Not the judgement, not the conversations, not the decisions about people. The task lists, the policy lookups, the four departments who haven't done their part of an onboarding yet.
This section comes first because in HR it matters more than the capability list. Ask any vendor for their version of it.
We build summaries that show a reviewer what an application says and where each point came from. We don't build a number that decides who gets read. Automated rejection carries real legal exposure in the EU, parts of the US and increasingly elsewhere — and it removes exactly the judgement hiring depends on.
Systems that rate people or flag who might leave are built on proxies that encode whatever bias produced the training data. If your organisation wants that, it should be a deliberate governance decision with legal review — not something an integration partner adds because it was technically possible.
Reading the tone of internal messages to gauge morale destroys the trust that made those messages honest. Run a survey people know about instead.
Agents make information retrieval frictionless, which is exactly the risk in HR data. Access follows your existing permission model, always, and we test that before launch rather than after.
Those three properties are the test. Work that fails any of them stays with your team.
Swipe →
Every joiner task fired into the owning system — IT, payroll, access, buddy assignment — then followed up. The one that hasn't moved by day three gets escalated to a named person. It's the laptop. It's always the laptop.
"How much leave do I have left", "can I carry it over", "what's the notice period" — answered from the current handbook, citing the clause, reading the individual's own record where the answer depends on it. Silent where the handbook is silent.
Each application reduced to what it says against the requirements you wrote, with every claim linked to where it appears in the document. A reviewer reads faster. A reviewer still decides.
Each team does its part competently. Nobody owns the gaps. The agent owns the gaps.
A policy agent is only as coherent as the policy behind it. We deployed one for a client whose handbook existed in three versions that disagreed on leave carry-over. By week four the agent was abstaining on nearly half of all questions — correctly, because there was no right answer to give.
We recommended switching it off. The organisation hadn't decided its own policy, and no system resolves a decision that was never made. They spent six weeks consolidating, then we turned it back on and it worked.
So we audit the handbook during the Sprint, before quoting. Occasionally that means telling a people team their real first project is a writing exercise, not an AI one. It's a worse sales conversation and a much better outcome.
The agent can never surface to a person something they couldn't already access in the HRIS. Tested explicitly before launch.
Where policy is silent or contradictory it routes to a person. It will not infer an entitlement your organisation never granted.
Grievances, disciplinary matters, harassment, health, resignation, pay disputes — recognised and handed to a named person, never answered.
Employee records often can't leave a jurisdiction. Retrieval can run entirely inside your network with open-weight models where that's the constraint.
Who asked what, what the agent read, what it answered. Exportable for a works council, a regulator or an internal review.
Nobody discovers by accident that HR answers come from an agent. Every deployment ships with a plain-language note to staff — we help write it.
For factual questions, most prefer it — asking a machine how much leave you have left avoids the small social cost of asking a person for the third time. For anything sensitive, they shouldn't have to trust it, which is why those topics route straight to a human by design rather than by the agent's judgement in the moment.
Employment is a sensitive area under the EU framework, and systems making decisions about recruitment, promotion or termination attract significant obligations. This is a large part of why we build summarisation and orchestration rather than decision-making — it keeps deployments outside the heaviest category. We're an engineering firm, not your counsel, so your legal team should confirm the classification for your specific configuration.
In Germany, the Netherlands and several other jurisdictions, almost certainly. We build the documentation that consultation requires — what data is read, by whom, what is logged, what the system cannot do — as part of delivery rather than as an afterthought when someone asks for it.
Common, and usually workable. Options range from a scheduled export the agent reads, to a thin read-only layer over the database, to working from the systems around the HRIS rather than through it. What matters is whether a reliable read path exists at all — that's a day-two question in the Sprint, and occasionally the answer is no.
Below roughly 200 employees, usually not for policy answers — the volume doesn't justify the integration. Onboarding orchestration can still pay off at lower headcount if you're hiring quickly, because the cost there is a bad first week for each joiner rather than HR's time. We'd rather run the numbers with you than sell you the smaller version.
Five days, no cost. We'll map where onboarding actually stalls, audit whether your policy is coherent enough to automate against, and tell you plainly if it isn't yet.