AI employees in compliance

The pile is the product

Compliance work often starts as PDFs. Someone reads them. Someone types the same fields into a form. Then they do it again tomorrow. That is not a strategy problem. That is a process you can write down.

At SmartComply we used structured output so the system had to return a shape we could check. The reviewer still owned the miss. Review time dropped by about half. That is the whole story, minus the slogans.

First pass is a job. Decision is a name.

I do not ship a reviewer-less system for this kind of work. The first pass can be automatic. The miss still needs a name next to it. If that sounds slow, try explaining a silent miss to a customer, or to a supervisor who asked for the file.

Speed without an owner is how compliance theater happens. You look faster in week one. You cannot reconstruct week four.

Make the output boring enough to check

A model that writes a paragraph is a liability. A model that fills a schema is a worker. You can test a schema. You can reject a missing field. You can send the weird case to a person without pretending the prose was a control.

Zod, types, a reject path. None of that is a demo. All of it is why the half we saved did not come back as cleanup.

• Return a fixed shape or fail.
• Cite the page or passage the field came from when you can.
• Never fill a blank with a guess.
• Queue exceptions with the original document still attached.

High-risk is already your life

If you are scoring people, filtering them, or feeding a regulated decision, treat the employee as high-risk whether or not a lawyer has used those words yet. The EU AI Act puts employment and certain credit uses in that bucket. You still need oversight, data you can defend, and a human who can override.

I build that way for document review even when the Act is not the headline. Logs and a named reviewer are cheaper than a story you invent after the miss.

Incident response is the hidden win

When reviewers are not buried, they can answer the ugly ticket. That is what got faster after the pile shrank. People talk about accuracy. I watch whether the team can still think.

If your AI employee creates a second pile called 'check the checker', you have not won. You have moved the hours. Tighten the loop or turn it off.

One document type, then the next

Pick the pack that arrives every day. Map the fields. Connect to the tools the reviewers already use. Hand it over with a miss-rate number you can say without a slide.

Then stop. The second document type is a different job. Do not celebrate a platform until the first loop is boring.

What I will not call a win

A dashboard that says documents processed is not a win. A reviewer who gets home without a second pile is a win. A miss you can explain on a call is a win. A field that was invented and then 'fixed' in cleanup is a loss you will pay for twice.

I like TypeScript, Node, and Postgres for this work because I can see the shape. I like a schema the way I like a ledger. If the employee cannot live inside that, it does not belong next to a policy binder.

If you have one document type that still runs on people copying things, that is usually the first job. We can look at it in thirty minutes. If the pack is too exotic to automate, I will say so. Most of them are not exotic. They are just unread a second time.