The hidden cost of letting staff pick their own AI tools

Why letting every employee pick their own AI tool costs more than it saves

Most small businesses never decided to adopt artificial intelligence. It arrived anyway, one employee at a time.

Someone in accounts started drafting supplier emails with a chatbot. Someone in sales began summarising client calls. The bookkeeper found something that reads invoices and pulls the totals out. None of it went through a decision, and none of it appears on a statement, because the free tier of most of these tools does enough to be useful.

By the time an owner notices, the business is running on six or seven tools that nobody chose, nobody administers and nobody can switch off.

The cost is not the subscriptions

Owners who go looking usually expect to find wasted spend. They rarely do. Free tiers are genuinely free, and the paid seats are small enough to disappear into a card statement.

The real cost sits in three other places.

The first is rework. Two people doing the same task with two different tools produce two different formats. The quote that comes back from one looks nothing like the quote that comes back from the other. Somebody then spends an hour making them match, which is an hour the tools were supposed to save.

The second is verification. A tool that is right most of the time still has to be checked every time, because you cannot tell from the output which time this is. Staff who trust the tool stop checking. Staff who do not trust it check everything twice. Neither group is faster than they were before.

The third is the one owners find last. Work quietly moves to whoever is best at the tool rather than whoever should be doing the job. The person who writes the best prompts starts handling the client correspondence, regardless of whether they know the client.

What goes into these tools is a business decision, not an IT one

There is a second problem underneath the first, and it is harder to see because nothing visibly breaks.

To get useful output, staff paste in whatever the task needs. A supplier contract. A customer complaint with a name and address in it. A payroll spreadsheet. A photograph of an invoice. The tool needs context to be helpful, so people give it context, and the context is your clients.

Nobody is being reckless. They are doing the thing the tool asks them to do.

This is where the free tier stops being free in the way that matters. Consumer plans and business plans often handle data differently, and the difference is rarely visible inside the product. An employee cannot tell from the screen whether the material they just pasted sits inside your account or somewhere broader. Most of them have never thought to ask.

The fix is not a policy document that lives in a shared drive. It is picking the tools, paying for the tier that gives you administrative control, and telling people which ones they are allowed to use for work that involves a client.

Start with the task, not the tool

The businesses that get value out of this do one thing differently. They start from a task they already understand and can already measure, and then ask whether a tool improves it.

That sounds obvious. Almost nobody does it, because the marketing for these tools is organised around capability rather than around work. You are sold a thing that can summarise, draft, extract and analyse. You are not sold a thing that shortens the time between a job finishing and an invoice going out, which is the problem you actually have.

Pick one task with an owner, a frequency and a number attached to it. Quotes that take two days to go out. Invoices that sit in a pile for a week. Enquiries that arrive after five in the afternoon and get answered the following morning.

Then measure it for a fortnight before you change anything. If you cannot say what the task costs today, you will not be able to say whether the tool helped, and you will end up arguing about impressions.

Where accountants landed, and why it is worth copying

Accounting is worth watching here, because it went through this earlier and more thoroughly than most sectors.

Accountants had the same sprawl everyone else had. What changed is that their work has hard edges: a reconciliation either balances or it does not, a lodgement is either on time or late, and a client file is either complete or it is missing something. That makes it unusually easy to tell whether a tool is helping, and unusually expensive to be wrong.

So the profession converged fairly quickly on a smaller set of tools that plug into the ledger rather than sitting beside it, and on a rule that client data does not go into anything that has not been approved. Anyone deciding what to standardise on can save themselves months by looking at which AI tools accounting firms have actually settled on before running their own trial from scratch.

The transferable part is not the specific tools. It is the shape of the decision. Choose few, connect them to the system where the work already lives, and be explicit about what may be put into them.

Three questions before you buy anything

Who owns this task today? If the answer is that everyone does it a bit, a tool will not fix that. It will give you the same confusion faster.

What happens when it is wrong? A tool that drafts an internal summary and a tool that drafts something a client reads carry very different risks. The second one needs a person in front of it, every time, and you should decide who that person is before you buy.

What goes into it? Write down, in one line, what staff are allowed to paste. If you cannot write that line, you are not ready to deploy the tool, and neither is your team.

The boring version works

The businesses doing well with this are not the ones with the most tools. They are usually running two or three, chosen deliberately, connected to the systems they already use, with one person responsible for each.

That is a far less exciting story than the one the market is telling. It is also the one that survives contact with a normal week, a sick employee and a client who wants their quote today.

The technology is genuinely useful. It just does not arrange itself, and waiting for it to do so is how you end up with seven tools, no owner and an hour a day spent making outputs match.