Before you buy anything

AI is already in your business. Not as a project, not as a decision anyone took, but as a habit. Someone in customer services is pasting complaint emails into a free chatbot to draft the reply. Someone in finance has found that the spreadsheet tool now writes formulas for them. Someone in sales has uploaded the price list to get it reformatted.

None of this is on a company account. Nobody has asked what data went in or where it now sits.

This is the starting point for almost every business we talk to. The question in front of you is not whether to adopt AI. It is whether adoption has already started and is currently visible, owned and under control, or invisible and quietly accumulating risk you cannot see.

Most businesses answer that question by buying something to solve a problem they haven't established what it's achieving. A supplier arrives with a platform, a demonstration and a list of the things it could automate. It is impressive. It is also the wrong first move, for a reason that has nothing to do with the supplier.

If you choose a supplier before you have written down what safe looks like, two things happen.

• •You cannot assure what you have bought. Nobody wrote down what adequate control meant, so there is nothing to test the supplier's system against. You are taking their word for it, and every supplier's assessment of your needs looks remarkably like what that supplier builds.
• •And you have chosen the answer before agreeing the question. The use cases that get built are the ones the platform is good at, not the ones the business plan needs. Six months later there is a working system, a monthly invoice, and no clear line from either to the profit and loss.

There is a third problem that sits underneath. A small technology firm running software inside your finance system, your CRM and your email is a dependency you have created. If that firm stops trading, changes its model, or loses its one senior engineer, what happens to the systems it built?

Businesses that start with the product rarely ask that until it matters.

So do not start there. Start with what good looks like.

Governance is a word that makes owners reach for a compliance consultant. Set it aside. What follows is not a framework. It is a list of conditions you could check yourself, in an afternoon, without a technical background.

You know AI is properly run in your business when:

• •One named person is accountable for it, and has the authority to stop something. Not a committee. A person with time set aside for it.
• •Every use of AI is on a register, with four things recorded against it: who owns it, what data it sees, what it is allowed to do, and when it is next reviewed. Not a list of software. A list of uses, because the software changes under you every month.
• •Every use is on a company account. Personal accounts mean the business has no visibility, no control over data retention and no way to turn it off.
• •There is a one-page acceptable use policy that people have actually read, in the language of the business rather than the language of legal, and everyone with access has done the thirty minutes of training that goes with it.
• •Use cases come from the business plan, not from what the technology can do. Each one is ranked on value, on feasibility, and on what happens when it gets it wrong.
• •Nothing goes live without four things. A named owner. A written boundary on what it may touch and what it may do. One or two measures of failure with a threshold set before launch, so somebody knows when to stop it. And a way to stop it.
• •Anything that writes into a business system has been tested by someone who did not build it. The people who designed the controls should not be the only people deciding whether the controls are adequate.
• •The board sees one page a month. What is running, what it touched, what it cost, what it saved against a baseline recorded before launch, and what went wrong.
• •You could turn all of it off tomorrow and still trade.

That last one is the test that matters most, and the one nobody asks.

The mistake in most governance thinking is to treat every use of AI the same way. A tool that drafts an internal email does not need the scrutiny of a tool that changes a price or sends a message to a customer. Treating them alike either buries the business in process or leaves the dangerous ones under-controlled. Usually both.

Three tiers are enough.

• •Low. The tool drafts, summarises or advises. A human checks everything before it is used. Register it, put it on a company account, move on.
• •Medium. The tool influences a decision or writes into a system that is not core: a shared document, a task list, a marketing draft. It needs an owner, a boundary, and a monthly look at what it did.
• •High. The tool touches money, personal data, a regulated obligation or a customer, or it acts without a human in the loop. Independent testing before launch. Failure measures with thresholds. The owner re-signs whenever the supplier changes the model, the terms or the way data is used.

The tier decides how much assurance a use needs. Decide the tier first and the rest of the argument about control mostly resolves itself.

Four pieces of work. They overlap, and they should, because people will not tell you what they are using if they think the only outcome is losing it. Show them the business wants their ideas at the same time as it asks what they have been doing.

1. 1.See it. Find out what is actually in use. Every tool, every account, every kind of data, in every part of the business. Do not rely on IT to discover it: expense claims, software logins and a confidential declaration from staff will find what a network scan will not. Two to three weeks. The output is one register.
2. 2.Contain it. Move everything onto company accounts. Turn off what should not be there. Publish the one-page policy. Name the owner. This is the cheapest risk reduction available to you and it needs no supplier and no budget beyond people's time. Start it as soon as the first high-risk use surfaces, not when the audit is complete.
3. 3.Choose it. Take the business plan and ask where AI removes real cost or wins real revenue. Score each candidate on value, feasibility and cost of failure. Pick two. For each, record the baseline now: hours per transaction, error rate, cycle time. Without that number, nobody will ever know if it worked.
4. 4.Build it under control. Write the control specification before you speak to a supplier. Then run one thing, with a named owner, a defined boundary, failure measures and a monthly report. Prove it, then widen. Twelve weeks is enough to know.

The one gate that does not move: no build starts, and no supplier is selected, until the owner is named, the policy is published and the tiers are agreed. Everything else can run in parallel.

Three roles, and they should not be the same person.

• •An executive sponsor who owns the risk appetite and the budget, and who outranks anyone whose project might need to be stopped. In most businesses that is the chair or the chief executive.
• •An operational owner who runs the register, the policy and the monthly report, with time protected for it in writing. Give someone this job on top of a full one and it becomes nominal inside a quarter.
• •A named person in each business unit who owns the uses in their area and the benefit they are supposed to deliver. The centre sets the standard. The units live with it.

The question that decides whether any of this is real: if the operational owner needs to stop a divisional director's pet project, can they? If the answer is no, fix that before anything else.

Suppliers belong in the fourth piece of work, after the specification exists, and there should be more than one of them.

The specification is the list of controls you are commissioning. Identity and accounts. Data boundaries. What the system is permitted to do and, more importantly, what it is not. Logging. Monitoring. A stop switch and a tested rollback. Security testing that includes the attacks specific to AI, such as instructions hidden in a document or an email that the system then follows. Change control when the model updates. What happens when it goes wrong.

Alongside the technical list, the commercial terms that protect you if the relationship ends: your data is not used to train their models, you own what is built for you, the code is held in escrow, and there is a documented route to run without them.

Then the question that tells you most about any supplier. Ask them to walk you through the most complex system they have put into live production, from the original workflow to what it does today. Ask what went wrong during deployment and what they changed. A team that has done this will talk immediately about messy things: permissions that did not behave, dirty data, users working around the controls, unexpected costs. A team that has not will describe the architecture that ought to exist.

One page. Monthly. It says what is running and at what tier, what each thing touched, what it cost, what it saved against the recorded baseline, any threshold that was crossed, and anything that was stopped and why.

If a report cannot say what was saved against a baseline, the number is a guess. If it cannot say what went wrong, nobody is looking.

Name the owner. Start the register. Move the personal accounts. Everything else follows from those three, and none of them cost anything but attention.

Then ask the question that should sit above every AI decision the business makes:

"If we switched this off tomorrow, could we still trade?"

If the answer is yes, you are in control of it. If the answer is no, you are not, whatever the supplier's slide deck says.