Is Anyone Watching How Your Team Uses AI at Work?

TL;DR: A policy tells people what to do. It can't tell you when they didn't. Ask your security team for a short watch list: five to ten AI events worth a ticket, each with an owner and a first check. It takes an hour to write, and it's the part of an AI policy someone can actually run.

Key takeaways:

• Josh Woodruff took six client calls between August 31 and October 8, 2026, where a security team asked how to watch for AI use at work.

• A policy says what's allowed. A watch list says what your team will notice and who acts on it.

• On October 1, 2026, a national retailer's internal audit team asked whether anyone was watching for sensitive content going into AI. The security team had been working through a committee for a year.

• On September 28, 2026, a state agency with no approved AI wanted controls running first, so leadership could say yes with guardrails already in place.

• NIST's incident response guide, SP 800-61 Rev. 3 (April 2025), treats detection as one job inside overall cybersecurity risk management. Leaders own part of it.

What does "watching for AI use" actually mean?

It means your security team keeps a short list of AI events it would act on, and has a way to notice each one. It stops well short of reading everyone's chats. A good list holds five to ten items. Each item has an owner and a first check. Think smoke detector.

Most leaders picture something bigger: a dashboard, a vendor, a team of analysts, and an archive of everyone's prompts.

The small version starts with one question for your security team: which AI events would you open a ticket for? If the answer comes back in a sentence, you have a watch list. If it comes back as a procurement request, you have a shopping trip.

Why can't a policy do the watching?

A policy is a promise people make. When someone breaks it, nothing happens unless a person or a system is looking. Turns out a policy can't watch anything. It can only hope.

Josh logged the same shape on five calls between September 21 and October 8, 2026: a policy on paper and no way to tell when it was broken. On one of them, the cost showed up as time. Analysts worked alerts one at a time, because nobody had written down which alerts counted.

Two of those calls show the pressure from opposite ends. At a national retailer, the internal audit team asked whether anyone was watching for sensitive content. A committee had handled AI by hand for a year, and leadership wanted something more formal. At a state agency, nobody had approved any AI yet. The security team wanted controls running before that happened, so leadership could say yes with guardrails already in place.

One team was asked about the past. The other was getting ready for the future. Both needed the same thing: a list of what they'd notice.

Which AI events are worth a ticket?

Pick the ones where someone would act the same day. Company files going to a personal AI account in bulk. A new AI app connected to company accounts. An AI agent reaching data the person asking couldn't see. Someone turning off logging, or switching on a connector for the whole company.

Here's why those four, in business terms:

• Files leaving for a personal account.

  Once they're there, the company can't pull them back or see who else reads them.

• A new app connected to company accounts.

  One approval click can give an app standing access to mail, files, calendars, and shared drives.

• An AI agent reaching past its requester.

  The person asking has limits. The agent shouldn't hand them a way around those limits.

• Logging switched off.

  Without logs, the other three become guesses.

Josh's post for security teams, what a SOC should alert on first for AI use, covers the logs behind each event. You can lead the work without that level of detail. You need to know the list exists and who owns it.

What should a leader ask the security team?

Ask four questions. Which five to ten AI events would you open a ticket for? Who owns each one? How would we find out today if it happened? Has anyone tested that? Four honest answers beat any dashboard. A blank answer is fine. It's a task with a name beside it.

The second question trips people up. Many AI platforms are run by a different group than security, often a workplace technology team. Security sees access through user groups and not much else. Someone has to own the list as a whole, and someone else owns each platform's settings.

The fourth question is the one that counts. A watch list nobody has tested is a document. Run one safe test event and time how long it takes a named person to see it. The post on testing whether you'd see an AI agent misbehave in real time shows how to run that clock.

NIST's SP 800-61 Rev. 3, published in April 2025, aims to help organizations prepare for incidents and make finding and fixing them work better. It ties incident response into overall risk management. That puts the watch list on your desk, too.

Do we have to read everyone's chats?

No. Prompt and chat text is private, and your legal and privacy teams should decide who can read it, and when, before anyone collects it. Reliable alerts come from other signals: who signed in, which apps got permission, what files moved, and where they went. Chat text is something an investigator opens after one of those fires.

On October 8, 2026, Josh heard a security team ask whether it should collect everything, chat included. His answer was to start from the failures. The team picks the AI events it would act on, then finds the least it needs to collect to spot each one.

This also protects trust. Tell staff what you watch and why, and keep chat text off limits until legal and privacy teams set the rules.

Frequently asked questions

Do we need to buy a new product first?

No. Write the list first. Two of the most useful alerts, large uploads to AI sites and new app approvals, run on logs most security teams already collect. After a couple of months, whatever the list can't see tells you what to buy.

Will employees feel spied on?

Watching for files leaving the company is a different act from reading what people type. Say so in plain words, and publish the list where staff can see it.

Who should own the watch list?

One named person owns the whole list and keeps it current. Security owns the alerts. Each AI platform has its own administrator who owns its settings. Everyone should know which name sits beside which item.

How often should we update it?

Review it every quarter, and any time you approve a new AI platform. New platforms bring new ways to fail. The list should grow by one line when that happens.

What if leadership is slow to block personal AI accounts?

Start with the watch list anyway. The alerts give you real numbers, and numbers make the blocking conversation easier. One of the teams Josh spoke with was in exactly that spot.

Write the list this week. It's short, it's plain, and it gives you a real answer the next time someone asks whether anyone is watching.