HR processes your AI agent can run.

An HR process your AI agent runs end to end, with no screen, under the permissions of the person who asked. Enroll a new starter from Claude.

Why AI help stops at the chat window

Your People team already asks an AI assistant for help. Usually that means pasting an export into a chat.

The answer comes back, and then the work starts. Someone opens the HR tool, finds next week's starters, and clicks through them one by one. The assistant gave advice. A person still ran the process, and employee data went into a chat to get there.

Giving the assistant more reach brings a different problem. HR agents tend to fail in the same few ways, and each is worth testing for before you sign anything:

  • It doesn't know when to escalate, so it answers a complex leave question instead of handing it to HR.
  • It can't explain its actions, so HR can't trace what it said, to whom, or why.
  • It makes up a policy that doesn't exist, and sounds sure about it.

Joris Luijke, Pyn's co-founder and co-CEO, lists these and three more in MCP for HR AI agents and AI tools.

So ask three things of any agent. What can it do? As whom? And where is that written down?

What headless means

Headless means there is no screen to click through. A headless process is an HR process your AI agent runs end to end through Pyn, under the permissions of the person who asked.

You ask in the assistant you already use: Claude, ChatGPT, or Cursor. The agent does the steps. Pyn checks each one and keeps the record.

One run, start to finish

Take a job someone on your team does every week: enrolling next week's new starters in pre-boarding.

You ask
In Claude, in plain words: “Enroll next week's starters in pre-boarding.”
The agent looks
It finds who starts next week from live data in Pyn, which reads your HRIS. There is no export, and nothing is pasted into a chat.
The agent acts
It enrolls each starter, as you. Pyn checks every call against your own role first.
Pyn records it
The change appears in Activity like any other change.
The flow takes over
From here it is ordinary Pyn. Each starter gets their pre-boarding on their own dates, from a real colleague, in Slack, Microsoft Teams, or email.

The agent didn't write the welcome or decide who gets it. You set that up once. It did the clicking.

How Pyn sets it up

The assistant reaches Pyn through MCP. That stands for Model Context Protocol: the standard way an AI assistant connects to another system, and the place you set what it can, should, and shouldn't do there. You don't need to read the protocol. You need to know who decides what.

An admin decides what the agent can do

A workspace owner or admin creates an MCP server in Pyn. Think of it as a list of the operations an assistant may use. Start from a preset, such as People Directory (read only) or Onboarding Coordinator (find new starters, review their progress, and manage permitted onboarding steps), or pick operations one by one. Each is marked Read or Write, and the admin sets the role a person needs to use the server.

Each person connects as themselves

Each person adds the server to their own assistant and signs in to Pyn. Before anything connects, a consent screen lists the tools the assistant will get, grouped into what it can read and what it can change.

A Skill tells the agent how your company does it

An agent needs what a new team member needs: a clear scope and a playbook. In Pyn that is a Skill, a named set of instructions that connected assistants follow. Write your way of running onboarding once, and every connected agent follows it.

“When you hire a new HR team member, you'd give them a clear scope and a playbook or detailed instructions.”

On the HRIS you already have

Nothing gets replaced. Pyn reads your HRIS, so the agent works on the same people data and the same flows that already run your onboarding. It works with BambooHR, HiBob, Rippling, and Workday. See every integration.

Your permissions, not a master key

The worry with any agent is that it sees everything. This one can't. Every call runs as the signed-in person, with their own permissions. Your AI assistant only sees what you could already see, and nobody shares a master key.

Offering an operation never grants access. If you can't do something in Pyn, you can't do it through Claude either.

Four things to settle before an agent touches people data

The four come from Joris's post, and they hold for any HR agent, not only one that runs on Pyn.

Permissions
What data and tools can it reach? In Pyn: only what the person asking could already reach.
Boundaries
What must it never do? The admin leaves those operations off the server, or keeps the whole server read only.
Role
What is the agent for? Put it in the Skill, in the words you'd give a new team member. Joris's example: “You are an Onboarding Assistant. Your job is to help new hires feel supported, not to answer legal questions.”
Handover
When does it stop and ask a person? Write that into the Skill too, then test it with a question it shouldn't answer.

For the rest of your diligence, take 12 questions HR leaders should ask AI vendors to every demo, including ours.

What to hand over, and what to keep

Hand over the steps that have a right answer. Who starts next week? Which flows is this starter in? Who hasn't finished their onboarding checklist? Then the actions that follow: enroll the starter, send the reminder, resend the message that bounced.

Keep the calls that don't have one. An agent can flag that a manager hasn't checked in. It can't tell you whether that manager needs a reminder, a conversation, or something else.

Where a headless process stops

“Every system has ways it can fail. The question is whether the vendor knows them and can articulate them.”

Here are ours.

  • It answers from the data in your HRIS. A wrong start date there is a wrong answer here.
  • It only changes things if an admin has allowed write operations, and only as far as the person's own role allows.
  • It is the wrong tool for legal questions, payroll changes, and anything about a termination. Leave those operations off, and tell the agent to hand them to a person.
  • It doesn't replace your automations. A step that should happen for everyone, on a date, belongs in an automation you set up once. A headless process is for the work in between: the question, the exception, the one-off.

The difference between the two is set out in AI agents vs. HR automation.

The AI Agent Runner is next

Today a headless process starts with someone asking. The AI Agent Runner is the next step, and it isn't available yet.

The plan is that you write a Skill and an agent runs it without being asked. It gathers data, spots the patterns that matter, and delivers the insight straight to Slack or Microsoft Teams. Nobody digs through a dashboard. An HR business partner gets early warning on turnover risk. An onboarding coordinator gets a daily health report with fix-it buttons.

These pages say how it is meant to work with each system:

Start with one process

Skip the AI strategy deck. Pick one process a person on your team runs every week, and hand over that.

  1. Pick the process.Choose one with a right answer and a weekly rhythm, such as enrolling next week's starters.
  2. Start read only.Have an admin add a read-only server such as People Directory. Ask who starts next week, and check the answer against your HRIS.
  3. Connect as yourself.Add the server to your assistant, sign in to Pyn, and read the consent screen. Then ask “who am I in Pyn?” It answers with your name, the workspace, and the server.
  4. Write the Skill.The scope, the playbook, and when to hand over to a person.
  5. Add write, with a person watching.Move up to Onboarding Coordinator. Keep a person on the first few runs, and check each change in Activity.

Book a demo. Bring your HRIS and the process you'd hand over first. We enroll a starter from Claude, as a named person, and show you where it was recorded.

FAQ

What does headless mean?
No screen. Nobody opens Pyn and clicks through the steps. An AI agent does them from the assistant you already use, and Pyn checks and records each one.
What is MCP?
Model Context Protocol: the standard way an AI assistant connects to another system. In Pyn, an admin creates an MCP server that lists exactly which operations an assistant may use, each marked Read or Write.
Which AI assistants work?
Claude, ChatGPT, and Cursor.
Can the agent see more than I can?
No. Every call runs as the signed-in person, with their own permissions. Your AI assistant only sees what you could already see, and nobody shares a master key.
Can the agent change things, or only read?
Only if an admin allows write operations, and only as far as the person's own role allows. You can keep a server read only.
Who sets it up?
Workspace owners and admins. They create the server and choose the role a person needs to connect.
Is the AI Agent Runner available?
Not yet. It is coming soon. What runs today starts in Claude, ChatGPT, or Cursor, when someone asks.

Your HR stack, with superpowers.

Pick the tools you need. Launch in a day. Cancel anytime.