Production ready
HRM Agent · AI-enabled HR
HR answers to the right person, and the people processes run for you.
An HR assistant for your published people material: the policies, the procedures and the guides your team has approved. It answers at the asker’s authority with the governing document behind every answer, and it runs the people processes you have already written down. Individual pay never leaves the vault, no matter who asks.
What it answers, what it runs, stated plainly
- Answers from your published HR material: policies, procedures and guides your team has approved. Every answer cites the governing document and its version, rendered at the asker’s authority
- Answers on how a process works and what a person is entitled to: leave, overtime, benefits, performance cycles, the discipline path
- Runs the people processes you have written down: onboarding and offboarding runs, probation milestone tracking, periodic HR reporting to the right audience. On request or on a schedule
- Individual pay and personnel files stay blocked at every tier, and nothing leaves the company without a named person approving the draft
- Live connections to your HRIS, payroll or attendance platform: an extended capability, designed and scoped to your request rather than shipped by default
- Account provisioning in your identity platform: a deliberate manual step today, with the request prepared for IT. Automating it is scoped with your security team
HRM Agent
Was this answer helpful?
Thanks. What was wrong with it? (optional)
✅ Thanks, this is recorded and will be reviewed.
Onboarding plan · Alex Carter
Was this answer helpful?
What this AI agent does for your people function
Policy answers scoped by authority
Leave, overtime, benefits, performance cycles, discipline procedure: each answer is rendered at the asker’s access level and cites the governing document and version.
Pay privacy enforced in code
Individual salaries and personnel files are hard-blocked at every tier. There is no configuration that can open them, and a deterministic screen backs up the AI reviewer.
Safe in mixed company
In a group chat the answer follows the least-privileged member present, and nobody learns which member caused the restriction.
Processes that run end to end
Onboarding, offboarding and internal moves: one command turns a multi-step, multi-role process into review cards. On confirmation the record is created, the profile folder prepared, and each role receives one consolidated task email with its own deadlines. Every step is drafted from your documented process and actioned only with your approval.
On a schedule, not only on request
The recurring people work runs by itself: a daily scan of every probation against its statutory milestones, the periodic HR report to its audience, the reminders nobody wants to be responsible for remembering. HR is reminded before the law decides for you.
Local compliance by design
Personal-data protection shapes what the system is even allowed to collect. Business-day arithmetic understands lunar-calendar holidays. Vietnamese and English are both first-class.
Access that mirrors your HR governance
Tiers derive from your records-governance policy, and the boundaries are enforced in the retrieval layer, before any model is involved.
Exec
- Full HR library, including compensation structure
- Performance and employee-relations records
- Reports and analytics
- Restricted archives stay blocked
Manager
- Procedures, recruitment, leave and attendance
- Learning and development, company assets
- No pay data, no personnel files
- No ratings, no disciplinary records
Staff
- Policies and procedures
- Learning and development
- No personal data of anyone
Guest / External
- Polite refusal
- Access request path explained
- No retrieval, no model call
Under the hood
The write path and the chat path are structurally separate, so the part that reads untrusted input can never hold a write credential. Usefulness is watched just as closely: every answer is rated by the person it served, so a wrong refusal becomes a number instead of a silence.
- 829 automated tests, plus a 31-case adversarial compliance evaluation against the live models
- Every answer rated by its reader: the vote lands on the answer’s own audit row, so a block a reader calls unhelpful, the over-refusal every HR assistant risks, becomes a queryable number rather than a silent success
- Three separate app registrations: the Q&A identity is read-only; writes go through a scoped writer locked to one library and one mailbox at the platform level
- Temporary passwords are structurally unleakable: no data structure in the pipeline can carry one to a card, a log or a model
- PII whitelist on every roster read: the system names the exact columns it may read and refuses identity, bank and address fields
- Audit rows are HR records: every question, refusal and onboarding step leaves an immutable row kept without expiry, holding the question keywords and the answer status rather than the wording of either
- Deterministic citation check: a source the answer names must have been retrieved this turn, or the answer is corrected
Three separate app registrations, so the part that reads untrusted input is not the part that can write. Only an approved plan crosses between them.
Common questions
Can it ever reveal someone’s salary?
No. Individual pay is blocked by three independent layers: personnel and payroll folders never enter the model context, a deterministic screen catches pay amounts tied to a person, and an independent reviewer model checks every draft. The refusal itself is logged for audit.
Does it file leave requests for employees?
It answers how leave works, what each person is entitled to, and where to file, always citing the governing procedure, then routes the action to your HR system of record. Deliberately, the request itself stays in the system your approvals already live in.
What exactly does onboarding automate?
The repetitive half: allocating the employee ID, creating the record, preparing the profile folder, and emailing each role one consolidated task list with deadlines computed in business days. Decisions stay human, and account provisioning is intentionally a manual step with a prepared request.
Can it run other people processes, or only onboarding?
Any process you have documented. Onboarding is the one we run in production today, and it is the pattern the others follow: the agent reads your written procedure, drafts each step with the deadlines and owners it computes, and returns an approval card before anything is created or sent. Offboarding, internal moves, probation rounds and periodic HR reporting are configured the same way, against your own procedure rather than a generic template. Each new process is scoped with your HR team and inherits the same tiers, the same guard and the same audit trail.
How is personal data protected?
By collecting less in the first place. The onboarding flow does not collect identity numbers, bank accounts or dependants; roster reads name the exact columns they may touch; small groups are never reported on; and usage analytics store topics, never the wording of questions.
How do you measure whether the answers are any good?
Every answer carries a thumbs up or down from the person who asked, recorded on the same audit row as the guard verdict, and a thumbs down opens an optional note of up to 300 characters. For an HR assistant this matters doubly: refusing is often the correct behaviour, so the only way to tell a right refusal from a wrong one is a reader saying the block was unhelpful. That over-refusal rate is reviewed alongside the security evaluations, and the note itself stays in the agent’s own table rather than in any report.
Next step
See the HRM Agent handle HR questions safely.
Book a 30-minute demo: policy answers, a refusal on restricted data, and a full onboarding run, live in production.
We map your HR records governance to access tiers in the same call.