sp

assume that there is no performance indicator for the asset enhancement

activity t2. The primaryactivity tI is to produceoutputfor sale in the current
period:its indicatoris x = u(t,) + ex, wherep is increasingand concave. We
assume that e, and e, are independent.
We consider two alternativeorganizationalmodes-contracting, in which
the change in asset value accrues to the agent, and employment,in which the
change in asset value accrues to the firm or principal. The crucial difference
between these lies in the incentives for the agentto engage in the two kinds of
activities. To focus on the most interestingcase, we will assume that it is
highly desirable to induce the agent to devote a positive amountof effort to
both activities. Let

z1 = Max B(tj) - C(tl),

tI

n72 = Max V(t2) - C(t),

t2

r12= Max B(t,) + V(t - tl) - C(f).

tI

Proposition 2. Assume that r12 > Max('r1,r2). Then, the optimal em-
ployment contractalways entails paying a fixed wage (a = 0). Wheneverthe
independentcontractingrelationis optimal, it involves "high-poweredincen-
tives" (a > 0). Furthermore,thereexist values of the parametersr, o2, and a2
for which employmentcontractsare optimalandothersfor which independent
contractingis optimal. If employment contractingis optimal for some fixed
parameters(r,a2,a2), then it is also optimalfor higher values of these param-
eters. Similarly, if independentcontractingis optimal, then it is also optimal
for lower values of these parameters.2

Proof. First, consider the case of the employment contract, where the
returnsV(t2) accrue to the firm. If the principalsets a > 0, the agent will
respondby setting t, so that a = C'(t,) and setting t2 = 0. The total certainty
equivalent wealth is equal to B(t,) - C(tI) - rao2a2< 7r1 < 7r12. However, if
a = 0, the agent is willing to spend time f in any proportionsand so a total
certaintyequivalentwealth of r'2 is obtained.Therefore,it is optimal for the
principalto set a = 0 in an employmentcontract.
For the independent contractor,the maximum total certainty equivalent
wealth is computedas follows. Let (t1(a),f2(a)) maximize acu(tl) + V(t2) -
C(tI + t2) - ra2/2; this representsthe agent'soptimalresponseto a. The total
certaintyequivalent wealth for any fixed a is

12. One can derive a similar result with a general quadraticcost function C(tl,t2). The only
difference is that the commission rate would not necessarily be zero for an employed agent,
though it would always be smaller than for an independentcontractor.
 Multitask
Principal-AgentAnalyses 37

B(t,(a)) + V(t2(a)) - C(it(a) + t2(a)) - -{( x[C'(i,(a) + t2(a))]2 + a2}

and the maximumsurplusis the maximumof this expressionover a. If we fix

a = 0, then this expressionis lower for the independentcontractorthanfor the
employment regime. Hence, whenever the independentcontractorregime is
optimal, it must be optimal to set a > 0.
Note that if a2 = a2 = 0, first best is achieved by setting a = 1 in the
independentcontractorregime. Since first best never can be achieved in the
employmentregime, the independentcontractorregime is betterin this case.
Letting ra2 grow large makes the payoff to the independentcontractregime
fall without limit, so there are also some parametersfor which the em-
ployment regime is better.
The last two sentences of Proposition2 follow from the observationthatthe
expression for the total certainty equivalent in the independentcontractor
regime is decreasingin ra2 and ra2, but these two termsdo not affect the total
certaintyequivalent in the employment regime. Q.E.D.

Proposition 2 is consistent with the evidence reportedby Anderson and

Anderson and Schmittlein. They attemptedto identify the reasons why firms
in the electronic components industryhave an employed sales force in some
districts and independentsales representativesin other districts. (Many, but
not all firms, used both forms of sales organization,suggesting that econo-
mies of scale play a lesser role.) They found that the perceived difficulty of
measuringsales of individualsalespeople (due to team selling or costly record
keeping) was the best empiricalpredictorof the use of an in-house sales force.
Transactioncost variables-such as specific training, with the exception of
confidentialinformation-were not significanteither alone or in conjunction
with performancemeasurement.If we suppose that one function of the sales
force is to build an asset that is impossible to measure, such as "goodwill"
(how satisfied and loyal are the customers?),then our model suggests that the
difficulty of measuringsales would lead to the patternof sales organization
that Andersonand Schmittleinobserved, and that commission rates would be
lower for company-runsales forces.13
Anderson also finds that the importanceof nonselling activities, such as
promoting new productsor products with a long selling cycle, is positively
related to the use of an in-house sales force. We can analyze this finding by
introducinga thirdactivity t3, which benefits the principal,but not the agent.
Since an independentcontractorwill spend no time on nonselling activities

13. It may be arguedthatrisk aversioncannot be a very relevantfactorif the independentsales

representativeis itself a large firm. Recall, however, that the cost of risk can equivalently be
derived from imperfectobservabilityof B, paired with a convex cost C, in a risk-neutralmodel
(see Baker). We can rely on work aversioninsteadof risk aversion. In this case, to make sure that
it is not optimal to transferB to the independent sales firm (i.e., make the manufacturera
subcontractor),one has to add an imperfectly observed input by the manufacturer.With two
equally important,equally costly, and imperfectlyobserved inputs, the residualreturnB will be
allocated to the party whose input is more difficult to measure.
 V7 Sp
38 The Journalof Law,Economics.&Organization,

