Ask most company owners which AI platforms their team uses and you will get a confident answer: the one or two the company pays for. The research says otherwise. In survey after survey, roughly 8 in 10 workers admit using AI platforms their company never approved, and virtually every organization studied - 98% in one 2026 report - has some unsanctioned AI in use.
This is called shadow AI, and if you are trying to control AI costs, it is the part of the iceberg under the water.
Why it happens (and why it is not malice)
People just want to get their work done. A salesperson finds a tool that writes proposals faster. An analyst pastes data into a free chatbot because the paid one was never set up for them. Nearly half of employees say they taught themselves AI with no company support, and 60% of businesses admit their staff never got proper AI training.
Shadow AI is what filling that gap looks like when the company does not do it first.
It is also self-reinforcing. The person who found their own tool gets faster. A teammate copies them. Within a quarter, the workaround is the workflow. By the time anyone asks "what are we actually using?", the honest answer lives in dozens of browser tabs and personal accounts - not in any system the company controls.
What it costs you
- Duplicate spend. Personal subscriptions on expense reports, three teams paying for three tools that do the same thing.
- Invisible usage. You cannot right-size or optimize spend you cannot see. Your real AI bill is bigger than the invoices you know about.
- Data walking out the door. Company documents pasted into free consumer tools live outside your control. That is a security question as much as a cost one.
What this looks like at a 50-person company
Put those percentages on a real team and the picture gets uncomfortable fast. Say you run a 50-person company. You pay for one sanctioned AI platform, and you assume that is the whole story.
- 8 in 10 workers use unapproved tools. On your team, that is roughly 40 people using AI you have never heard of.
- About half are unaware of the risks. Call it 20 people pasting customer lists and draft contracts into free tools without a second thought.
- 60% of businesses never trained their staff. If you are one of them, nobody ever told those 20 people where the line is.
Now the money side. Across companies, 10-20% of paid AI seats typically sit idle. On 50 seats, that is 5 to 10 people you pay for who never log in - and some of them are the same people expensing a personal tool that does the same job. You are paying twice: once for the seat nobody uses, once for the workaround. We wrote up that first half of the problem separately in the hidden cost of idle seats.
None of these numbers make your team unusual. They make it typical. That is the point.
The wrong fix and the right one
The instinct is to ban unapproved tools. It does not work - usage just goes further underground, and you lose the productivity along with the visibility. Companies that handle this well do the opposite:
- Offer good sanctioned tools so nobody needs to go around you.
- Get visibility first, judgment later. Find out what is actually in use and what it costs before deciding what stays.
- Train people. Companies with real AI training see far fewer incidents and much better usage habits.
Notice what is not on that list: policing individual prompts, or asking people to log their own usage by hand. Neither survives contact with a busy week. Visibility has to come from the billing and admin side, where the data already exists and nobody has to remember anything.
How to check this in your own company this week
You do not need a security audit to size your own shadow AI. Five checks, a few hours total:
- Sweep the expense reports. Search the last three months for AI platform names and for anything vague labeled "software" or "subscription". Personal subscriptions are the paper trail shadow AI leaves behind.
- Check sign-in logs. If your company runs on Google Workspace or Microsoft 365, the admin console lists third-party apps people have connected to their work accounts. AI platforms show up there, even the free ones.
- Count seats against active users. Open the admin page of the tool you do pay for. If 10-20% of seats have not been touched in a month, that is your idle-seat number - and often a hint that those people do their AI work somewhere else.
- Run the amnesty survey. One question, anonymous if that helps: "Which AI platforms do you actually use for work?" Say clearly that the goal is funding better tools, not punishing anyone.
- Compare the list to your invoices. The gap between what people use and what you knowingly pay for is your shadow AI, in one line.
"If people pay out of pocket, isn't that free for us?"
The most common pushback, and it sounds reasonable: if an employee covers a subscription personally, the company saves money. Three problems with that.
First, it rarely stays personal. Subscriptions migrate onto expense reports one by one, and duplicate tools pile up with nobody comparing them. Second, the data risk does not care who pays. A contract pasted into a free consumer tool is outside your control either way. Third - and this is the cost angle - you cannot fix what you cannot see. Industry research finds 30-40% of company token spend is waste, and 79% of companies overshot their AI budget. Part of that overshoot is exactly this: spend and usage nobody had on the map.
Shadow AI is not a side story to cost control. It is the reason the AI cost management picture never adds up when you only look at the invoices you know about.
Where Optimize fits
Optimize connects to your company billing and the official AI providers, which surfaces the subscriptions and usage you already pay for - including the ones nobody mentioned. Full picture first, then the savings plan.
The mechanics matter here: connecting takes about 15 minutes, access is read-only, and the first report lands within 48 hours. Nobody installs anything on laptops, and nobody gets named and shamed. The output is a map - tools, seats, spend - including the 10-20% of paid seats that typically sit idle. Trimming those alone is usually worth 5-15% of the bill, before you touch a single workflow.
From there, the usual path is the one described in how companies cut their AI bill 40%: see everything first, then right-size, then keep watching. Shadow AI is step zero. You cannot start the savings until the map is honest.