Bengt Holmström Prize Lecture: Pay for Performance and Beyond

---

413

Pay For Performance and Beyond

Prize Lecture, December 8, 2016

by Bengt Holmström

*

Massachusetts Institute of Technology, Cambridge, MA, USA.

I

n this lecture, I will talk about my work on incentive contracts, especially incen-

tives related to moral hazard. I will provide a narrative of my intellectual jour-

ney from models narrowly focused on pay for performance to models that see 

the scope of the incentive problem in much broader terms featuring multitask-

ing employees and firms making extensive use of nonfinancial instruments in 

designing coherent incentive systems.

I will highlight the key moments of this journey, including misunderstand-

ings as well as new insights. The former are often precursors to the latter. In the 

process, I hope to convey a sense of how I work with models. There is no one 

right way about theorizing, but I believe it is important to develop a consistent 

style with which one is comfortable.

I begin with a brief account of the roundabout way in which I became 

an economist. It reveals the origins of my interest in incentive problems and 

accounts for my life long association with business practice, something that has 

strongly influenced my research and style of work.

*

 I want to thank George Baker, Robert Gibbons, Oliver Hart, Paul Milgrom, Canice 

Prendergast, John Roberts and Jean Tirole for years of fruitful discussions on the topic of 

this lecture and Jonathan Day, Dale Delitis, Robert Gibbons, Gary Gorton, Parag Pathak,  

Alp Simsek, David Warsh and especially Iván Werning for comments on various versions 

of the paper.

---

414 

The Nobel Prizes

I did not plan to become an academic. After graduating from the University 

of Helsinki, I got a job with Ahlstrom as a corporate planner. Ahlstrom was one 

of the ten biggest companies in Finland at the time, a large conglomerate with 

20–30 factories around the country. The company took pride in staying abreast of 

recent advances in management. I was hired to implement a linear programming 

model that had recently been built to help management with long-term strategic 

planning. It was a huge model with thousands of variables and hundreds of con-

straints, describing all the factories within the conglomerate, their production 

activities, investment plans and financial and technological interdependencies. 

The company had invested heavily in mainframe computers and had high hopes 

that the planning model was going to be one of the payoffs from such investments.

The most urgent task for me was to arrange the data collection process. I 

began visiting factories to discuss the data requirements. It did not take me very 

long to see that the people providing the data were deeply suspicious of a 23-year-

old mathematician sent from headquarters to collect data for a planning model 

that would advise top management on how to spend scarce resources for invest-

ments. They wanted to know what numbers to feed into my “black box” model 

to ensure that their own plans for their factory would receive the appropriate 

amount of resources. Data were forthcoming slowly and hesitantly.

After some months, I came to the conclusion that the whole enterprise was 

misguided. Even assuming the best of intentions at the factory level, reasonable 

data were going to be very difficult to obtain. The quality of the data varied a lot 

and disagreements about them often surfaced, especially in cases where facto-

ries were interconnected. Also, I grew increasingly concerned about gaming. 

The integrity of the data therefore seemed questionable for technical as well as 

strategic reasons.

I suggested that we give up on the grand project and that I instead would 

focus on two things: (i) smaller models that could help each factory improve its 

own planning process and (ii) try to deal with incentive problems in investment 

planning at the corporate level.

My first recommendation met with some success. I worked out small linear 

programs for factories that seemed amenable to such models. I used these models 

as an economist would: I tried to replicate what the factories were doing. This 

involved a lot of back-and-forth. I would get the data, run the linear program 

and then go tell the factory what the program was proposing as the “optimal 

solution” given their data and, importantly, why the model was proposing the 

solution it did. This last step—explaining how the model was thinking—was the 

key. It made clear that I was not there to recommend a mechanical solution; I 

was there to understand what may be missing in my model specification. This 

---

Pay For Performance and Beyond 

415

lesson, on the usefulness of small models and of listening to them, would stick 

with me throughout my academic career.

My second recommendation to think creatively about the incentive problems 

surrounding investment planning was a failure. I made all the mistakes one is 

likely to make when one tries to design incentives for the first time. My thinking 

was guided by two principles: factories should pay for their borrowing and the 

price should be obtained at least partly through a market-like process so that 

funds would be allocated efficiently. In today’s language, I was suggesting that 

we “bring the market inside the firm” in order to allocate funds to the factories.

Today, I know better. As I will try to explain, one of the main lessons from 

working on incentive problems for 25 years is, that within firms, high-powered 

financial incentives can be very dysfunctional and attempts to bring the market 

inside the firm are generally misguided. Typically, it is best to avoid high-pow-

ered incentives and sometimes not use pay-for-performance at all. The recent 

scandal at Wells Fargo explains the reason (Tayan 2016). Monetary incentives 

were powerful but misaligned and led some managers to sell phony accounts to 

enhance their bonuses. Kerr’s (1975) famous article “The Folly of Hoping for A, 

While Paying for B” could have served as a warning, though the article is rather 

thin on suggestions for providing alternative ways to provide incentives. I hope 

to show that our understanding of incentive problems has advanced quite a bit 

since the days Kerr published his article. In order to appreciate the progress of 

thought, I will start with the early literature on principal-agent models.

I. THE PRINCIPAL-AGENT PROBLEM

A. The One-Dimensional “Effort” Model

Early contributions to the principal-agent literature on moral hazard include 

Wilson (1969), Spence and Zeckhauser (1971), Ross (1973), Stiglitz (1975) and 

Mirrlees ([1975]1999). Wilson and Ross asked under what conditions the prin-

cipal’s and the agent’s preferences over risky lotteries will be perfectly aligned 

when sharing risk optimally. This is possible if the principal’s and the agent’s 

utility functions are such that linear risk sharing is optimal. Spence and Zeck-

hauser (1971) studied insurance contracts under varying information assump-

tions including the case of moral hazard as well as adverse selection.

Mirrlees was the first to study in generality the case where the agent provides 

a service to the principal and the issue is how to motivate the agent to work dili-

gently. The agent’s action is often referred to as “effort” though this interpretation 

should not be taken literally. The model applies in a wide range of situations: an 

---