The Tools Nobody Filed: Auditing Unapproved AI Use in Your Business

The tool that arrived without a purchase order

Somewhere in your business this week, someone pasted something into a chatbot. It may have been a draft tender response, a supplier's latest terms, a rambling customer complaint, or a spreadsheet column of order references. They did it because it saved twenty minutes and because no one had explained why they should not.

This is not a story about reckless employees. It is the ordinary consequence of capable, free software meeting a busy team with deadlines. The productivity gain is often genuine. So is the exposure.

Shadow information technology is not new. Spreadsheets ran finance departments for years before anyone called it a governance problem. What has changed is the speed and the intimacy of it: a large language model can now be handed a contract, a customer list or a set of payscales, and it will answer instantly, cheerfully, and without asking whether it should have been told.

For a managing director, the useful question is not whether this is happening. It is where, how often, and with what kind of information. That is an audit question, not a disciplinary one.

Good intent is not the same as good governance

Three patterns tend to recur in businesses of twenty to two hundred people.

First, the tools are consumer-grade. Free tiers of well-known assistants, browser extensions, transcription services that convert a client call into minutes, and image tools that remove a background from a product photograph. None of these has been procured, none has been security-reviewed, and most are governed by terms written for individuals rather than organisations.

Second, the use is invisible to the person who owns the risk. The marketing team knows about the copywriting assistant. Finance knows about the spreadsheet macros. Nobody knows about both, and nobody has asked whether the copywriting assistant was ever shown a customer's address.

Third, the intentions are sensible. An operations manager summarising long supplier emails, a salesperson drafting follow-ups, a bookkeeper extracting totals from a scanned invoice. Each is a reasonable attempt to do more with a small team. Each is also outside any policy, because there may not be a policy yet.

A blanket ban is the tempting response and usually the wrong one. It drives the practice further underground, removes the productivity gain, and leaves you with less visibility than you started with.

What UK data protection law actually asks

You do not need a technical education to grasp the obligations. Under the UK GDPR and the Data Protection Act 2018, four questions matter when personal data leaves your systems.

Is there a lawful basis? If an employee feeds customer information into a tool for a business purpose, that processing needs a basis, and your privacy notice should not contradict it. Legitimate interests often applies, but it must be assessed rather than assumed.

Who is the processor? You remain the controller. If the tool's provider is storing or learning from the data, you need a contract that reflects that, and a free consumer account almost certainly does not provide one.

Where does the data go? Many services process data outside the UK. Transfers need appropriate safeguards. This is a question for the supplier's documentation, not for guesswork.

Do you need an assessment? Some uses, particularly large-scale or sensitive ones, call for a data protection impact assessment. The Information Commissioner's Office publishes guidance on AI and data protection that is readable without a legal background, and it is worth an hour of someone's time.

There is a fourth consideration that is not strictly legal but matters commercially: confidentiality. If your client contracts contain a clause about keeping their information secure, a chat window is a disclosure.

Step one: a discovery conversation, not an investigation

Announce the review openly and frame it as a mapping exercise. Say plainly that no one is in trouble, that the aim is to understand what is working, and that any tool found to be unsuitable will be replaced with something approved rather than simply withdrawn.

Then ask a short set of questions in team meetings, not by questionnaire alone:

• Which tools do you use that help you write, summarise, transcribe, translate or analyse?
• What kind of information do you normally put into them?
• What did you do before you had them?
• What would you miss most if it disappeared tomorrow?

That last question matters. It tells you which workflows are worth formalising, and it tells the team that their judgement is being taken seriously.

A useful cross-check is the browser history of shared machines, the software list on finance's payment records, and any expense claims for subscriptions. Expect a handful of surprises, usually harmless.

Step two: rate the data, not the tool

The instinct is to rank tools as safe or unsafe. A better method is to rank the information being shared and let that determine the answer.

Tier one — public. Published marketing copy, press releases, generic research. Low risk, minimal restriction.

Tier two — internal but impersonal. Meeting notes with no names, process documentation, job descriptions. Moderate; acceptable in an approved tool with a business account.

Tier three — personal data. Customer names and addresses, employee records, contact details, complaint correspondence. Requires an approved tool, a proper contract, and a clear business purpose.

Tier four — sensitive or contractual. Special category data, financial details, signed agreements, pricing schedules, anything covered by a confidentiality clause. Default position: not permitted, unless a specific workflow has been formally agreed and recorded.

Place every use you discovered on this grid. Most will be tier two. A few will be tier three. One or two will be tier four, and those are the conversations that need to happen this week rather than next quarter.

Step three: a one-page approved-tools policy

Keep it to a page. Longer documents are not read.

State the purpose in a sentence. List the approved tools by name. State the tiers above in plain language. Say what must never be entered. Explain how to request a new tool, and commit to a turnaround time, so that the answer is rarely no and usually yes within a fortnight. Name the person who owns the list.

Include two practical rules that prevent most incidents: turn off any setting that allows a provider to train on your inputs, and never use a personal account for business information.

Step four: formalise two or three workflows

Formalisation is where the value sits. Pick the two or three uses with the clearest benefit, build them properly, and let the rest of the organisation see how a good one looks.

Typical candidates in an SME: drafting routine customer correspondence from a template and a short brief; summarising long supplier or client emails into a standard note; extracting invoice data into your accounting system with a human check before posting.

For each, document what goes in, what the tool is, who reviews the output, and what must never be included. A single page per workflow is enough. The review step is not bureaucracy; it is the control that makes the productivity gain safe to keep.

Give the review a rhythm

Shadow use reappears whenever a new tool becomes fashionable, so treat this as a repeating exercise rather than a project. A short review each quarter, attached to an existing management meeting, is sufficient. Ask the same four questions, update the approved list, and note anything that has changed.

One person should hold the list. It should not be a committee. A named owner with a fortnightly half-hour is more effective than a working group with a mandate.

The point of the exercise

The aim is not a tidy inventory of software. It is to keep the productivity that your team has already discovered, while removing the small number of uses that could become a serious conversation with a client, a regulator or a customer.

Most businesses find that eight out of ten uses are harmless and worth keeping. The audit exists to find the other two, and to make sure that when someone reaches for a new tool next month, they know exactly where to look for the answer.