Skip to content

Connecting HiBob, Rippling and Personio to Claude and ChatGPT safely

5 min readFor people operations, HR systems owners and security reviewers connecting HR data to AI assistants

Connect HiBob, Rippling and Personio with user-level access: users sign in with their own HiBob, Rippling or Personio account (or your SSO), and the assistant inherits their existing HR permissions, so a manager sees only her own team. Then narrow the fields your questions need, and agree what headcount, leaver and absence mean before anyone compares numbers.

HR questions are easy to ask and easy to get wrong

People teams field the same questions every month: headcount by site at month end, who left and why, how much vacation is still unused before year end. In HiBob, Rippling and Personio the data is there and the questions are short. That makes them good candidates for an assistant such as Claude or ChatGPT, reached through an MCP server.

They are also among the most sensitive questions in a company. An HR record carries compensation, health-related absence, termination reasons and home addresses. A connection that is fine for headcount can expose much more if nobody narrowed it. Most of the work is deciding what the connection may read before anyone types a question.

Whose access an answer uses

All three systems already decide who sees which employees and which fields. A connection that keeps user-level access lets those rules carry through: users sign in with their own account (or your SSO), and an answer covers only the people and fields that user is already authorized to see.

SystemAPIWhat decides each user’s access
HiBobThe Bob public APIThe user’s permission groups in Bob: which people and which field categories they can see
RipplingThe Rippling REST APIThe user’s Rippling role and permissions
PersonioThe Personio APIThe user’s Personio role, including which employee attributes it can read

Permissions follow the user

With user-level access, a manager who sees only her own team inside HiBob sees only her own team in an answer too. An HR business partner sees the populations they support, and an analyst without compensation access gets no salary figures. Nexatron inherits these existing permissions, so the HR system stays the authority.

Be wary of any description that says an HR connector respects HR permissions without explaining how. Ask whose account each query runs under, and test it with two users whose access differs. Both should get different results, or a refusal, for the same question.

Narrow the fields first

Start from the questions and work back to the fields. Most headcount, attrition and absence questions need far less than the full employee record.

  • HeadcountStatus, start and termination dates, department, site or office, employment type, and the effective-dated work history that places a person in a department on a given date.
  • AttritionTermination date and reason, plus whatever field marks a departure as regretted. Manager history only if the questions ask about it.
  • AbsenceAbsence type, dates, status, and working hours for full-time equivalents. Leave notes and medical details are rarely needed and are worth excluding.
  • Leave outCompensation, bank details, national identifiers, home addresses and free-text notes, unless a specific approved question needs one of them.

Apply it in each system

In HiBob, review the permission groups that cover sensitive field categories. In Personio, check which roles can read which employee attributes. In Rippling, check which roles can see compensation and leave details. Write the choices down with the reason for each field, so the next reviewer can tell a deliberate gap from an oversight.

Then test with a question that should fail for a user without compensation access, such as average salary by department, and confirm it does. A refusal that names the missing field is the result you want. A plausible-looking number built from whatever happened to be readable is the result to fix.

Agree the definitions before comparing numbers

Two correct answers to how many people work here can differ by dozens. One counts contractors and one does not. One counts people on long-term leave. One reads the current profile while the other reads work history as of month end. When the assistant’s number and the HR report disagree, the cause is almost always a definition, and people lose trust in the assistant regardless.

Write down headcount, leaver, voluntary, regretted and absence day once, with an owner in HR named, and make the assistant’s answers say which definition they used. That one page saves more arguments than any technical control.

One readingHeadcount on 31 MarchEmployed status, from work historyContractorsExcludedOn leaveCountedAnother readingHeadcount on 31 MarchCurrent profile, every active workerContractorsIncludedOn leaveExcludedOne readingHeadcount on 31 MarchEmployed status, from work historyContractorsExcludedOn leaveCountedAnother readingHeadcount on 31 MarchCurrent profile, every active workerContractorsIncludedOn leaveExcluded
Both readings are defensible. Agree one, and make every answer say which it used.

Claude and ChatGPT add their own layer

Claude and ChatGPT reach remote MCP servers as connectors. On business plans, an owner or workspace admin usually controls which connectors exist and who can add them, and users then connect with their own account. With user-level access, the same account decides what the HR system returns, so the assistant never sees more than the user already can.

Check two things in testing. First, that two users with different HR permissions get different results, or a refusal, for the same question. Second, that the tools the assistant can call only read. An assistant that can update an employee record needs a separate approval.

How Nexatron handles HR systems

Nexatron connects to HiBob, Rippling and Personio through each vendor’s API with user-level access: users sign in with their own account (or your SSO), and Nexatron inherits their existing HR permissions. People ask in the Nexatron app or from Claude, ChatGPT and Microsoft Copilot through Nexatron’s hosted MCP server. Questions only read the HR system. Each answer shows the reading and definition it used, with the other readings one click away, and it reads Verified only once a named person on your team has reviewed the logic.

A question outside your rules is declined with the reason and what to do next. The HiBob headcount and attrition pages, the Rippling headcount and leave pages and the Personio absence and turnover pages list example questions and the definitions to agree first. For people cost, the HiBob and NetSuite page shows how headcount and posted salary expense meet in one answer.

Claude is a trademark of Anthropic, PBC.

Bring three business questions

Explore your sources, definitions, and review criteria with our team. Please keep credentials and sensitive records out of the inquiry.