1. Managers need context before they can make good pay decisions

Compensation decisions don’t happen in a vacuum. A manager needs more than a salary number and a merit budget — they need the full picture of the person in front of them.

That means understanding pay history, where someone sits in their range, the available budget, performance, and talent potential. Clare’s point was that comp can become so numbers-heavy that the person disappears from the decision.

Managers are already carrying some of that context in their heads. The better system is one where every manager gets the same relevant information before making a call.

“You need the pay history… you need to know where they sit currently within their range… and I think you need as well that performance and talent potential information.”

The takeaway: don’t just give managers a worksheet. Give them the context required to exercise judgment consistently.

2. Managers should own the decision — but not pretend they control the entire system

One of the hardest parts of compensation is getting the manager’s role right.

Managers are closest to the work, so they should have meaningful ownership over the recommendation. But the final outcome also sits inside a much larger system: budgets, compensation rules, job levels, country requirements, calibration, and company performance.

“The manager owns the decision, but they don’t own that envelope.”

A better conversation is transparent: “Here are the constraints I’m working within. Here’s the judgment I made inside those constraints. And here’s what I want to see from you going forward.”

That clarity makes the manager more credible, not less.

3. Build HR technology around the manager’s job — not the complexity of the system

Managers don’t need to see every piece of complexity sitting underneath a compensation process.

They need the few things that matter most, presented clearly enough that they can make a decision quickly — with the ability to drill deeper when necessary.

Sébastien’s advice starts with understanding the user. An HRBP and a manager are different personas with different needs, so forcing both through the same experience creates unnecessary friction.

“You need to make sure that the managers are willing to enter the system, and they don’t see it as an additional burden.”

The design principle: surface the three or four things that matter most, flag exceptions, and hide complexity without hiding information.

4. AI should be the autopilot, not the pilot

When it comes to compensation, Sébastien used a useful metaphor: think about AI like the autopilot on a plane.

You want it doing a huge amount of work behind the scenes. You want it helping manage complexity, surface recommendations, and make the operator more effective.

But you still want a human in control when the stakes are high.

“You need all the electronics… but you would still like to have a pilot in control.”

That’s especially important in performance and pay decisions because the outcomes directly affect people’s careers, rewards, and experience at work.

The practical test for HR teams: if the AI makes a recommendation a manager disagrees with, can the manager understand it, challenge it, and override it?

If not, the system has too much authority.

5. Responsible AI needs to be explainable, collaborative, and controllable

beqom’s framework for what it calls “intentional AI” comes down to three requirements:

- Explainable: You can trace how a recommendation or decision was reached.

- Collaborative: A human stays in the loop and has the final say.

- Controllable: You decide which AI capabilities are used, where, and for what.

“If you miss one, you don’t have intentional AI.”

That matters because each principle protects against a different failure mode. An explainable system that nobody can control still creates risk. A collaborative system that nobody can explain asks managers to approve something they can’t inspect.

For people leaders evaluating AI tools, these three questions are a useful starting point: Can we explain it? Can a human intervene? Can we control where it operates?

6. Don’t use the same kind of AI for every problem

One of the most practical parts of the conversation was Sébastien’s breakdown of three different types of AI — and the jobs each is best suited for.

LLMs are strong at language: explaining policy, summarizing information, and helping people understand complex concepts.

Statistical models are better suited for analyzing variables and generating recommendations — but they need supervision because they can learn from historical patterns and biases.

Rule-based systems handle the less glamorous but highly auditable work: eligibility rules, floors, country requirements, and formulas.

“LLMs are built to be plausible, but comp needs to be correct.”

The mistake is asking one type of AI to do everything.

Instead, stack the right tools together: deterministic rules where precision matters, statistical models where prediction helps, and LLMs where language and guidance add value.

That architecture gives HR the benefits of AI without turning critical compensation decisions into a black box.

‍