FixControl
HR questions

HR questions answered with AI, under the permission model HR needs

An employee of your organization asks about leave, a payslip line or a procedure. FixControl resolves who is asking, which sources that person may be answered from, and how sensitive the question is. A routine answer can be delivered straight away; a sensitive one becomes a case that only an explicitly assigned reviewer can open, and that a second person has to approve before it goes out.

Book a demoTalk to us
Sensitive cases only for an assigned reviewerSecond person approves before deliveryAFAS Profit and Visma Raet Youforce read-only
The demo

Watch the governed HR flow

From an HR question in Teams to a delivered answer: who may see it, what the HR system is allowed to return, and the second person who has to approve.

This video is spoken in Dutch, with Dutch captions.Dutch captions available — turn them on in the player if you want them.
Read the video transcript (in Dutch)

Een HR-vraag komt binnen

Een medewerker vraagt in Microsoft Teams hoeveel ouderschapsverlof zij nog heeft.

Eerst: wie mag dit zien?

FixControl bepaalt eerst wie dit mag zien: de medewerker en de behandelaar, niet het hele team.

Alleen-lezen, en eerlijk als er geen cijfer is

FixControl leest het saldo alleen-lezen uit het HR-systeem: alleen het veld dat nodig is. Is er geen cijfer, dan verzint FixControl niets — een mens neemt het over.

Een gevoelige vraag gaat een andere route

Een gevoelige vraag gaat naar een bevoegde beoordelaar. Wie toewijst, ziet alleen kenmerken — niet de inhoud.

Vier-ogen vóór levering

Een gevoelig antwoord gaat pas de deur uit na goedkeuring door een tweede persoon. Die goedkeuring is 72 uur geldig.

Vastgelegd en herstelbaar

Elke stap is vastgelegd. Klopt iets niet, dan volgt een gepubliceerde correctie.

Sneller, met controle

Zo krijgen medewerkers sneller antwoord, terwijl HR de controle houdt.

Questions HR people ask

What an employee and an HR manager want to know

Who sees an employee’s question?+
That employee does, always. For a normal case, the HR colleagues on your team whose project role covers that work see it too. For a sensitive case the circle is smaller: only the reviewer the case was explicitly assigned to can read it. Someone with the permission to assign sees the case exists — status, sensitivity, route — without its content.
What happens to a sensitive question?+
It does not get a direct answer. It becomes a case at a higher sensitivity level, and that level determines which permission a reviewer needs before the case can even be handed to them. Medical, legal and confidential employee-relations questions always take that route.
What does the AI see of an employee’s personnel file?+
The question text, and nothing else. That is what the classification proposal is given, and it is kept fully separate from your organization’s own data. The employee’s file, retrieved documents and the answer body are never sent to a model. When personal data is needed for an answer, the server reads named fields directly from the HR system.
Can an answer just go out to the employee?+
A general, current, non-contradicting policy answer above the confidence threshold can be delivered directly. Anything touching that employee’s own situation becomes a draft or a review case. In that case an approval is required, it is bound to that exact answer text, the person who requested it cannot approve it, and it expires after 72 hours on the server.
What if the answer turns out to be wrong?+
The employee can report it as incorrect, and so can your HR staff within what they may already read. The answer goes under review and a correction task is created. A correction becomes a new version with a link to the original, which is kept and marked superseded. If there is no safe replacement, the answer is withdrawn instead of quietly rewritten.
What do the connections to the HR systems read?+
AFAS Profit and Visma Raet Youforce are connected read-only and write nothing back. Each call filters on one person and projects the response onto a fixed set of fields. If a value is missing or unreadable, no number is invented and no leave balance is defaulted to zero — the question is handed to a person instead.
Is what happens recorded?+
Yes, on identifiers. The record holds project scope, case and answer identifiers, classification, the source versions used, the policy decision and its reason code. Question text, summaries, medical content, employee references and answer bodies are deliberately not in it, so the trail can be reviewed without reopening the case content.
Through which channels can an employee ask?+
Microsoft Teams in a personal chat, Slack in a direct message with the bot, or your own existing chatbot through an authenticated API. The API key is bound to one organization, the chatbot may state who is asking and nothing more, and the same permission checks then run for that employee.
How it works

From a question to an answer someone is accountable for

Every step is a server-side check, not a setting in the chat window. The channel the question arrives on carries no permissions of its own: Teams, Slack and your chatbot all call the same command, and that command decides.