(just as practitionersclaim), it is easy to see that an increase in the value of

this activity will work in favor of an owner-runsales force in our model.
Another piece of evidence consistent with our model comes from the fast-
food industry.Firms such as McDonaId'sand BurgerKing own about30% of
their stores and franchisethe rest. The differencein incentives between fran-
chisees and owner-managedfirms is striking. Franchiseespay royalties that
are at most 10% of sales, correspondingto at least a 90% commission,
whereas managers of company-owned stores typically receive no explicit
incentives either on profit or sales (Krueger, Brickley and Dark). The dif-
ference in incentives is all the more remarkable,consideringhow similar the
two types of stores are in all other aspects. According to our theory, the
discontinuousshift in residualreturns[V(t2)]associated with franchisingand
the attendantshift in attentiontowardlong-termasset values and cost contain-
ment, forces the franchisecontractto increase short-termincentives sharply.
Or, looked upon the other way, short-termincentives for employed managers
must be muted to prevent them from allocating their attention away from
important,but hard to measure, asset values.

4. Limits on Outside Activities

Ourpreviousanalysis emphasizesthe importanceof studyingthe full rangeof
the agent's activities for analyzing incentives. If activities interact in the
agent's cost function, incentive strength can be predicted only once the
agent's whole portfolioof tasks is known. An equally importantimplicationis
that the principalcan influencethe agent's incentivesby choosing the agent's
portfolio of tasks. In the next section, we will study the optimal allocationof
tasks between two agents. In this section, we consider how the principal
might try to manage the agent's access to outside (private) activities.
Even casual observation makes it clear that the rules governing outside
activities depend on the job. It is a commonplaceobservationthat employees
in "responsible positions" are allowed more freedom of action than other
employees, and that they use that freedom in partto pursuepersonallybene-
ficial activities. To analyze the issues that this observationraises, we begin
with the assumptionthat it is easier for an employer to exclude an activity
entirely than to monitor it and limit its extent. For example, a rule against
personal telephone calls during business hours is found in many offices and
seems to be motivatedin partby its ease of enforcementcompared,say, to a
rule that limits the percentageof business hours devoted to personal calls to
2%. Althoughgeneralizationsaboutemploymentall seem to have exceptions,
a common featureof employmentcontractsis thatthe employerhas authority
to restrictthe employee's outside activities duringbusiness hours, and some-
times after hours as well.
Assume then that the agent has a finite pool K = {1,...,N} of potential
activities, which the principalcan control only by exclusion. The returnsto
these tasks, which we will refer to as the agent'spersonal business for short,
are assumed nonstochasticand to benefit the agent alone (in principle, these
tasks could benefit the principal,too, but the analyticswould be more compli-
 Multitask Principal-Agent Analyses 39

cated). The principalcontrols the agent's personal business by allowing the

agent to engage only in a subset of tasks A C K. Within the set of allowable
tasks, A, the agent can engage in as much or as little personalbusiness as he
pleases, but none outside A. To focus on the interactionsbetween the agent's
workplaceactivities and personalbusiness, we representworkplaceactivities
simply as a single task in which performanceis imperfectlymeasured.
Let t denote the attentionthe agent devotes to the principal'stask and tk the
time he devotes to personalbusiness k. We model the personalbenefitsthatthe
agent derives as an offset against, or deduction from, his personal cost of
effort, as follows:

= C(t + - (8)
c(ttl,...tN) 2Ktk) XKVk(tk).

The notation XK stands for summation over k in K. Here C is the agent's

privatecost of the total attentionhe devotes to all his (permitted)personalac-
tivities. The returnfrom personalactivity k is measuredby the functionvk(tk);
these functions are assumed to be strictly concave with vk(O)= 0. If k EA,
then tk = 0, so we could replace ZK with ZA in (8).
We make the simplifying assumptionsthatthereareconstantreturnsto time
both in generatingprofits and in improvingmeasuredperformance:

B(t,t, ...,tN) = pt, X(t,tl,...,tN) = t + e. (9)

The variance of e is a2.

The principal's control instrumentsare the commission rate a and the
allowed set of personal business tasks A C K. We will study the principal's
problem in two stages. First, we fix a and consider the optimal choice of A,
denoted A(a), and then we determinethe optimal a.
Given the parametersa and A, the agent chooses t and t. to maximize

at + 2AVk(tk) - C(t + ZAtk).

Assume for the moment that this problem has an interiorsolution. Then the
first-orderconditions that characterizethe agent's optimum are

a = C'(t + 2tk), (10)

a = v,(tk)- (11)

We note from (11) that the amountof time the agent chooses to spend on task
k, denoted tk(a), only depends on a and not on A. Also, the total time spent
working, t + 2Atk, is independentof A. Consequently,if the agent is allowed
more personaltasks, without a change in a, all the time for those tasks will be
reallocatedaway from the principal'stask;this is the convenienceof assuming
(9) together with a cost function that only depends on total time. It makes it
very simple to determinewhich personal tasks the agent should be allowed
40 The Journalof Law,Economics,& Organization,
V7 Sp

for a given a. The benefit of allowing the agent to spend time on task k is
vk(tk(a)),while the (opportunity)cost is ptk(a). Therefore, the optimal set of
allowable personal tasks is

A(a) = {k E Kv,(tk(a)) > pt,(a)}. (12)

