In a typical mid-size company in 2026, finance runs an agent that matches invoices to purchase orders, and sales drafts proposals in a chatbot. Support switched on the AI features in its help desk, and HR screens CVs with a plug-in in the applicant tracking system. None of it went through IT, and every team would fight to keep it.
The surveys say that's common. In the 2026 Enterprise AI Trends Study that FTI Consulting ran for Smarsh, released in July, 55% of enterprises said they're actively deploying AI, but only about a quarter said their governance is fully aligned with that pace. Employees describe the same gap from their side. In a global survey Okta published in May, 52% of employees admitted to using unapproved AI, often through personal accounts.
Budgets for next year get drafted in the autumn. Before you fund anything new, find out what you're already running.
Why a ban backfires
The first instinct is a policy that only approved AI may be used. The people using a personal chatbot account to clean up a spreadsheet keep using it after the memo goes out, and they leave it out of any inventory you run afterwards. Run the inventory before you write the policy.
A form and a week
Send every department head the same short form and give them a week. For each AI app, agent or feature their team uses, ask:
- What does it do, in one sentence?
- Who pays for it, and on which card or contract?
- What data goes into it? Customer records, financials, contracts, source code, nothing sensitive?
- Can it act on its own, or does a person review every output?
- Who would notice first if it stopped working?
People tend to leave the fourth question blank. Chase it. An agent that drafts and waits for a person carries a different risk from one that sends or pays, and you can't sort the list without knowing which kind you have.
Expect gaps. People forget the AI features inside software they already own, like the summariser built into the CRM or the writing help in the email client. Run a second pass two weeks later and ask specifically about features inside existing products, then cross-check against the finance team's card statements, which show subscriptions nobody remembers signing up for.
Sort the list by exposure
Rank each entry by what goes into it and what it can do. A summary of public web pages for a marketing draft is low risk on almost any service. Put customer contracts into an agent and it moves up on the data alone. Anything that can move money or contact customers without review is at least medium, and high if it also touches sensitive data.
Three levels are enough for a first cut, and our guide to tiering AI use cases covers how to draw the lines. In short, low entries get registered and a check of the vendor's data terms. Every medium entry needs a named owner who tests it against past cases and reads its logs monthly. High entries also get a person approving what they do until the logs show a track record, and a written record of how they were evaluated.
Do the sort before any reviews start, so a month doesn't go on the meeting-notes summariser while an agent issuing refunds runs unchecked.
Make the approved path faster than the workaround
People give up on governance at the approval step. A team waits six weeks for an answer and goes back to the personal account.
Publish a pre-approved list of apps and services for low-risk work that anyone can use from day one, and give medium-risk requests an answer within two weeks from a named person. When you approve something, approve it for the company, so the second team that asks for the same thing gets a yes the same day.
Give the teams that moved first some credit. They found real uses without a budget line, and a workflow someone built on their own time is often the best candidate you have for a properly engineered system. Bring those people into the review.
One owner for the list
Governance needs one person, usually in IT or enterprise architecture, who owns the inventory and the approval path. Each use case keeps an owner in the business, because finance knows what the invoice agent should do better than anyone in IT does.
Approvals go through that owner. A committee meeting monthly can't keep up with every department shipping AI on its own schedule, and teams stop sending it requests.
Start with the form
Send the five questions to every department head this week, with a deadline, and tell people the goal is to know what's running so the company can support it. Leave the policy for later, and publish the pre-approved list for low-risk work at the same time, so nobody waits for the inventory to get a yes.
When the answers come back, pick the three entries that take in the most sensitive data and can act without a person, and find out who reviews what each one does before it reaches a customer or a ledger.
Frequently asked questions
What is shadow AI?
It's any AI app or service people use for work that the company hasn't approved or doesn't know about. That includes a personal chatbot account used on a customer contract, or a team subscription IT never reviewed.
Should we ban unapproved AI apps?
A ban moves the usage somewhere you can't see it. In a global survey Okta published in May 2026, 52% of employees admitted to using unapproved AI, often through personal accounts. Register what's in use first, then close the uses that put sensitive data at risk.
Where do we start if we have no AI inventory at all?
Ask each department head for one list: what they use, who pays for it, what data goes into it, whether it can act without a person checking, and who'd notice if it stopped. A spreadsheet is fine.
How do we sort AI use cases by risk?
By what data goes in and what the agent is allowed to do with its output. Start with low, medium and high, and put anything that acts without a person's approval at medium or above.
Who should own AI governance?
One person, usually in IT or enterprise architecture, owns the inventory and the approval path. Each use case keeps an owner in the business, because the people who run the work know what the agent should do.