← Blog/Point of view

An agent should never be able to do more than the person who ran it

Service accounts are why security teams say no. Here is the permission model that gets them to yes.

Tony Darugar//1 min read
run #2318
09:14:02readnetsuite.vendor#4471
09:14:03deniedworkday.salary
09:14:03flagneeds a person

This is a placeholder post. It exists so the blog shows its real layout; Rayyan replaces it, or deletes it, before launch.

The service account problem

Placeholder text for this section. The finished post will make one point here, in plain language, with the numbers or the example that backs it up.

Running as the person

Placeholder text for this section. The finished post will make one point here, in plain language, with the numbers or the example that backs it up.

More on the blog

Point of viewThe last two months

Why most enterprise AI stalls before production

The proof of concept is the easy part. The last two months, where security, permissions, and the systems of record get involved, are where projects go quiet.

Five questionsBefore it touches the ledger

The five questions a controller asks before an agent touches the ledger

None of them are about the model. All of them are about the record.

10×apart in cost per correct answer, at the same accuracy

Accuracy is the wrong number to pick a model on

Two models that score the same on a benchmark can cost ten times apart per correct answer. Here is the number a finance team should look at instead.

From the research

Benchmark · Sep 18Contract review: where fourteen models hold up, and where they don'tRead the paper →Whitepaper · Sep 2Cost per correct answer on enterprise tasksRead the paper →Benchmark · Aug 20The monthly close, model versus modelRead the paper →Whitepaper · Aug 5When agents should stop: measuring the flag rate finance teams actually wantRead the paper →

Bring us a process your team still does by hand.

We’ll give you automated workflow agents.

Get a demoRead the research