PersaliaAI for HRFoundations

Can AI get into your HR system? Only if there’s a door.

How an AI actually gets into an HR system, and why it all comes down to the key. Written for HR people, from my own build with 429 fake employees.

Field
AI for HR
Track
Foundations, 1 of 3
Published
Reading time
13 min
Also on
LinkedIn

Someone’s going to ask it. Probably in a leadership meeting, probably right after a demo. “Can’t we just hook the AI up to the HR system?”

And the answer is yes, most of the time. It depends on whether the system allows it, how it allows it, and who you allow it for.

I spent the last few weeks building this myself, with a fake company of 429 employees, so I’ll walk you through it the way I had to learn it. No jargon you don’t need.

The doorAPIThe part of the HR system that lets another system talk to it. No API, no way in.
The room behind itMCPA control room built in the HR system. What’s installed is what’s there.
What gets you inThe keyUnlocks the door. What it’s attached to decides what you can do inside.
The three parts of this piece. Each one gets its own section below.

First, you need a door

The door is the API. That’s the part of the HR system that lets another system talk to it. Payroll already uses it. Your login provider probably uses it.

Think of the HR system as a room. If the room doesn’t have a door, you can’t get into it. So if there’s no API, there’s no way in.

Sure, someone with access can take screenshots and show them to the AI. We all do that. I hope. If not, it’s just me. But that’s not access. That’s a person manually moving information into the AI, and then having to put it back by hand if something needs changing. And it’s limiting, because you only ever get what that one person can see.

Okay, technically there’s a window. Some AI tools can drive a browser like a person, using that person’s login. It’s slow, it breaks easily, and it gets everything that person can see. So, a window. Not a door.

And here’s the part most HR people don’t know. This door is different from your other logins. It’s not locked behind single sign-on, a company email or two-factor authentication. It’s locked by a key.

And this key is special. So special that many systems only show it to you once. If you don’t save it, you have to make a new one. And that key can take any AI straight into the system without any of your other login measures. It never sees the login page at all.

The room behind the door

This is the MCP. It stands for Model Context Protocol, but nobody calls it that, so we’ll stick with MCP. Anthropic created it, and since December 2025 it’s run by a foundation under the Linux Foundation, with OpenAI and Block as co-founders. So no single company owns it.

The people behind it call it “a USB-C port for AI applications”. I look at it as a control room.

The room is built in the HR system. Sometimes the vendor decides everything in it, sometimes you get to. Either way, what’s installed is what’s there. If there’s no button that can change salaries, then nobody changes salaries through it. It doesn’t matter who you are.

Every control room works the same way, and that’s the whole point. Claude, ChatGPT, Copilot, they all know how to read the panel and press the buttons. So the vendor only has to build the room once, and every AI can use it.

But vendors build very different rooms.

Simployer built one room with one option. Their guide says it plainly: the MCP “exposes no data and no operations beyond those already available through the Public API”. But their keys are owner level. So this control room controls everything. It can see the data and it can change the data.

Frappe HR, the open-source system I used, works differently. In Frappe you build your own room. A room with only monitors is a monitor room. Add buttons, and now it’s a monitor and change room. You decide how many monitors there are, and how many buttons. You can give it everything: read, write, change, update, whatnot. Or, on the safer side, which is what I did, you tick the box that says it can only read. All monitors, no buttons. I could also have gone further and only put up the monitors for names, emails and addresses, and left out salaries and other sensitive information. In Frappe you can do that. In Simployer, you get the one room: see everything, do everything.

And some vendors haven’t built a room at all. I couldn’t find an MCP from Huma or CatalystOne. Then you’d have to go through a third-party service that builds one for you, and now there’s more than one company handling your employees’ data. That’s not something I’d advise.

Which is why, with AI, it’s not enough to check that your HR data is good. You also have to look at how that data can be reached, and by what.

Who gets the key?

So you have the room. You have the door. How the hell do you get in? Do you run at it? Knock it down? Yell at it?

No, no, no. You get a key. The key unlocks the door, you walk in, and you get whatever the control room lets you do.

There are two ways to get one.

A key made in advance. Usually an admin generates it, copies it and hands it to the AI. It works until someone deletes it.