Figure 1 shows the determinationof A(a). The pt line representsthe returns

from spendingtime on the principal'stask. The v1 and v2 curves representthe
returnsfrom two privatetasks. Both privatetasks are socially valuable in that
the vk curvesrise above thept line on a positive intervaltkE [0,4], where4 is
definedby the intersectionvk(tk) = p4. However, for the chosen a, only task
1 is worth keeping; it is optimal to exclude task 2 since t2(a) > tF-that is,
time t2(a) yields more in the principal'stask than it yields in task 2.
The geometry of Figure 1 makes it evident that A(a) expands as a is
increased. This follows because tk(a) is decreasingas v. is strictly concave.
As a is raised, the agent will spend less time on privatebusiness. This brings
more projectsinto the efficientregion tk(a) ' t4,which is characterizedby the
condition thattime tj(a) in the privatetask yields more than the same amount
of time spent in the principal'stask. Furthermore,we see thatthe criticalvalue
of a at which private task k will be excluded is entirely determinedby the
slope of vk at the point where vkintersectsthe pt line. This follows since tk(a)
^ a.
< t4 if and only if v^.()
We record these observationsin the following proposition.

Proposition 3. Assume that a is such that t(a) > 0. Then the following
statementshold.
(i) It is optimal to let the agent pursue exactly those private business
opportunitiesthat belong to A(a) defined in (12); that is, those tasks k for
which the resulting average product v(t,(a))/tk(a) exceeds the marginal
product p in the principal'stask.
(ii) The higher is the agent's marginalrewardfor performancein the main
job, the greateris his freedom to pursuepersonalbusiness. Formally,if a
a', then A(a) D A(a').
(iii) If it is optimal to exclude task k, then it is also optimal to exclude all
tasks m, for which v,'(ti) > v(t4), where tf is defined by vj(t) = pi.

It is possible that for small enough a it will be optimal to set t(a) = 0 and
hence A(a) = K. In that case, there are no gains from tradeand the principal
will not employ the agent. Such a solution may be optimal if the cost of
bearingrisk becomes sufficentlylarge. One could exclude thatcase by assum-
ing that the agent's privatebusinesses are less productivethanworkingfor the
principal with zero incentive (as we saw earlier, zero incentive does not
preclude productive work), but there is no need to make such a restriction.
Obviously,if t(a) = 0, then t(a') = 0, for all a'<a. Therefore,job separation
will occur, if at all, below a critical cutoff value for a.
Part (ii) of Proposition 3 articulatesa familiarand fundamentalprinciple:
 Multitask Prtncipal-Agent Analyses 41

Return
($)

t1(c0) t, t t ((0)
11 2 2
Effort
Figure1. Itis optimalto allowtask 1 but to exclude task 2, because tl(a) < fi but f2 <
t2(a). Notice that this is true even though the social returnsto task 2 are everywhere
higher than those to task 1.

responsibilityand authorityshould go hand in hand. It is optimal to give the

agent more freedom to pursuepersonalbusiness when he is financiallymore
responsible for his performance.In the extremecase, when performancecan
be measured without error and it is optimal to make the agent a residual
claimant(a = p), the agent will be free to engage in whateverprivatebusiness
he deems desirable. The responsibilityprinciple again underscoresthat the
agent's incentives can be influencedindirectlyby alteringthe opportunitycost
for supplying desired inputs. It is readily seen that the agent's marginalcost
(but not total cost) of spendinga given amountof time t in the principal'stask
42 The Journal of Law, Economics. & Organization, V7 Sp

is reduced by excluding private tasks. Exclusion will be more extensively

used the more costly it is to provide financialrewards.14
Part (iii) of the Proposition shows that the social value of a personal ac-
tivity, and the likelihood that it will be excluded, need bear little relationship
to each other.For instance, in Figure 1 the social value of task 2 is higherthan
that of task I for any given amount of time spent on either, yet it is task 2
ratherthantask 1 thatis excluded. The reasonis thattask 2 more easily invites
excess attention.
Before turning to the choice of a, there is a point that deserves to be
emphasized. The amountof personalbusiness A that the principalwill allow
for a fixed a, as characterizedin (12), does not dependdirectlyon r or o2, nor
on the cost function C. These factors affect A only through a. Therefore,
given data aboutr, a2, a, C, andA, it is econometricallycorrectto regressA
against the endogenous variable a. Proposition3 predicts that the extent of
agent freedom will be positively related to a, irrespectivelyof which of the
model parameters(other than vk) are viewed as exogenous. The parametersr
and a2 are a naturalsourceof cross-sectionalvariationin a as indicatedby the
following.

Proposition4. Assume thatthe optimalsolutionfeaturest(a) > 0. Then the

following statementshold.
(i) The optimal value of a is given by

a = pl[1 + ro21(dtlda)l,

where dtlda = 1/C" -+ A(a)(l/V1).

(ii) If it becomes easier to measurethe agent's performance(a2 decreases),
or the agent becomes less risk averse (r decreases), then the agent's marginal
reward a will be raised and his personal business activities will be less
curtailed.
(iii) Any personal task that would be excluded in a first-bestarrangement
[v*(0) ' p] will also be excluded in a second-best arrangement.For suffi-
ciently high values of ra2, some tasks that would be included at the first best
will be excluded at the second best.

