Is a Copilot or ChatGPT licence enough?
Giving your staff a Copilot, ChatGPT or Claude licence gets people using AI. It doesn't get your business using it. Here is what's missing and what becomes possible when you build a harness around it.
It's enough to get people using AI. It isn't enough to get your business using it. A Copilot, ChatGPT or Claude licence gives each person a clever assistant. Copilot can even read your files. None of them has been taught your rules: how you price, how you word a quote, what you never say. So the answers come out generic, the drafts need fixing every time and what one person learns stays with them. An administrator can see who is using it. Nobody can easily see whether the work followed your rules. The answer is to keep the licences and build what's called a harness around them, which means written rules, a memory you can read, checks that run every time and jobs that start without being asked. Because all of that lives in files you own, you can change AI provider later without starting again.
Buying everyone a licence feels like the job’s done. Each person has a Copilot, ChatGPT or Claude login, someone has sent the email telling them to give it a go and off they go. Months later the answers are still generic, the drafts still need rewriting and nobody can tell you what difference it has made. The licence is fine. It’s what isn’t around it that causes the trouble.
Why does AI need fixing every time?
Because it hasn’t been taught how you work. The AI behind a licence has read a vast amount of the internet. Copilot can read your files and emails as well, which helps it find things. But nobody has told it which price list is current, how you word a quote, which customer gets a phone call instead of an email or the phrase your managing director can’t stand. So it writes the average of everything it has read, which is why it sounds fine and sounds like everyone else.
Which one you buy matters less than you’d think. Copilot makes most sense if your business already runs on Microsoft 365. ChatGPT and Claude are strong general assistants. All three start with the same gap.
Memory helps less than you’d hope. ChatGPT, Claude and Copilot all remember things now, but what they keep are notes about one person, written in the AI’s own way. Think of a temp who remembers your name and your coffee order and has never seen the staff handbook.
Then there’s being wrong. AI gets things wrong in exactly the same confident tone it gets things right, so every output has to be read by someone who would spot the mistake. When they do spot it, they fix the document and move on. The AI is never told. Tomorrow somebody else gets the same mistake.
PwC’s latest survey of 2,023 UK workers found that 44% of people using AI report a heavier workload. It doesn’t say why. A tool that drafts in seconds and then needs correcting every time would be one good place to look. Our prompt, What Doesn’t Your AI Know?, takes the last few things you had to fix and shows you what was missing.
Why isn’t my whole team getting better?
Because what one person learns stays with them. Give a team a licence each and you can easily end up with a few explorers doing clever things and everyone else having tried it twice before going back to the old way. The explorers’ best work lives in their chat history and their heads. When one of them leaves, it leaves too.
Even when everybody is using it, you’ve got ten people doing it ten ways. Ten versions of your tone of voice. Ten sets of instructions, all invisible to each other. Nobody’s improvement reaches anybody else, so the business stands still while individuals move.
I saw a Microsoft Copilot advert this week about exactly this: the person who has been in the business longest and gets asked everything. The pitch was that everyone asks Copilot instead, so that person can get on with their job. It tells you the problem is real. But a tool like that can only find what has already been written down. It doesn’t get the knowledge out of that person’s head or apply your rules to the answer.
How do I control what AI is doing?
With a licence on its own, mostly you can’t. Business plans give an administrator some useful controls over who can log in and how your data is treated. They won’t tell you whether the letter that went out this morning followed your policy.
Some of it won’t be on your licences at all. Personio’s Workforce Pulse survey found that 26% of UK employees use AI without their employer knowing. An AI policy is a sensible start, but a policy is a request and requests get forgotten at ten to five.
I’ve spent over a decade as an internal auditor. The first thing an auditor asks of any process is how you know it happened the way it was meant to. Ask that of a business where everyone has a licence and there isn’t an answer yet.
Use a business plan with a promise not to train on your data and a data processing agreement, which is the contract setting out how they handle it. Don't use personal accounts. Strip out what the AI doesn't need, such as names and account numbers. Where a contract or regulator demands it, keep processing in the UK or EU or run a freely available AI model on a server you control. Our guide to where your AI data goes covers what to check.
What is a harness around AI?
The word comes from horses. A horse is enormously strong and without a harness that strength goes wherever the horse fancies. The harness doesn’t make it any stronger. It points the strength at the cart. That’s what this is, except the horse is an AI model and the cart is your business.
In practice it’s a set of things built around the licences you’re already paying for, which is what an AI assistant built around your business comes down to:
- Your rules, written down. How you price, what you never say, the templates you use, all in plain files the AI reads before it does anything.
- A memory you can open. What happened, what was decided and what got corrected, kept in documents you can read and change.
- Lessons that travel. When one person’s assistant learns something the business should know, the others can get it too.
- Checks that can’t be skipped. The rules that matter most become automatic checks, that we call judges, which run every time whether anyone remembers or not.
- Routines. Jobs that start without anyone asking.
The big change is what happens to a correction. It stops being a one-off fix and becomes part of how the thing works. We told our own assistant never to put a comma before “and”. That’s now an automatic check on everything client-facing, so nobody has to say it twice.
What can it do that a licence can’t?
A licence waits to be asked. You type, it answers, you type again. Turn by turn, every time, with you doing all the remembering. A harness lets the work start without you.
We run this business on one, an assistant called Otto. Every morning there’s a briefing waiting that nobody asked for: what needs a decision, what’s already drafted and the things we said we’d do and may have forgotten.
It also does what a good colleague does and speaks up. It challenges work it doesn’t think is worth doing. It looks at where it went wrong and proposes changes to how it works. It suggests areas of the business worth exploring and avenues we hadn’t thought of. We say yes or no to each one.
It’s the same AI on the same licence. The difference is between a clever stranger and someone who has worked alongside you for a year. It’s the idea behind Smart Agents, our custom AI assistants and agents for UK businesses: an intelligent AI assistant for everyone in the company, built around its own rules.
What happens if I change AI provider?
With a licence on its own, you lose more than you’d think. Claude can import memory from other AI services, although Anthropic describes that as experimental. So the notes may come with you. The way of working doesn’t. The custom set-ups, the routines and months of corrections have been living inside one company’s product.
A harness built properly lives in plain files you own. The AI reads them, whoever made the AI. If a better model turns up or a price changes or you decide sensitive work should run on an open source model on your own server, you point the new one at the same folder. Ours is already read by tools from two different AI companies.
It isn’t effortless. Each tool enforces checks in its own way, so those need rewiring. Open models can’t yet match the top ones on every job. But you’re rewiring, not starting again, which is why it’s worth owning the set-up instead of chasing the best model.
The licence is the part you can replace. What you’ve taught it is the part worth keeping.
- A licence gets people using AI. It doesn’t get your business using it.
- What one person learns stays with them until something shares it.
- A policy is a request. A check that runs every time is a control.
- A harness lets the work arrive drafted before anyone asks for it.
- Keep your rules and memory in files you own and you can change provider without starting again.
- Start with the last three things you had to fix. They’ll show you what your AI is missing.
Drafted by Otto, the Perkins SmartOps AI assistant. Reviewed, edited and published by David Perkins, the human.
