HR work is checklist work across systems that do not talk to each other. Onboarding touches identity, payroll, benefits, equipment, and the team calendar. Offboarding runs the same list in reverse, under time pressure, with security consequences when a step is missed. Meanwhile the same twenty policy questions arrive on Slack every week, and leave balances live in a spreadsheet one person truly understands.
The failure mode is rarely dramatic. It is a leaver whose accounts stayed live for three weeks because a deprovisioning ticket sat in a queue. It is a new hire spending day one waiting for logins. It is a benefits answer that varies depending on who happened to reply. None of this is hard work. All of it is unforgiving of dropped steps.
Runbooks become durable workflows. Each step in each system is a node; the run parks at approval gates and survives restarts, so a process that spans a notice period does not depend on anyone keeping a tab open. Every action lands in an append-only audit log. When security asks whether a leaver’s access was removed, the answer is a link to the evidence, not a recollection.
The load-bearing trust point for HR is data handling. Employee records are personal data by definition, so guardrails redact PII before anything reaches a model, and guardrail events are hash-chained. The policy agent on Slack can only search knowledge bases it was explicitly granted: the search_kb tool’s scope is fixed server-side, so a question about parental leave has no path into salary data or grievance files. Agents reach your HR systems through the credential vault — short-lived, scoped references, never raw secrets.
Operational records move out of spreadsheets into Data Tables: typed columns, views, and rows that agents and workflows read and write directly, with no export-import loop.
Every access grant and every revocation is approved by a person before it executes. Managers decide leave requests; the workflow only makes the decision cheap to make. Policy answers the knowledge base cannot ground go to a person in the Conversations inbox rather than to a guess. And the conversations that define HR — performance, grievances, exits — are not agent work at all. The platform’s job is to make the mechanical parts verifiable so your team spends judgment where it matters.
Start with one workflow. The free tier is enough to run a real one — agents propose, your team decides.