Proof. The equationin (i) is a special case of (5); the expressionfor dtlda
follows from the agent's first-orderconditions. Revealed preference paired
with Proposition3 implies (ii). Part (i) implies that a c p and that a goes to
zero when ro2 goes to infinity;this proves (iii). Q.E.D.

If we assume that the agent's cost and benefit functions are quadratic,we
see from (i) that the agent's responsivenessto incentives, dtlda, increases as
the set of allowable tasks, A(a), expands. Consequently,viewing A(a) as

14. One can also show that by excluding privatetasks, the agent becomes less responsive to
increases in the commission a. A less flexible job design is associatedwith weaker incentives as
we mentioned earlier.
 Multitask
Principal-AgentAnalyses 43

exogenous, it is optimal to raise a in response to an increase in the agent's

degree of freedom. Of course, A(a) is not exogenous; it expands with an
increase in a. We see then that a and A(a) are complementaryinstruments:
increasing either leads to an increase in the other.15
Part (ii) is the most interesting one. It predicts that there will be more
constraintson an agent's activities in situations where performancerewards
are weak because of measurementproblems. The rigid rules and limits that
characterizebureaucracy,in this view, constitutean optimalresponse to diffi-
culties in measuringand rewardingperformance.Among the "personalbusi-
ness" activities that bureaucraciestry to limit are collusion (Tirole; Holm-
strom and Milgrom, 1990; Itoh, 1989) and influence activities (Milgrom,
Milgrom and Roberts). The restrictions on trade between employees that
Holmstromand Milgrom (1990) recommendand the restrictionson commu-
nications that Milgrom and Roberts propose are examples of optimal exclu-
sion of activities that would be permittedor perhaps even encouragedin a
first-bestworld.
The desire to exclude activities provides a second possible explanationof
the empirical results of Anderson and Schmittlein. Here, rather than dis-
tinguishingthe roles of employee and independentcontractoron the basis of
ownershipof productiveassets, the focus is on the discretionaryauthorityof
the employerto preventsalespeople fromoutside activities, such as selling the
products of other manufacturersduring business hours. As the difficulty of
monitoring performance(measured by a2) rises, Proposition4 asserts that
there is an increasingdegree of exclusivity in efficientcontracts.If exclusivity
is easier to enforce within firms than across firms, then poor sales measure-
ment and employment are positively related.
Our two explanations of the Anderson-SchmittIeinevidence are distinct
but closely related. In the first, the extra incentive from employmentcomes
from transferringto the firm the returnstream associated with the goodwill
created by customer satisfaction. In the second, the extra incentive comes
from eliminating (ratherthan transferring)a returnstream-that associated
with personal business. In each case, eliminating the agent's direct profits
from an activity reduces the opportunitycost of work to the employee and
lowers the cost of providing incentives.

5. Allocating Tasks between Two Agents

In the single-agent model, the commission rates a, serve threepurposes:they
allocate risk, motivate work, and direct the agent's efforts among his various
activities. A trade-offarises when these objectives are in conflict with each
other: Optimal risk-sharingmay be inconsistent with motivating work, and
motivatinghardwork may distortthe agent's allocationof effortsacross tasks.
Among the instrumentsavailable to the principalto alleviate these problems

15. Nonlinearitiesin the principal'stask would not alterthe conclusion thata is reducedwhen
r or a2 is increased;this partis just a revealedpreferenceargument.However, the set A(a) would
be harderto characterizeas the exclusion of tasks would interactwith each other as a result of
integer problems. One could even find that a personal task is included when a is reduced.
 44 V7 Sp
The Journalof Lav. Economics.&Organization,

arejob restructuringand relative performanceevaluation:The formerallows

the principal to reduce the distortions in how attention is allocated among
activities, while the latterenables the principalto lower the cost of incentives
by using a more sensitive measureof actual performance.

5.1 OptimalGroupingsof Tasks into Jobs

Here we initiate the study of how incentive considerationsmight affect the
grouping of tasks into jobs. We use a model that eliminates other important
effects, such as differences among the agents and complementaritiesamong
task assignments. There are two identical agents, indexed i = 1,2, who
allocate their attentionacross a continuumof tasks indexed by k E [0,1]. Let
ti(k) denote the attentionagent i devotes to task k. We assume that the two
agents can sharea task andthattheirlaborinputsare perfectsubstitutes.Thus,
profitB(t) is a functionof the total time vector t {t(k):k E [0,1 ]}, where t(k)
= t,(k) + t2(k).Likewise, the performancesignal from task k, t(t(k),k), only
dependson the total attentiont(k) devotedto it. The errorvarianceof task k is
a2(k) > 0 and the errorsare assumed independent.
Agent i's total labor input is given by

ti = J t,(k)dk. (13)

His privatecost is C(i,); the cost functionis assumeddifferentiableand strictly

convex.
Since the ex ante specificationof the model is symmetricin the roles of the
two agents, if the problem entailed a concave objective and convex con-
straints, we would expect the optimal solution to be symmetric. However, as
we shall see, the optimal solution is not symmetric,so we must be careful to
deal correctly with the inherentnonconvexities of the problem.
We begin by studying the problem of implementing, at minimum cost, a
given vector t = {t(k)} of total attentionto be devoted to the various tasks,
given the constraintthat the total attentiondevoted by agent i is tf. Denoting
the commission paid to agent i for task k by a,(k), this problemis describedby

Minimize C(i,) + C(t2) + 2 f [al(k) + a2(k)]a2(k)dk,

al(),a2(-) (14)

subject to (12), (13), and the incentive constraints

a,(k)u'(t(k),k) <C'(i ), if t,(k) = 0,

a,(k)'(t(k),k) = C'(,i), if t,(k) > 0, i=1,2, k E [0,1].

