AI in HRFriend
AI does the work. People decide.
Connect Claude or ChatGPT and your assistant works in HRFriend with exactly your permissions. Where HRFriend uses AI itself, it suggests and shows why, and a person decides.
Agents
Your agent works in HRFriend.
HRFriend has an MCP server over a full API. Your assistant signs in as you and can do what you can do in the app: register absence, find out who is free, and approve what waits for you.
Where an agent stops
Exactly your permissions
A token is one member in one workspace, with that member's role and reach. An employee's agent sees their own absence, not the team's.
Read means read only
A Read token cannot change anything. The agent is not even offered the tools that write.
Revoke it, and it stops
Revoke a token or a connected app and the agent stops at its next call. Owners see every token in the workspace and can revoke any of them.
AI drafts
AI drafts. People decide.
Claude reads a meeting's template and its last three meetings and suggests what the next one should cover, each item with its reason. Nothing is on the agenda, in a mail or on anyone's link until the meeting's owner, co-owner or an admin accepts it.
- Off until a workspace owner switches it on in Settings, Privacy
- While it is off, no meeting content is sent to Anthropic
- At most eight suggestions, and never an item already on the agenda
Candidate check
Evidence, not a verdict.
Ask for a check on one candidate, and Claude reads the CV and the application against the position's requirements and your hiring profile. For each requirement you see the evidence, quoted, and what is missing, and then questions for the interview. No score and no ranking: the team's own rating decides.
- Off until an owner switches it on
- Runs only when someone asks, for one candidate at a time
- The candidate is told on the application form
- Every check is logged with the model, the prompt version, who asked and when
Research
Only the professional, always with a source.
Ask for research on one candidate, and HRFriend finds their public professional footprint: the links they gave, GitHub, their own site, articles, talks, press and roles in CVR, the Danish company register. Every finding has its source, and a finding that may be about someone else with the same name is flagged.
- Off until an owner switches it on
- The candidate is told on the application form
- You read the findings and decide what they mean
Context
The context is already there.
An agent is only as good as what it can read. Ask who is free next week, and the answer comes from real data, not from a form someone forgot to fill in.
Already in HRFriend
- People, teams and managers
- Absence and holiday balances
- Capacity and planning by week
- Positions and candidates
- Meetings, agendas and actions
- Rooms, bookings and equipment
Synced from your tools every hour
More about the integrationsWhat AI never does in HRFriend.
HRFriend calls Claude in three places: the agenda draft, the candidate check and research on a candidate. Each is off until an owner switches it on, and no decision about a person is handed to a model.
No score, no ranking, no rejection
The candidate check shows evidence and gaps, never a score or a verdict. It doesn't sort the board, move anyone or reject anyone. The team rates candidates itself, from Strong no to Strong yes.
Nothing private, nothing sensitive
Research looks only at a candidate's public professional footprint. Never private profiles on social media, and never health, religion, politics, union membership, sexuality, ethnicity, family or criminal records.
No predictions, no performance scores
No guess at who will leave and no risk score. The figures in Insights are what GitHub, Linear, Google and Adversus count, shown as they are.
No change without a person
An agenda suggestion waits until someone who runs the meeting accepts it. A check or a research run moves no candidate and sends no mail. HRFriend's own AI changes nothing by itself.
Questions about AI
Which AI assistants work with HRFriend?
Claude and ChatGPT connect with the MCP address and a sign-in. Claude Code, Cursor and other MCP clients that can send a header use a personal token. Scripts use the same token against the API.
What can an agent see and do?
Exactly what the member it acts for can see and do, in one workspace: the same role, reach and rules as in the app. When an agent approves absence, the person gets a mail, just as when you press Approve yourself. A Read token can only read.
Can we switch it off?
Yes. Each member revokes their own tokens and connected apps in Account settings, and an owner can revoke any of them in Settings, API. The agent stops at its next call. Removing a member's sign-in revokes their tokens too. AI drafts, the candidate check and research are each off until an owner switches them on in Settings, Privacy, and off again in the same place.
Where does our data go?
HRFriend is hosted in Germany. Once an owner switches an AI feature on, what is listed above for that feature goes to Anthropic in the USA, which makes Claude, on HRFriend's own key. Nothing else in HRFriend is sent to an AI model. What your own agent reads goes to the assistant you use, under your agreement with its provider.
What is the candidate told?
Once an owner switches on the candidate check or research, the application form says so before the candidate applies: that HRFriend's AI can read the CV and the application against the position, and that you can look up the candidate's public professional footprint.
Do the tests use AI?
No. The result of a personality test or a case is worked out by fixed rules. It shows on the candidate, never sorts the board and rejects no one.
What about the EU AI Act?
AI that helps assess candidates falls under the AI Act's high-risk rules. HRFriend is built for those rules: a person always decides, the candidate is told, every check and every research run is logged, and nothing is on until an owner switches it on. HRFriend is not certified under the Act.