That’s what I did. I made a user called AI Auditor and gave it read-only access. Then I pulled an API key for that user and connected it to Claude. So Claude can see everything AI Auditor can see, and do everything AI Auditor can do. That key stays there until I delete it. And anyone who gets hold of it can reach everything AI Auditor can reach, which is why these keys need to be kept safe.

Signing in as yourself. This is OAuth. You’ve done it a hundred times. “Sign in with Microsoft. This app wants to read your calendar. Allow?” That’s it. You click Connect in Claude, the HR system’s own login page pops up, you sign in with your work login, and you click allow. Claude gets a key that’s yours, and it runs out after a while.

So which one is best?

In my demo I used a key made in advance. But I’m one person, so that’s fine.

Now picture ten HR advisers and that one key. Everyone sees what the key sees. Take a couple of scenarios.

The junior adviser sees everything the HR director sees. Maybe the director should have more access than the junior. Doesn’t matter. Same key, same view.

Or there’s someone on the team who’s a bit nosy, and they use the key to look up things they shouldn’t. The system only logs what AI Auditor did. So you can see that someone has been snooping. But you’ll never find out who.

And lastly, someone leaves. You want to shut their access, and what you usually do is delete the key. But delete that key and the whole team loses access, and you have to make a new one.

So then the question is: if you have a company of 400 people, do you make 400 keys?

No. This is where signing in comes in. An admin switches sign-in on once, and only the people who actually use the AI connect. Probably the HR team, not all 400. Maybe the tech team too, another 20 people. That makes access a lot easier, and it’s locked down.

But here’s what’s important. Signing in only works if the system can break access down and follow each user’s profile. Simployer doesn’t offer sign-in for its MCP. Everyone would get owner level anyway, and that just doesn’t work.

Not every key is the same

We’ve touched on this, but this is the bit that surprised me. I assumed a key was a key. It isn’t. It depends on what the vendor attached it to.

Simployer One. The key isn’t attached to anything. One key, owner level, everything, as we’ve already gone through. Their own guide says the tokens “carry owner-level rights across your tenant and cannot currently be scope-limited”. Read, and in several modules write, for every employee.

Hailey HR and Frappe HR. The key is a user, which is what I’ve described so far. You create a user, call it AI Auditor, give that user access in the system, and connect the key to it. That’s how you build the room. It sees what the user sees. Hailey says the agent “sees exactly what the connected user is allowed to see. Nothing more.”

Huma, CatalystOne and SAP SuccessFactors. The key gets its own role. So instead of being a user in the system, the key itself gets a role, and someone decides what it can touch. A new key in Huma can’t do anything at all until you give it one. But Huma and CatalystOne have the door, not the control room. I couldn’t find an MCP from either. So Claude has no panel to read, unless a third party builds one.

Workday. The most complex of them all, and the most flexible. The agent and the user together. The AI only gets what both of them are allowed.

Take an example. The AI can see names, birthdays and emails across the whole company. That’s the AI’s side. The user can only see their own team. So when that user asks the AI, they get names, birthdays and emails for their own team. They still can’t see salaries, because that’s not part of the AI. And they can’t see outside their team, because that’s not part of their own rights.

Same question. Four answers. So the question for your vendor isn’t “do you have an API?” It’s “what’s your key attached to? What kind of control room are my people walking into?”

The key is attached toSystemsWhat that means in practice
NothingSimployer OneOne owner-level key for the whole tenant. It can’t be limited, so whoever holds it can see and change everything.
A userHailey HR, Frappe HRThe AI sees what that user sees. Make the user read-only and the AI is read-only.
Its own roleHuma, CatalystOne, SAP SuccessFactorsSomeone decides what the key can touch. Huma and CatalystOne have the door, but I found no control room.
The agent and the user togetherWorkdayThe AI only gets what both the agent and the person asking are allowed.
Same question, four answers. A summary of this section, from each vendor’s own documentation linked above, as I found it in November 2026.

So who asked?

Imagine AI Auditor did something weird on a Tuesday. Pulled a whole department’s records. Who asked it?

With one shared key, you can’t tell. The HR system’s log just says AI Auditor. Anthropic says it straight in their own docs: with a shared key, “every member’s requests carry the same key, so don’t use it to identify which user is calling.”

With OAuth, the log shows the actual person. Workday goes further and logs both. The agent did it, on behalf of this specific person.