The incentive constraintscan be correctlydescribedby first-orderconditions,

because the agent's choice problem is a concave maximizationproblem. As
usual, the implementationcost reflectsboth the direct cost of work as well as
the cost of risk-bearing,since both costs are deductedwhen determiningthe
total certaintyequivalent of the parties.
 Multitask Principal-Agent Analyses 45

We shall say that the principalmakes the two agentsjointly responsiblefor

task k if al(k) > 0 and a2(k) > 0. Similarly,agent i is solely responsiblefor
task k if a,(k) > 0 and aj(k) = O, i $ j.

Proposition5. In the model describedabove, it is never optimalfor the two

agents to be jointly responsible for any task k.

Proof. Let k be a set of tasks for which thereis joint responsibility-that is,
a1(k)a2(k) > 0, for k E K-and suppose K has positive measure.Let ti(K)
fKti(k)dkand choose K' C K such that fK,t(k)dk= t (K). Define a new set of
attentionallocations and commission rates {It(k),ai(k)}so that these coincide
with the original specificationfor k ? K. For k E K', set ? (k) = t(k), da(k) =
a,(k), and t2(k)= d2(k) = 0; for k E K\K', set t1(k)(k)d(k) = 0, t(k) = t(k),
and a2(k) = a2(k).
The total attentiondevotedto each task as well as the total attentionof each
of the two agents is unalteredin the new scheme. By construction,therefore,
the first-orderconditions (15) hold and the new scheme is feasible. The new
scheme strictly improves the objective function as some of the commission
rates are lowered to zero for a set of tasks of nonzero measure. Q.E.D.

This proposition reflects our earlier observationthat providing incentives

for an agent in any task incurs a fixed cost as the agent assumes some
nontrivialfraction of the risk associated with that task (or its measurement).
Since we have assumed that the tasks are small relative to the agent's ca-
pabilities, assigning joint responsibility for any task would incur two fixed
costs unnecessarily.As the proof demonstrates,if one begins with an arrange-
ment in which some tasks are shared, it is possible to split the same tasks
among the agents without affecting either the total effort requiredof either
agent or the total effort allocated to any task. This rearrangementmakes it
possible to eliminate some of each agent's responsibilities[setting a,(k) = 0],
thereby reducingthe risk that the agent must bear and so increasingthe total
surplus of the three parties.
Having establishedthateach task will be assigned to just one employee, we
next turn to the issue of how the tasks will be grouped. With this in mind,
it is convenient to redefine our variables. We reinterpretai(k) to be the
hypotheticalcommission rate that the principalwould need to pay in orderto
elicit the desiredlevel of effortt(k) from agent i if he were assigned task k [see
(17) below]. We also define a task assignmentvariableI,(k), which is set equal
to unity if agent i is assigned task k and is set equal to zero otherwise. Then,
the actual commission rate paid to agent i for task k is a,(k),i(k); that is, it is
a,(k) if i is assigned the task and it is zero otherwise. Proposition3 implies
that at the optimum, ti(k) = I,(k)t(k). We can now state the principal'stask
assignment problem as follows:

rI
Minimize C(t1) + C(t2) +
l
j [ l (k)a2(k) + I2(k)a(k)]a2K)dk (16)
(k),12(k)
46 The Journalof Law,Economics,& Organization,
V7 Sp

subject to

a,(k),u'(t(k),k)= C'('i), i = 1,2, k E [0,1], (17)

f I(k)t(k) = ii, i = 1,2. (18)

l(k) + 12(k) = 1, k [0,1], (19)

I,(k) = 0 or 1, i = 1,2 and k E [0,1]. (20)

Constraint(17) merely defines a,(k), since t(k) and ti are fixed. If tf = f2,
then it is clear from (17) that al(k) = a2(k), and hence thatthe objective (16)
is independentof the task assignment:All feasible assignmentsthen yield the
same total certaintyequivalent wealth. As we will see below, the important
case is the asymmetricone, so let us assume that fi < t2.
To solve program(16)-(20), we first solve the relaxed programin which
(20) is replaced by the less restrictiveconstraint

I,(k) > 0, i = 1,2 and k E [0,1]. (21)

In the relaxed problem, the objective and constraintsare all linear (hence,
convex) in the choice variables I,(k), so first-orderconditions fully charac-
terize the optimum. Let y, be the Lagrangemultiplierassociated with con-
straint (18). Then, optimizing in the usual way, we find that

l,(k) = 0, if (r/2)[a2(k) - a2(k)]a2(k) + (y - y2)t(k) > 0

and (22)

11(k) = 1, if (r/2)[a2(k) - a2(k)]2(k) + (7y - y2)t(k) < 0.

