// Resources
Which AI Tools Is My Team Already Using, and How Do I Find Out?
The five places to look without an IT department, what the 2026 survey data says about use nobody approved, the four questions to put to your team, and the four decisions to make about everything on the list.
By Victor Xu, Founder · LinkedIn
Last updated · 2026-10-27
The short version
Assume your team is using AI and go and find out where. Five places give you almost the whole picture in an afternoon, none of them needing an IT department: the third-party apps people have signed into with their work Google or Microsoft account, card and expense records, browser extensions on company laptops, the AI features your approved software switched on without telling you, and the question asked directly in a one to one. The scale is not a guess. Verizon's 2026 Data Breach Investigations Report found that 45% of employees regularly use AI on corporate systems, up from 15% the year before, and that 67% of those users reach it through accounts the company cannot see. The output of the exercise matters more than the detection. A list of what people reached for is a list of the jobs they cannot get done the current way, which is the cheapest demand research a business will ever run, and it is where a sensible AI roadmap starts.
Which AI tools is my team already using?
The tools your team is already using are almost certainly a general assistant, a note taker and something inside software you already pay for. That is the pattern in nearly every business we look at: one or two people on ChatGPT, Claude or Gemini in a browser tab, a meeting recorder that somebody added to a calendar invite, and AI features inside an existing CRM, help desk or accounting package that arrived in an update.
The three categories behave differently and matter differently. A general assistant in a browser sees whatever a person pastes into it, and that is where client data goes. A note taker joins meetings and keeps recordings and transcripts somewhere, which is a records question as much as a security one. Features inside approved software are the quietest of the three, because nobody made a decision to adopt them and your existing contract may or may not cover them.
Underneath those sit two that most owners miss. Browser extensions, because an extension does its work inside the page rather than on a website you would see in any log. And anything a developer or a technical person has wired up, including coding assistants and the small scripts people build for themselves, which tend to have the broadest access of anything on the list.
None of this means somebody did something wrong. People reach for tools when the work is heavier than the time available. Which is why the list is useful beyond the security answer: it tells you where the work is heaviest, in the words of the people doing it.
How many people are actually using AI without telling anyone?
More people are using AI without telling anyone than most owners expect, and the surveys converge on a figure between a half and three quarters. Verizon's 2026 Data Breach Investigations Report, summarised by eSecurity Planet on 21 May 2026, found that 45% of employees are now regular AI users on corporate systems whether approved or not, up from 15% the previous year, and that 67% of those users accessed AI services through non-corporate accounts on company devices.
The employee-side surveys land in the same place. The Upwork Research Institute's 2026 Future Workforce Index, a survey of 2,400 United States workers run in March and April 2026 and written up by Upwork, found 55% of full-time employees admit using unapproved AI tools at work. A Cybernews survey of more than 1,000 United States employees in 2025 found 59%. And Resume Now, in a release of 24 June 2026, reported 76% of 1,020 employed United States adults surveyed in May 2026 had used AI tools they personally found and signed up for. Different definitions, same direction.
Set that against what companies say about themselves. The United States Census Bureau's Business Trends and Outlook Survey, reviewed over data collected from 14 December 2025 to 3 May 2026, found overall AI use hovering between 17% and 20% of businesses, and under 20% among firms with four or fewer employees. Both numbers can be true at once, because they measure different things: the Census asks the business what it uses, and the other surveys ask the people what they do. The gap between those two answers is the thing this exercise is looking for.
Where do I look first if I do not have an IT department?
Look in five places, in this order, and the first two cost nothing and take about an hour. No security product is needed for any of them.
Third-party app access in your Google Workspace or Microsoft 365 admin console. Most browser based AI tools ask people to sign in with their work account, which creates a record of that grant and what it was allowed to read. This single list is usually the biggest surprise in the exercise.
Card and expense records. Search the last six months of statements and expense claims for AI vendor names and small recurring charges. Personal subscriptions reimbursed as expenses show up here and nowhere else.
Browser extensions on company laptops. Check what is installed, because an extension works inside the page and leaves no trace in the places people usually look. Do this on managed devices only, tell people you are doing it, and take advice first if you are in a jurisdiction with employee monitoring rules.
The approved software you already pay for. Open each tool's settings and read what AI features are switched on and what data they use. Approving a product in 2024 did not approve the features it shipped in 2026, and this is the category most likely to be touching customer records.
The conversation. Ask people. More tools surface in a ten minute one to one than in any automated sweep, especially the ones running on personal accounts and personal devices, which no company system can see at all.
What should I actually ask my team?
Ask your team four questions, and ask them in a way that makes an honest answer safe. The useful version of this conversation is about the work rather than about compliance, because a person who thinks they are being investigated will tell you about the tool they are allowed to use.
The four: which AI tools do you use for work, what do you use each one for, what do you paste into it, and what would you stop doing by hand tomorrow if it kept working. The last one is the question nobody asks and the one that pays for the exercise.
Say at the start that nobody is in trouble. This is one place where a security framing actively costs you information. If you open with policy, the honest answers go underground and you are left relying on logs that cannot see personal accounts, which is where two thirds of the activity sits according to the Verizon figures above.
