If AI can measure everything, what should we actually reward?

For the past few years, companies have treated AI as something employees and executives need to adopt. That made sense while the technology was new: people had to experiment with it, processes had to be redesigned and organisations wanted evidence that management was taking the opportunity seriously.

Some employers have even put money behind that objective. The law firm Shoosmiths, for example, set a collective target of one million Microsoft Copilot prompts, with achievement making an additional £1 million available to its firmwide bonus pool, subject to the firm’s financial performance. The target was reached four months early. Its next initiative moved on from simple usage towards AI fluency and accreditation.

As a short-term way of breaking inertia, that is perfectly intelligible. But it also illustrates how quickly an adoption metric can date.

Once AI becomes part of the normal operating environment, using it stops being exceptional performance and becomes part of the job. Nobody receives a bonus for using Excel effectively and we do not usually create an executive KPI for using the company’s accounting software. AI will increasingly move into the same category.

Indeed, the legal industry is already encountering the next problem. As usage scales, firms are beginning to scrutinise the economics of AI consumption (tokenmaxxing anyone?), including whether expensive models are being used for relatively low-value tasks. I predict a future of choosing from a shopping list of small (not large) language models to achieve different tasks more cost effectively. 

There is a useful incentive-design lesson in that transition. A metric that makes sense when adoption is scarce can become counterproductive once usage is widespread and the constraint shifts towards cost and value.

But there is a second, potentially more important, consequence of AI. As it becomes less suitable as something to reward in its own right, it is becoming much more powerful as a way of measuring the people we reward.

That creates a paradox: we should probably reward AI use less, even as AI allows us to measure - and potentially reward - much more.

From AI adoption to algorithmic management

For decades, one constraint on performance-related pay was that employers could not observe enough. AI and algorithmic management may create the opposite problem.

This is already less hypothetical than it sounds. An OECD survey published in 2025 found that 90% of US managers said their firms used at least one algorithmic tool to instruct, monitor or evaluate workers. Across France, Germany, Italy and Spain, the average was 79%. The OECD definition is broader than AI - it covers software used to automate or support managerial tasks - but that is precisely the point. Algorithmic management is already becoming part of ordinary management infrastructure.

Employers can increasingly observe response times, client interactions, workflow, output, conversion rates, utilisation, collaboration patterns and potentially thousands of other signals about individual performance. At first sight, that looks like an incentive designer’s dream. Better information ought to allow performance to be rewarded more accurately.

The difficulty is that the moment a measure affects someone’s pay, it stops merely observing behaviour and starts changing it. Put response time into a bonus and people will respond faster. Reward client contact and recorded client interactions will increase. Pay for utilisation and people will optimise hours.

None of that makes the measures necessarily wrong. Incentives are supposed to change behaviour (a topic I have written about before, click here). The real question is whether improving the metric improves the business.

We have already started to automate parts of management. The next question is what happens when we automate parts of reward.

Prediction is not causation

This is where AI creates a particular danger, because AI can be extremely good at finding variables that predict performance, while incentive design needs something subtly different: variables that help cause performance.

Suppose an algorithm discovers that the most profitable employees respond to clients faster, collaborate with a wider range of colleagues and have more frequent client contact. That may be useful information, but it does not follow that everyone will become more valuable if they answer emails faster, attend more internal meetings or record more client conversations.

Successful people may exhibit those behaviours because they are successful. The distinction is therefore between:

prediction → explanation → incentive

AI may become extremely good at the first, but that does not mean it has established the second. Without the second, moving to the third can be dangerous.

A variable that predicts who performs well is not necessarily a variable we should pay people to maximise. The risk is not simply that employees game the target. It is that the organisation takes a sophisticated correlation and turns it into a crude instruction.

Observing work is not the same as attributing value

There is a further complication.

Imagine someone doubles their output after being given access to a powerful AI system. Some of the additional value may reflect that individual's judgement or ability to handle greater capacity, but some comes from capital supplied by the organisation: the model, proprietary data, workflow design, technology infrastructure and accumulated knowledge of the business.

Performance increasingly becomes a joint product of individual judgement and organisational capital. That weakens the assumption that measurable individual output is necessarily a good basis for allocating individual reward.

This creates an important distinction: AI may solve the problem of observing work without solving the much harder problem of attributing value.

Indeed, as the technology becomes more capable, the scarce human contribution may increasingly shift towards things that remain difficult to count: choosing the right question, challenging the model, exercising judgement, taking responsibility and building trust.

Those qualities may become more valuable precisely as everything around them becomes easier to measure.

Can the algorithm explain the bonus?

There is also a more basic incentive problem. An algorithm might assess an employee using thousands of datapoints, peer comparisons and statistical relationships and the resulting score could be highly predictive. But an incentive is not simply a mechanism for allocating money; its purpose is also to affect future behaviour.

For that to work, there has to be an intelligible relationship between what someone does and what they receive. A system can therefore be very good at pricing performance while being quite poor at incentivising it.

This is no longer just theoretical. The Government is currently consulting on the fair and transparent use of workplace monitoring technologies. Current ICO guidance, which is under review following recent data-law changes, also expressly contemplates automated monitoring being used to assess performance and determine pay.

Explainability is therefore not simply an AI governance or data-protection issue, it is part of the incentive architecture.

Five questions before a metric becomes a target

Better data should improve reward decisions. But before converting a new measure into an incentive target, I would ask five questions:

1. Is it causally connected with value or merely correlated with successful employees?

2. Can the individual materially influence it?

3. Does improving it actually create economic value?

4. What will people rationally do once money depends on it?

   And perhaps the best test of all:

5. Would we still want this behaviour if nobody were measuring it?

That last question helps distinguish genuine performance from performance of the measurement system.

Better measurement may mean simpler incentives

The natural instinct will be to use AI and richer data to build increasingly precise incentive systems. I fear that may be exactly the wrong conclusion.

As the cost of measurement falls, good remuneration design may need to become more selective rather than more granular: fewer metrics, stronger causal logic, greater emphasis on economic outcomes and more willingness to preserve judgement where individual contribution cannot sensibly be reduced to a datapoint. Although measuring judgement won't exactly be easy either (you can read more for my thoughts on that here).

The move from adoption targets towards scrutiny of AI consumption captures the wider point. When adoption is scarce, encouraging usage can make sense. When usage becomes ubiquitous and carries a marginal cost, the objective changes. The same should be true of performance data: just because something has become measurable does not mean it has become worth rewarding.

For decades, incentive design has been constrained by scarcity of information. The next problem may be abundance.

AI may make performance measurement dramatically more granular. Good remuneration design may need to become less so.

At Burges Salmon

AI is beginning to change not only how businesses operate, but how performance itself can be defined, measured and rewarded. We are working with companies on the wider performance and reward architecture around emerging technologies - including when a new incentive measure is genuinely needed, when existing financial measures should capture the value and how to avoid turning better data into worse incentives.