By (17), a, < a2 as F, < t2; therefore, (22) implies that y, > y2- Since li(k)
takes values 0 and 1 at the optimum of the relaxed programwith constraint
(21) in place of (20), Equations (22) also characterizethe solution to the
original problem and identify the marginal tasks. A marginal task is one
where the advantageof assigning the task to agent 1, in termsof the lower risk
premium required, is just offset by the higher marginal value of agent l's
time. The firstof these costs varies with the measurementerrorattachedto the
task and the second varies with the amountof time the task requires.These
observations suggest an alternativecharacterizationof the optimum assign-
ment policy.
Define the noise-to-signal ratio of task k by n(k) = a2(k)/tu'(t(k),k)2and the
informationcoefficient by o(k) = n(k)/t(k). Let

= (2/r)(y, - -
2)[C'(2)2 C'(il)21- (. (23)
 Multitask Principal-Agent Analyses 47

We can then restate (22) as follows.

Proposition6. Suppose thatthe two agents devote differentamountsof total

attentionto theirtasks (i.e., tl < t2). Then, tasks are optimallyassigned in this
model so that all the hardest-to-monitortasks are undertakenby agent 1 and
all the easiest-to-monitortasks are undertakenby agent 2. That is, agent 1 is
assigned all the tasks k for which e(k) > e, and agent 2 is assigned all those
with g(k) < p, where g is defined in (23).

Corollary. Suppose that 1 < t2, the requiredallocation of attentionis uni-

form [i.e., t(k) = 1 for all k] and the signal functionsare identical [i.e.,/u(t,k)
= u(t)]. Then there exists a x such thatagent 1 will optimallybe given all the
tasks k for which a2(k) > a2(x), and agent 2 all the tasks for which a2(k) <
o2(x).

These results provide, in purely incentive-theoreticterms, an account of

how activities might be grouped, with some employees specializing in ac-
tivities that are hard to monitor and others in activities that are easily
monitored. 6 Separatingtasks accordingto their measurabilitycharacteristics
[g(k)] allows the principalto give strong incentives for tasks that are easy to
measurewithout fearingthatthe agent will substituteeffortsaway from other,
harder-to-measuretasks. The present model oversimplifies these issues by
assuming that there are no restrictionson how the principalmay group tasks.
In the case of piece rates discussed in Section 3, it might not be possible to
separate the tasks of providing high output from those of providing high
quality: The worker might always be able to substitutespeed for attentionto
details. Nevertheless, the results of Proposition6 are suggestive.17
The appearanceof Q(k) in these results is unfamiliar, and seems worth
reviewing in detail. Let realized performancein task k be measuredby

Xk(t) = 4u(t,k) + ek,

where Ek is distributednormally with zero mean and variance a2(k). The

normalizedperformancemeasure,

Xk(t) = [u(t,k) + Ek]/['(t(k),k),

provides the same informationand has errorvariance equal to the noise-to-

16. Minahan derives a result that is related. In his model there are four tasks: two easy to
measure and two hard to measure. He shows that it is better not to mix the tasks. The main
differencebetween his model and ours is that in his model the principalcannotprovide incentives
on individualtasks, just on the sum of the tasks. This would greatlysimplify our analysis. On the
other hand, Minahan's analysis deals with nonlinear incentives and general utility functions,
which adds to the complexity.
17. One manifestationof the task allocation principle may be found in the organizationof
R&D activities in firms [see Holmstrom(1989)].
48 The Journalof Lav. Economics,&Organization,
V7 Sp

signal ratio n(k). If we let ii(k) denote the commission paid based on nor-
malized performance,it follows from (17) that

ai(k) = C'(t) = ai, for all k. (24)

Thus, normalizedcommissions di(k) must all be equal for an agent. This is an

implicationof the assumptionthat attentionto various tasks are perfect sub-
stitutes in the cost function. Since all commissions are equal, the risk cost
from allocating task k to agent i is (r/2)a2n(k). Taskk requiringattentiont(k)
will be optimally assigned to the agent with the lowest price per unit effort.
The risk cost for agent i per unit effort is (r/2)ai2(k) and the value of the
agent's attentionin its best alternativeuse is y,. Therefore,task k is optimally
assigned to agent i if i's total cost of y, + (r/2)2}g(k) per unit effort is less
thanj's correspondingcost. From this observationand (24), it is evidently
optimalto assign the highere(k) tasks to the agent with the lower Tiand to pay
that agent a lower "normalizedcommission rate." This observationis incor-
poratedinto the next proposition,but the proposition'smain purposeis differ-
ent: It verifies that even though the two agents in our model are identical ex
ante, an optimal solution necessarily treats them asymmetrically,requiring
them to specialize in differenttasks.

Proposition 7. Suppose that the informationcoefficients g(k) are not all

identicaland considerthe variantof program(16)-(20) in which the variables
ti (i = 1,2) are added to the list of choice variables. This programhas no
symmetricoptimal solution (ti 7 t2). There is an optimumat which agent 1 is
assigned less strenuous work (TI < t2), takes responsibilityfor the hard-to-
measuretasks [those with Q(k) > L], and receives lower "normalizedcom-
missions" [il(k) < d2(k)].

