428
The Nobel Prizes
the commission rate on x
2
will be α
2
< p
2
, trading risk against incentives for a
second-best outcome.
Now, assume that the tasks are substitutes: the more time the agent spends
on one task, the higher is his marginal cost of spending time on the other task
(the cross partial C
12
> 0.) In this case it is no longer optimal to give the agent
first-best incentives for choosing quantity (α
1
= p
1
), because lowering α
1
slightly
will at the margin cost nothing in lost
value on the first task, but will be strictly
beneficial for providing incentives on the second task. The agent can be given
the incentive to supply the same amount of quality for a lower α
2
, which reduces
the cost of risk. Or alternatively, when α
1
is lower the marginal cost of spending
time on quality has gone down, making the agent spend more time on that task.
Either way, lowering the incentive on quantity is advantageous for the supply
of quality, by how much depends on the precision with which quality can be
measured and how substitutable the tasks are in the cost function. If the two
tasks are perfect substitutes (i.e., the agent just allocates his time between the
two tasks so the cost function is C(e
1
+ e
2
)) and attention to quality is essential
(this will require a nonlinear benefit function B), any incentive on quantity will
have to be matched by a correspondingly strong incentive on quality so that α
1
= α
2
. This makes the agent indifferent between spending time on either task and
he will choose whatever allocation is best for the principal. The cost is that the
incentive for both tasks will have to be reduced because quality is poorly mea-
sured. If there is no quality measure at all (i.e., the variance of ε
2
is infinite) then
no incentive for either task is optimal. If C′(0) < 0, the agent will still choose a
positive level of total effort.
B. Misalignment and Manipulation
Misalignment is an important variation on the multitask theme (Baker 1992,
2002). It can also be analyzed with the general multitask model.
Suppose again that the principal’s value is B(e
1
,e
2
) = p
1
e
1
+ p
2
e
2
, and there
is just one performance measure x(e
1
,e
2
) = g
1
e
1
+ g
2
e
2
. Even though there is no
uncertainty in the measure, there is a nontrivial incentive problem if the vec-
tors p and g are not aligned. For example, suppose x = e
1
+ e
2
but B(e) = e
1
.
The principal only values effort on the main task 1, but the agent can produce
measured output x with e
1
as well as e
2
. The second activity can be interpreted
as “manipulation,” which the agent may feel a moral dislike for, but if the prin-
cipal pushes too strongly on measured performance x in the hope of getting the
agent to work hard on the main task, the agent will engage in manipulation as
well. The principal ends up compensating the agent for worthless performance,
Pay For Performance and Beyond
429
so
misalignment causes waste, more so the higher α is. It will be optimal to set
α < 1, because at α = 1 the marginal cost of reducing α is strictly negative (e
1
is
first-best while e
2
is not).
Both variants of multitasking, disparities in measurement errors as well as
misalignment between performance measures and the value created, are relevant
variants of the general multitask model. It depends on the context which one is
more natural to use.
C. “You Get What You Pay For”
Gross manipulation of performance measures was behind the recent Wells Fargo
scandal (Tayan 2016). Wells Fargo had avoided the banking scandals associated
with the financial crisis. It was known for prudence in lending, making profits
by emphasizing its retail banking and customer service. Its branch managers
were highly incentivized toward cross-selling: getting its regular banking cus-
tomers to buy a range of products, such as credit lines. This part of their banking
business had steadily grown and been highly profitable. But continued growth
also required new customers and eventually, as the sales goals were tightened
(the branch managers’ performance was measured daily), some of the managers
opened new accounts for its customers without the customers’ knowledge. Shell
accounts were like a second activity in the two-task model described earlier. It
improved measured performance and generated bonuses, but since there was no
real activity in the accounts this activity generated minuscule profits for Wells
Fargo ($2.6 million according to Tayan 2016). Shell accounts caused minimal
costs for customers (an estimated $2.50 per account), but they were of course
illegal. Eventually the scam was discovered, causing the firing or resignation of
thousands of employees and eventually also the resignation of Mr. Stumpf, the
CEO. Wells Fargo had to pay $185 million in penalties, but the biggest cost by
far was the enormous damage to their stellar reputation.
16
The explanations for the BP oil spill in the Gulf of Mexico point in many
directions, but the fact that BPX—the exploration arm of BP—was encouraged
to be more aggressive in its exploration activity was likely one of the culprits
(Garicano and Rayo 2016). Measurable results were pushed hard (implicitly or
explicitly), compromising safety. This is an example where there are many activi-
ties, some easy to measure (successful exploration), and others not so easy. While
safety can be monitored, the intensity of monitoring whether rules are being
strictly followed is not easy. Slight delays in service or in repair of faulty parts,
especially in the case of backup systems and checks, may appear to carry minimal
risk and therefore be subject to trade-offs under pressure.