Write down the answers in the person's own words rather than tidying them into categories. "I paste the client's email thread in and ask it to pull out the dates" is a specification. "Uses AI for admin" is not, and the difference matters when you later decide what to build.
One practical note on scope: ask about customer data, contracts and anything under an NDA specifically, by name. People do not think of a client email as customer data until the question is put that directly. Our page on whether AI is safe with your business data covers the controls that follow, and whether a small business needs an AI policy covers the document.
What do I do with the list once I have it?
Sort every tool on the list into one of four decisions, and make all four on the same day so the exercise ends in something. Leaving a list unsorted is how this turns into a filing cabinet item.
| What you found | Decision | What it means in practice |
|---|---|---|
| Useful, no sensitive data, cheap | Keep it, write it down | Put it on an approved list with who uses it and for what, and move on |
| Useful, but seeing customer or client data | Keep the job, change the tool | Replace the personal account with a business account on terms you have read, or move the work into something scoped and logged |
| Doing a job three or more people do every week | Build or buy it properly | This is a system, not a subscription, and it is where the hours are |
| Risky and replaceable | Stop it, with a substitute ready | Removing the tool without replacing the job sends it underground |
The third row is the one worth the afternoon. A job several people are doing by hand with a chat window, every week, is an identified bottleneck with a volunteer user base and a worked example of the output, which is a better starting brief than any workshop produces. When we built the specimen database for the Universidad Maimónides 3D-scanning lab, the project started from exactly that shape of observation, 15 hours a week of manual data work that the people doing it could describe precisely, which came to about $39,000 a year of skilled time at $50 an hour. The case study shows what the finished thing looked like.
What the list is not is a scorecard on your team. Treat it as inventory plus demand research, keep it somewhere it gets updated, and re-run the five checks every quarter, because the approved software column changes with every release.
Should we block the tools we find?
Blocking is the right answer for a small number of the tools you find and the wrong answer for most of them. Block where there is a specific, nameable risk: a tool with no business terms that is seeing regulated data, anything that keeps recordings you cannot delete, and anything a client contract forbids.
The case against a general ban is in the breach data. IBM's Cost of a Data Breach Report 2026, which analysed 602 breaches across 17 industries between March 2025 and February 2026 and is summarised by Baker Donelson, found incidents involving shadow AI more than doubled to 43%, from 20% a year earlier, and cost more when they happened, $5.39M against $4.63M. The part that decides the argument is what those organisations had in place. Of the ones that suffered an AI-related incident, 92% lacked proper AI access controls and 68% lacked AI governance policies, itself up from 63% the year before. The organisations in trouble were not the ones whose staff used AI. They were the ones with no sanctioned path and nothing watching.
A ban also has a cost that does not appear in any report. The work does not stop, so it moves to a phone, a personal laptop or a home account, where you can see nothing and delete nothing. You have traded a visible problem for an invisible one.
The sequence that works is a sanctioned option, then a short written rule, then enforcement. Give people one approved assistant on business terms, say in a paragraph what may and may not be pasted into it, and then remove the rest. Enforcement without a substitute is the step that fails, and how to get your team to use AI covers the adoption half of that.
Frequently asked questions
How much does it cost to find out what AI tools we are using?
Nothing but time, if you do the five checks above. An afternoon in the admin console, an hour in the card statements, and a round of short conversations covers a business of up to a couple of hundred people. The paid discovery products on the market are built for enterprises with a security team to run them, and they find things the five checks miss, mostly inside developer environments and API traffic. For an owner-run firm, the limiting factor is deciding what to do with the list rather than producing it.
How do I find AI use on personal devices and personal accounts?
You cannot see it directly, and this is the honest limit of the exercise. Company systems only show company accounts, and the Verizon figures above say two thirds of AI users reach the services through non-corporate accounts. What works instead is asking, in a conversation where the answer is safe to give, combined with being specific about what is not allowed to leave the business. A rule people understand covers more ground than a tool they can route around.
Is a shadow AI inventory the same as an AI audit?
No. An inventory is a list of tools and who uses them. An audit works out where AI is worth applying in the business, what each opportunity is worth, and in what order to do it. The inventory is one input to the audit, and a useful one, because it shows what people already tried.
Can we run this without upsetting people?
Yes, and the framing decides it. Say what you are doing and why, lead with the work rather than the policy, and tell people what happens next. The thing that upsets a team is finding out afterwards that their browsing was reviewed, or losing a tool they depend on with nothing in its place. Announce the exercise, run it in the open, and come back within a fortnight with decisions.
What if we find someone has been pasting client data into a chatbot?
Deal with the exposure first and the person second. Establish what was pasted, over what period, which provider it went to, and what that provider's terms say about retention and training. Then check whether any client contract or regulator requires you to say something, which depends on your sector and your jurisdiction and is a question for your own counsel. Then give that person a sanctioned way to do the same job, because they were solving a real problem and will keep trying to solve it.
What is an AI audit sets out what ours covers, and the inventory section of it is usually the fastest part.
Turn the list into a plan
We map how work and information actually move through your business, put a number on where they leak, and tell you which of the jobs your team is already hacking together are worth building properly. Then our engineering team builds what the roadmap calls for and we run it for you from there. We own the outcome, not the deliverable.