Proof. First, we show that there is no optimumwith tI = t2 = t. If there

were, then-in view of (16), (17), and (19)-every feasible allocation of
tasks to agents leads to the same total payoff. In particular,thereis an optimal
solution in which agent 1 is assigned the high e(k) tasks; that is, all tasks for
which Q(k)> Q, whereg is set to just exhaustthe attentionti. Agent 2 is then
assigned the remaining[low e(k)] tasks.
Now considerthe family of feasible solutions, parameterizedby e, in which
ft() = f- e, t2(e) = f + e, and all the highest g(k) tasks are assignedto agent
1 until the total attentionrequiredis t (e). In orderfor the symmetricsolution
to be optimal, it is necessarythatthe derivativeof the objective with respectto
e be zero. The following calculation shows that the derivative is negative-
that is, that it would be better to specify that the agent who is assigned the
hard-to-measuretasks work a bit less thanhis counterpart.Indeed, the deriva-
tive of the objective with respect to e at e = 0 is equal to

' i r o2(k)
2C"
2C"~ 22 f0''2tk)
(12(k) l1(k))
- 2(t( k),k)
k
 Principal-AgentAnalyses 49
Multitask

= rC" 2 C I12(k)Q(k)t(k)dk - I I(k) e(k)t(k)dk

< rC"tC' O(t2 - ,) = 0.

L(k)s -<
The last step uses (18) and the facts thato(k) > g when i l(k)= 1 and
when I2(k) = 1 [and that g(k) is not constantso that the inequalityis strict].
The remainderof the propositionis verifiedin the paragraphsprecedingthe
proposition. Q.E.D.

5.2 Caveats
The model presentedin the previous subsectionrepresentsmerely a first pass
at studyingthe optimal groupingof tasks intojobs. Althoughit providessome
interesting insights, we have omitted so many key elements of the problem
and made so many special assumptionsto simplify an alreadycomplex analy-
sis thatit is well to make a preliminarylist of these featuresand omissions and
to speculate about how they may have affected our analysis.
First, we had assumed that all tasks are "small" and that the principalhas
perfect freedom to group them in any way to form a job. Neither of these
assumptionsis particularlyattractive.The assumptionthat all tasks are small
could be replacedby the assumptionthatthereare a finite numberof tasks that
all requiredthe same amountof time [t(k) constant];this, however, introduces
the possibility that i, = t2, in which case all task assignments are equally
good. When tasks require nonnegligible amounts of time and vary in size,
then the need to minimize costs borne by the agents by equalizing workloads
may reversesome of our conclusions. Moreover,tasks like maintainingquali-
ty and producing output cannot always be separated. In short, our model
exaggeratesthe principal'sability to group tasks into homogeneous measure-
ment classes and in so doing caricaturesthe problem of how jobs are con-
structed.The main virtue of our model is that it is structuredso that incentive
considerationsalone determinethe optimalsolution, so thatit lends some new
insights into the very limited question of how incentive concerns may affect
job design.
Second, we had assumed that the errors of measurementin the agent's
various tasks are all independent.We know from previous analyses, such as
Holmstrom(1982), that when errorsare positively correlated,separatingthe
tasks among the two agents allows the use of comparativeperformanceeval-
uation, which can help to reduce the risk premium incurredin providing
incentives. It is not hard to see that even without comparativeperformance
evaluation, separating tasks with positively correlated measurementerrors
creates a betterdiversifiedportfolioof tasks thatreducesthe risk thatthe agent
must bear. Similarly, grouping tasks in which performanceis negatively cor-
related reduces the agents' risk premium. So, even in the incentive domain,
our present model is highly incomplete.
 50 The Journalof Law.Economics,&Organization,V7 Sp

Third, the attentionallocation model that we have used throughoutis itself

a simplification, which forces all activities to be equal substitutes in the
agent's cost function and excludes the possibility that some activities may be
complementary.In our discussion in Section 2 of the issue of how teachers
should be compensated, we found that complementaritiesin the agent's pri-
vate cost of attentioncan have an importanteffect both on how jobs should be
designed and how agents should be compensated,but that complementarities
among the same variables in the productionfunction have no similar effect.
These are subtle distinctionsthatourtheory,in its attentionallocationversion,
cannot address.
Fourth, the models we have studied assume that the agents focus their
attentionon the same tasks for all time. As discussed in Section 2, the model
we are using is explicitly temporal,and issues of job rotationare an important
aspect of real job design. Our preliminaryanalysis shows that these issues
may be susceptible to analysis using an extension of the Section 2 model, in
which the players are uncertainabout the difficultyof productionand use the
past performanceto learn about it. We hope to be able to discuss these issues
more fully in follow-up work.

6. Conclusion
The problem of providing incentives to agents and employees is far more
intricatethan is representedin standardprincipal-agent models. The perfor-
mance measures upon which rewardsare based may aggregatehighly dispa-
rate aspects of performanceinto a single numberand omit other aspects of
performancethat are essential if the firm is to achieve its goals. Commonly,
the principal-agent problemboils down to this: Given a highly incomplete
set of performancemeasuresand a highly complex set of potentialresponses
from the agent, how can the agent be motivatedto act in the social interest?
Our approach emphasizes that incentive problems must be analyzed in
totality;one cannotmake correctinferencesaboutthe properincentives for an
activity by studying the attributesof that activity alone. Moreover,the range
of instrumentsthat can be used to control an agent's performancein one
activity is much wider than just deciding how to pay for performance.One
can also shift ownershipof relatedassets, vary restrictionson the ways a job
can be done, vary limits and incentives for competingactivities, grouprelated
tasks into a single job, and so on.
In a related article (Holmstromand Milgrom, 1991), we study the simul-
taneous use of various instrumentsfor controllingagents to derive new, test-
able results from the theory of organization.Our emphasis there is on how
cross-sectionalvariationsin the parametersthat determinethe optimaldesign
of jobs, the optimal intensity of incentives, and the optimal allocation of
ownershiplead to covariationsamongendogenousvariablesthatare similarto
the patternswe find in actual firms.
Most past models of organizationfocus only on one instrumentat a time for
determiningincentives and a single activity to be motivated.Newer theories,
 Multitask
Principal-AgentAnalyses 51

such as ours, that explicitly recognize connections between instrumentsand

activities, offer new promise to explain the richerpatternsof actual practice.