And this isn’t just nice to have. Under the EU AI Act, if you use AI for decisions about people at work, that’s considered high risk. And one of the rules is that the employer has to keep the system’s logs for at least six months. I’ll get into the AI Act properly in a separate piece.

When does that start? It was meant to be August 2026. Then the EU passed the Digital Omnibus, a package of changes to the AI Act, and that pushed it back, thank God. But not far. For employment systems, it’s now 2 December 2027.

Was meant to apply from
August 2026
Now applies to employment systems from
2 December 2027After the Digital Omnibus changed the AI Act
When the AI Act’s high-risk rules reach HR. From Article 26 and the Digital Omnibus.

That’s not long. Picking a system, switching on sign-in, getting six months of logs in place. That’s a year of work. And if you’re reading this when it comes out, a year is about what you’ve got.

Wait, what logs?

And this surprised me. When I went looking for proof of what Claude had done in Frappe, there was nothing on screen. No log. I had to open a terminal on the server running Frappe and pull it out with commands. This is what came out:

Server log lines pulled out of Frappe HR with commands.
The log, pulled from my own server with commands. There was nothing like it on screen in Frappe.

Turns out Frappe has a switch for it. Log API Requests, in System Settings. It’s off by default. Frappe says switching it on records each user’s API calls. In my case, that user is AI Auditor. So it still wouldn’t tell you who asked.

But I could only dig this out because it’s my server. If your HR system lives in the cloud, you probably can’t. No vendor is going to open their terminal and dig out what your AI did last Tuesday. So the log has to be switched on inside the system itself, before the AI goes in. Afterwards is too late. And even then, check what it actually records, because it might only ever say AI Auditor.

Hard rules and soft rules

So where are the guardrails? We’ve touched on this too.

The hard ones live in the HR system. The system checks who the key belongs to and says no. No, you can’t access your manager’s payslip.

That’s why in my demo, if Claude had tried to change anything, it couldn’t. I built it a room with monitors and no buttons. There was nothing to press.

The soft ones live in Claude. First, an admin decides which connectors your organisation can use at all. Then, for every tool inside a connector, you pick one of three: always allow, needs approval, or blocked. So you could let Claude read freely, make it ask you before it changes anything, and block deleting altogether.

They’re useful. But anyone holding the key outside Claude walks straight past them. If the HR system can’t enforce it, you’re hoping the AI tool will. And one thing we’ve seen is permission fatigue: so many pop-ups asking for permission that people stop reading them and just click allow.

I’ll go deeper on guardrails in a separate piece.

Simployer connects in five minutes. That’s the problem.

So could you connect Simployer One to Claude today?

Technically, yes. You paste in their address, add the token as a header, done. (The header option is still in beta in Claude, so not every organisation sees it yet.)

But think about what you’ve just done. Everyone in your organisation who connects gets owner-level access to every employee. There are no personal logs in Simployer. It’s all one token.

For a sandbox with fake data? It’s great, it’s easy, and it’s exactly what I want to test next. For real data? No. Not until that key can be tied to a user or a role. Until then, it might be something the HR department keeps to the few people who actually need it.

What I did

I built a fake company, Tindra Industrier AS, in Frappe HR. 429 people, filled in badly on purpose. I planted mistakes that should slip past the system’s own checks. Some didn’t, and the system stopped them at import, so I had to fix those. 127 made it in. A salary missing a zero. A wrong phone number. An employee with no manager, because theirs had left.

Then I connected Claude through a read-only key. One user, me. It caught 124 of the 127. The three it missed were small judgement calls on postcodes.

So AI is really useful here. Cleaning up an HR system, finding the mistakes nobody has time to look for. And that’s just one use. There are plenty of others already, and there’ll be a lot more.

But it’s not simple. So before you pick a system, or connect the one you’ve got, ask your vendor these questions:

  1. What’s the key attached to? Nothing, a user, or a role?
  2. Can I make it read-only?
  3. Can people sign in as themselves, or is it one shared key?
  4. Is there a log of what the AI did, and is it switched on?
  5. Does the log show my name, or the robot’s?
  6. If I pull one person’s access, does everyone else lose it too?
  7. Can I see every key that’s out there, and switch one off?

Written by Ronny W. Andersen. Everything I tested ran on my own server, with a made-up company and made-up people. If something here is wrong or out of date, I want to know.