  1. 01The question
    An employee of your organization asks in Teams, Slack or your own chatbot

    In Microsoft Teams the question is only handled in a personal chat, in Slack only in a direct message with the bot. A shared channel gets a short explanation and nothing is processed — the question text does not even reach the model. Your own chatbot can ask on an employee's behalf through an authenticated API.

  2. 02Who is asking
    Identity is resolved before anything else happens

    The chat account has to be linked to a FixControl user, and that user needs an active membership of an HR-enabled project with permission to ask. An unlinked account gets a link instruction instead of an answer. A missing identity can only make the check stricter, never looser.

  3. 03Sensitivity
    The question gets a sensitivity level it cannot lose

    A deterministic classification sets a safety floor. A model may propose a stricter reading on top of it — more sensitive, or human review required — but it can never argue a question down to a lighter route.

  4. 04Minimisation
    The model is shown the question text, not the file

    The classification proposal receives the question text and nothing else, fully separated from your organization’s own data. Personnel-file data, retrieved documents and answer bodies are never sent to a model; the answer itself is drafted from templates and server-selected sources.

  5. 05Assignment
    A sensitive case waits for an explicitly assigned reviewer

    Someone with the assign permission sees a restricted case as metadata only — status, sensitivity, route — and can hand it only to an active project member whose role carries the permission that sensitivity requires. Assignment is the bridge between a case existing and anyone being allowed to read it.

  6. 06Approval
    A second person approves, then the answer is delivered

    An approval is bound to the exact answer text and version. Whoever requested it can never decide it, the approval expires on a server-side 72-hour window clients cannot stretch, editing the answer invalidates it, and delivery consumes it in a single step so a double click cannot send twice.

What you get

The parts an HR pilot usually discovers late

Permissions, minimisation, approval and correction are the product here, not a phase after the answers work.

A visibility contract, not a shared queue

The employee who asked sees their own question. Your own HR staff see the work their project role allows. A restricted case is readable only by the reviewer it was explicitly assigned to — tenant administration is not HR access, and a technical administrator sees empty lists.

Three sensitivity levels with different routes

Normal, sensitive and highly restricted. Medical, legal and confidential employee-relations content takes the restricted human route. Missing, expired or contradicting sources, an unavailable connector, low confidence or an unclear question never produce an invented answer.

Named fields, not a personnel file

The connectors read a fixed, explicitly projected set of fields — employment percentage and type, collective agreement, remaining leave hours with an as-of date, a payroll explanation code. A widened view on the HR side cannot leak extra data into an answer.

AFAS Profit and Visma Raet Youforce, read-only

Both connectors read and write nothing back. Credentials are per organization, a project has to activate a connector explicitly, and a connection check makes a revoked token visible before an employee runs into it.

Report, correct, or withdraw

The employee who asked — and your HR staff within what they may already read — can mark an answer as incorrect. A correction is published as a new version while the original is kept and marked superseded. When there is no safe replacement, the answer is withdrawn instead.

An audit trail on identifiers, not on content

Decisions record project scope, case and answer identifiers, classification, source versions, the policy decision and its reason code. Question text, case summaries, medical content, employee references and answer bodies stay out of the audit record and inside the HR tables.

Build or buy

The chatbot is the easy half

Where an internal HR pilot usually stalls
Many organisations start their own AI pilot for HR questions. The chatbot is there quickly; then the real work begins: who may see which question, what may a model see of a personnel file, who approves a sensitive answer, and how do you show that afterwards? That work is rarely underestimated on purpose — it lands next to everyone's day job, and it needs decisions from HR, privacy, security and IT at the same time. At FixControl that governance foundation is the product, not the finishing touch. You bring the policies, the HR systems and the people who may approve; the permission model, the approval flow and the audit trail are already there.
What the model is allowed to do

AI proposes a reading; it decides nothing

The boundary around the model
The model has no database access and no free HTTP tool. It cannot resolve identity, cannot grant membership, cannot lower a sensitivity level, cannot choose which sources an answer rests on, cannot reach a connector, cannot evaluate the policy, and cannot approve or deliver an answer. Source references come only from rows the server selected. If the model provider is unavailable, the deterministic route is used and the question keeps moving.
Keep exploring
AI approval ledger

Who approved what, from which channel, on which grounds.

Microsoft Teams integration

Questions and approvals in Teams, in a personal chat.

Slack integration

Questions in a direct message, approvals in the thread.

See the permission model before the chatbot

Book a demo and follow one sensitive HR question through assignment, approval, delivery and correction.

Book a demoTalk to us