Personaas

The AI agents in your company that nobody approved

Shadow AI isn't a threat on the horizon. It's the analyst three desks over, quietly wiring a language model into your vendor contracts this afternoon.

A finance analyst on your team has a script open right now. It reads through vendor contracts, sends the text to a chatbot, and pulls back a tidy summary of the payment terms. It saves her about two hours a week. She built it in an afternoon. She didn't ask anyone, because it never occurred to her that she should — it felt about as significant as writing a spreadsheet formula.

Multiply that by every capable person in your organization. The support lead who set up an agent to draft replies from customer emails. The recruiter pasting résumés into a model to rank them. The ops manager whose Zapier flow now runs half his reporting through an API he found on a Tuesday. None of these people are reckless. Most of them are your best people — the ones curious enough to try the new thing and skilled enough to make it work.

That's the uncomfortable part. The problem isn't bad actors. It's good employees doing exactly what you'd want them to do, using tools that happen to touch data you'd never knowingly hand to an outside service.

The stuff you can't see is the stuff that hurts

When people talk about "shadow IT," they usually mean someone expensing a SaaS subscription that never went through procurement. Annoying, but visible — it shows up on a credit card statement eventually. Shadow AI is quieter than that. A prompt isn't a purchase. There's no line item, no login for IT to audit, no vendor to negotiate a data-processing agreement with. Just a person, a text box, and whatever they decided to paste into it.

Ask yourself three plain questions about the analyst and her contract summariser:

  • Where did those contracts actually go, and is a copy sitting in a training set now?
  • If the summary quietly gets a renewal date wrong, who catches it before the company misses the window?
  • When she leaves, does the script leave with her, or does it keep running against systems nobody remembers granting it access to?

Most companies can't answer any of the three. Not because they're careless, but because there was never a moment where someone stood up and said "we are now an organization that runs AI agents against confidential data." It just happened, one helpful afternoon at a time.

You didn't decide to adopt AI across the business. Your business decided for you, and forgot to tell you.

Banning it doesn't work — and you know it

The instinct, once this lands, is to lock it down. Block the domains, send the stern email, add a line to the handbook. It feels like control. It isn't. People who found a tool that saves them a morning a week don't stop using it because of a memo — they just stop telling you they use it. You don't remove the risk. You remove your visibility into the risk, which is strictly worse, because now the agents are still running and you've taught everyone not to mention them.

The other failure mode is the year-long committee: a working group, a policy document, a procurement cycle, a pilot that finally launches after the technology it was evaluating has been replaced twice. By the time governance arrives, the ungoverned version has been load-bearing in three departments for months.

Neither extreme actually engages with what's happening. The demand for this help is real and it isn't going away. The only useful question is whether it runs somewhere you can see it.

A harness, not a fence

Here's the reframe we keep coming back to. You don't govern AI by building a taller fence around it. You govern it by giving people a version that's genuinely better than the one they'd cobble together on their own — and making the safe path the easy path.

That means an AI coworker that already knows which systems it's allowed to touch and which it isn't. One where every action it takes is logged, so "what did it access?" has an answer you can pull up in seconds. One with a name, an owner, and a defined job, instead of an anonymous script that outlives the person who wrote it. When the sanctioned option is faster and less fiddly than the DIY one, the analyst uses it — not because policy forced her, but because it's the better tool.

The shift in one line Governance stops being the thing that says no, and becomes the thing that lets people say yes safely.

This is what we mean when we call Personaas an enterprise AI harness rather than another AI tool. A harness isn't a cage — it's the thing that lets you put real power to work without it dragging you somewhere you didn't intend to go. The agents are coming into your organization either way. The only decision you actually get to make is whether they arrive with names, owners, and an audit trail, or whether you find out about them the hard way.

Start by assuming they're already here. Because they are.

More from the blog