References
Alchian, Armen, and Harold Demsetz. 1972. "Production,InformationCosts and Economic
Organization,"62 American Economic Review 777-95.
Anderson, Erin. 1985. "The Salesperson as Outside Agent or Employee: A TransactionCost
Analysis," 4 ManagementScience 234-54.
--- , and David Schmittlein. 1984. "Integrationof the Sales Force: An EmpiricalExamina-
tion," 15 Rand Journal of Economics 385-95.
Baker, George. 1989. "Piece Rate Contractsand PerformanceMeasurementError,"manuscript,
GraduateSchool of Business, HarvardUniversity.
, Michael Jensen, and Kevin Murphy. 1988. "Competitionand Incentives: Practice vs
Theory," 43 Journal of Finance 593-616.
Barzel, Yoram. 1982. "MeasurementCosts and the Organizationof Markets,"25 Journalof Law
and Economics 27-48.
Brickley, James, and FrederickDark. 1987. "The Choice of OrganizationalForm:The Case of
Franchising," 18 Journal of Financial Economics 401-20.
Coase, Ronald. 1937. "The Natureof the Firm," 4 Economica 386-405.
Farrell,Joseph, and Carl Shapiro. 1989. "OptimalContractswith Lock-in," 79 AmericanEco-
nomic Review 51-68.
Greenwald, Bruce, and JosephStiglitz. 1986. "Exteralities in Economies with ImperfectInfor-
mation and Incomplete Markets," 101 QuarterlyJournal of Economics 229-64.
Grossman,Sanford, and Oliver Hart. 1986. "The Costs and Benefits of Ownership:A Theory of
Verticaland LateralIntegration,"94 Journal of Political Economy 691-719.
Hannaway,Jane. 1991. "Higher OrderThinking, Job Design and Incentives:An Analysis and
Proposal," American Education Research Journal (forthcoming).
Hart, Oliver, and Bengt Holmstrom. 1987. "The Theory of Contracts," in T. Bewley, ed.,
Advances in Economic Theory, Fifth World Congress. Cambridge:Cambridge University
Press.
Holmstrom, Bengt. 1982. "Moral Hazard in Teams," 13 Bell Journal of Economics 324-40.
- . 1989. "Agency Costs and Innovation,"12 Journalof EconomicBehaviorand Organiza-
tion 305-27.
, and Paul Milgrom. 1987. "Aggregationand Linearityin the Provisionof Intertemporal
Incentives," 55 Econometrica 303-28.
--- , and -- . 1990. "RegulatingTradeAmong Agents," 146 Journal of nstitutionaland
Theoretical Economics 85-105.
, and . 1991. "MeasurementCost and OrganizationTheory," working paper,
Stanford University.
---- , and Joan Ricart i Costa. 1986. "ManagerialIncentives and CapitalManagement,"101
QuarterlyJournal of Economics 835-60.
Itoh, Hideshi. 1989. "Coalitions, Incentives and Risk Sharing,"mimeograph,Kyoto University,
Japan.
. 1991. "Incentives to Help in Multi-AgentSituations," 59 Econometrica611-37.
Krueger, Alan. 1991. "Ownership,Agency and Wages: An Examinationof Franchisingin the
Fast Food Industry,"106 QuarterlyJournal of Economic Theory, forthcoming.
Laffont, Jean-Jacques,and Jean Tirole. 1989. "Provision of Quality and Power of Incentive
Schemes in Regulated Industries,"in W. Barnett, B. Coret, C. d'Aspremont,J. Gabsewicz,
and A. Mas-Colell, eds., EquilibriumTheory and Applications. Cambridge:CambridgeUni-
versity Press.
Lazear, Edward. 1989. "Pay Equalityand IndustrialPolitics," 97 Journal of Political Economy
561-80.
52 The Journalof Law,Economics,&Organization,V7 Sp

MacDonald, Glenn. 1984. "New Directions in the Economic Theory of Agency," 17 Canadian
Journal of Economics 415-40.
Milgrom, Paul. 1988. "EmploymentContracts,Influence Activities and EfficientOrganization
Design," 96 Journal of Political Economy 42-60.
, and JohnRoberts. 1988. "An Economic Approachto InfluenceActivities and Organiza-
tional Responses," 94 AmericanJournal of Sociology S154-79.
Minahan, John. 1988. "ManagerialIncentive Schemes and the Divisional Structureof Firms,"
mimeograph, University of Massachusetts.
Riordan, Michael, and David Sappington. 1987. "Information,Incentives and Organizational
Mode," 102 QuarterlyJournal of Economics 243-64.
Simon, Herbert. 1951. "A FormalTheory of the EmploymentRelationship,"19 Econometrica
293-305.
Tirole, Jean. 1986. "Hierarchiesand Bureaucracies:On the Role of Collusion in Organizations,2
Journal of Law, Economics & Organization 181-214.
Williamson, Oliver. 1985. The Economic Institutions of Capitalism. New York: Free Press.