Aon recently expanded its Radford McLagan Compensation Database to report AI-specific job families — head of AI, applied research scientist, machine learning engineer, AI ethics — alongside a frank observation about why the update was needed in the first place. As the firm put it, organizations are redesigning jobs to keep pace with AI, “roles, skills and expectations are changing in real time,” and the traditional frameworks for defining and valuing work are struggling to keep up. Job scope, Aon’s Radford team noted, is now changing faster than the role definitions and compensation structures that were built to contain it. You can read the announcement here.

The headline is a benchmarking story — new job families, better market data, AI-driven pay premiums to track. But underneath it is a quieter problem that lands squarely on the desk of anyone who runs a bonus plan. A short-term incentive plan is a promise made in January about what a job will look like, what it will be measured on, and what good performance will earn, paid out the following year. That promise assumes the job holds reasonably still for twelve months. When job scope starts moving in real time, that assumption is the first thing to break — and the bonus plan is usually the last thing anyone thinks to check.

Here’s what we think this means for HR and Total Rewards leaders designing variable comp programs for a workforce whose roles are no longer sitting still.

1. Your bonus plan assumes the job stays still for a year. That assumption is now the risk.

Most STIP designs we see treat the job as a fixed object for the plan year. You set the target, you pick the metrics, you map the payout curve — all against a role description that was current the day the plan was approved. In a stable year, that works. In a year where an analyst’s work is half-automated by Q2 and her remaining time is reallocated to model oversight nobody scoped in advance, the plan is now measuring a job that no longer exists.

In our experience, this is the failure mode that hides the longest. The plan still runs. Payouts still process. But the link between what the person actually did and what the plan rewards has quietly snapped, and you usually only find out at calibration, when managers start arguing that the goals “weren’t fair anymore.” They’re often right. The fix is not a better metric. It’s a faster check-in: any role where AI has materially changed the work mid-year needs its STIP goals re-examined before year-end, not relitigated after it.

2. Pay premiums for AI skills are a base-pay story that becomes a bonus-fairness story.

Aon’s data points to real premiums for AI-driven skills, and the natural response is to handle that on the base-pay side — re-benchmark, adjust ranges, move the hot roles up. That’s correct, and it’s where the market data earns its keep. But the second-order effect lands in the bonus plan, and it lands as a fairness problem.

When you pull a cluster of roles up the pay scale because the market moved, you’ve also changed the size of their target bonus, since target incentives are usually a percentage of base. Do that quietly and you’ve handed one team a materially larger bonus opportunity than the team next to it, for reasons that have nothing to do with the performance the plan is supposed to reward. In our experience, employees forgive a pay gap they understand far faster than a bonus gap they don’t. If AI premiums are reshaping your base structure, walk the change all the way through to target incentive — and be ready to say out loud why one role’s opportunity grew and another’s didn’t.

3. Benchmark data tells you what to pay. It doesn’t tell you what to reward.

A 50-year-old compensation database getting sharper on emerging roles is genuinely useful, and we’d never argue against grounding pay decisions in better market data. But market matching and plan design are two different jobs, and the speed of this AI shift makes them easy to conflate. Knowing the market rate for a machine learning engineer tells you where to set the salary. It tells you nothing about what that engineer’s bonus should be measured on.

That second question is the one that actually drives behavior, and it’s entirely yours to answer. The market can’t tell you whether to reward model accuracy, deployment velocity, cost-to-serve, or the governance work that keeps the whole thing defensible. Benchmark data is an input to the pay number. The metric is a design choice about what you want more of. When roles are moving fast, the temptation is to let the survey do both jobs. It can’t.

4. The faster roles move, the more the plan needs a stable spine.

The instinct, when job content is in flux, is to make the plan more elaborate — more individual metrics, more frequent re-scoping, a custom scorecard per role to capture all the change. We’d push the other way. The faster the work moves, the simpler and more durable the plan’s backbone needs to be, or you’ll spend the whole year administering it instead of running it.

In practice that means anchoring the plan on a small number of outcomes that survive a role’s evolution — team or business-unit results an employee can still influence even as their day-to-day tasks shift underneath them — and using a shorter measurement rhythm where the work is genuinely volatile. Two six-month cycles with a calibration in between will track a fast-moving role far better than one annual goal set in January against a job that’s three versions out of date by summer. The point isn’t more measurement. It’s measurement that doesn’t expire mid-year.

5. Defensibility cuts in two directions.

Aon frames the AI challenge for compensation leaders as one where “defensibility matters as much as speed” — you have to move fast on emerging roles while standing behind the decisions under board and regulatory scrutiny. We’d extend that frame one rung down. Defensibility to a regulator and defensibility to the employee are not the same test, and the bonus plan has to pass both.

A plan can be perfectly defensible on paper — clean documentation, market-grounded ranges, a tidy audit trail — and still feel arbitrary to the person whose job changed in March and whose bonus didn’t reflect it. The board-facing version of defensibility is about process and data. The employee-facing version is about whether the person can look at their payout and recognize their own year in it. When AI is rewriting the work that fast, the second test is the harder one, and it’s the one that determines whether the plan still motivates anyone.

What we’d tell HR leaders watching the job map redraw itself

Aon’s update is a useful signal precisely because it’s coming from the benchmarking layer. When the firms whose entire business is measuring jobs tell you the jobs are moving faster than the measurements, that’s worth hearing — and the implication for variable pay is sharper than the base-pay headline suggests. Your bonus plan is the part of the comp system most quietly dependent on the job standing still, and it’s the part nobody re-checks until the complaints start.

So the question for 2026 isn’t whether to track the new AI roles. The market data will handle that. The question is whether your STIP can survive a year in which the jobs it measures don’t hold their shape — whether the targets still mean something in June, whether the fairness logic survives a wave of base-pay adjustments, and whether an employee whose role was rebuilt around them can still see themselves in the payout. If the answer is shaky, the plan needs a redesign before it needs a benchmark.

If you’re looking for tools to simplify how you design, administer, and communicate variable pay — especially when the underlying jobs won’t sit still long enough for a static plan to keep up — let’s talk.