UbertAI engineering for SMEsBook an intro call
Solution

Rolling out Copilot: from licence to working tool

Rolling out Copilot is not an IT job; it is a change in how people work. Treat rolling out Copilot as handing out licences and it will fizzle out within a month: the icon sits in the toolbar, two enthusiasts use it, everyone else carries on as before.

This page is about the rollout, not the tool. What you agree up front, how to start with a small group, why connecting it to your own data is the difference between a toy and a tool, and what goes wrong.

The honest caveat straight away: we do not sell the licences. Not Microsoft’s, not OpenAI’s, nobody’s. So we earn nothing whether you buy ten or two hundred.

Book an intro call Take the free AI scan
4 platforms built by us and running€110 / €165 per hour, fixed hoursInsured for professional and public liability

Buying a licence is not the same as rolling it out

Buying has become the easy part. Your administrator switches it on, billing runs per user per month, and technically you are done. That is where the idea creeps in that the rollout is done too.

  • What you buy: access. An assistant inside the programs you already use, and a window someone can type questions into.
  • What you roll out: a different way of working. Which work changes, who does it, and what that person still checks.
  • What gets forgotten: the rules. Which data may go in, which may not, and what happens to what comes out.
  • What nobody measures: whether it is used. Without agreeing what you want to look different in three months, you can never tell whether it did anything.

Why a rollout without rules fizzles out

It goes the same way almost every time. In week one some people try something. In week two someone gets an answer that looks right and is not. In week three things get busy. After that it stops being a subject.

That is not the tool’s fault. There was no reason. “Give it a go” is not an instruction, and someone under pressure takes the route they know. On top of that, nobody takes a chance while it is unclear which data may go in. That is sensible behaviour, not resistance.

At an accounting firm we saw it in its simplest form: everyone had access, nobody knew whether a client document could be pasted into a chat window, so nobody did. Quite right. You answer that question before the rollout, not after.

What to sort out up front

Five things go on paper before the first licence is handed out. One page everyone reads beats a policy document nobody opens.

  • Who may do what. Does everyone get access, or do you start with one department? Who may connect folders and systems, and who decides that?
  • Which data may go in. Be specific per type: client documents yes or no, personnel files no, third-party documents only anonymised. Five rules work better than “be careful”.
  • Where the answers end up. What is stored, where, for how long, and who can get to it later. This differs per product, per subscription and per setting. Ask your own supplier or administrator and put the answer in writing. Do not rely on a blog post, including this page.
  • What goes into policy. A short usage guideline: what it is for, what must never go in, that the human stays responsible for what they send, and who to ask when in doubt. Update your record of processing activities.
  • Who owns it. One person who treats this as their subject. Without an owner the guideline gets written, published and forgotten.

How a rollout works in practice

Start with a small group that genuinely uses it, and only then the rest. Not because a pilot looks good, but because the first weeks show you which agreements are missing. You would rather learn that with eight people than with eighty.

  1. Decide on the work, not the tool. Which tasks come back every week and cost time without needing professional judgement? At a wholesaler the quote requests arriving by email; at an installation company the job sheets.
  2. Pick eight to twelve people. Deliberately include someone sceptical and someone who has done the work longest; those two find the weak spots.
  3. Put the agreements on paper before anyone starts. The five points above. It costs an hour and prevents most of the trouble.
  4. Practise on real work, on site. No sample case, but the documents that have to be finished that week anyway. Let the model get something wrong on purpose; people remember that better than a warning.
  5. Record what works: a one-page working instruction and saved prompts ready for next time. Without that step it has evaporated after three busy weeks.
  6. Decide after six to eight weeks. What is genuinely running differently? Only then do you expand, drop it, or conclude that what is needed here is not an assistant but an integration.

Connecting it to your own data: toy or tool

An assistant with only general knowledge is a writing aid. An assistant that can reach your own quotes, price lists and project files is a tool. That is the point where people stop saying “nice” and start using it, because it beats searching themselves.

That connection is also the hardest part. Your data sits scattered across email, a network drive, your ERP or accounting package, and the heads of a few people. What can be connected to Microsoft 365, Exact, AFAS, Moneybird or HubSpot differs per version, per licence and per supplier. We check that case by case, and you should ask your own supplier before promising anything to your team.

More important still: an assistant that can reach your files exposes what was already wrong with your permissions. Folders opened to everyone “temporarily”, a drive that also holds payroll. It invents nothing; it simply finds faster what an employee could have found by clicking around. Cleaning up permissions before the rollout is often the largest part of the preparation.

AI literacy and the EU AI Act

The moment you give an assistant to your whole workforce, you are an organisation using AI systems. The European AI Act asks such organisations to take measures so their people have sufficient AI literacy: enough understanding to grasp what the system does, what the risks are and how to judge whether an outcome holds up.

The Act prescribes no particular course, no number of hours and no certificate. So we do not sell a rollout with a compliance stamp on it; take legal advice on what applies to your organisation. What is practical: record who received which explanation. Our page on AI training for companies sets out what that looks like. The rollout and the literacy are the same project: people who cannot tell when an answer is unreliable end up checking everything, and then the time saved is gone.

What goes wrong when rolling out Copilot

Four things keep coming back. None of them is a software problem.

  • A licence for everyone and a reason for no one. The expensive version: you pay per user per month for access that is barely used.
  • Sensitive data in the wrong environment. Someone pastes a file into the first chat window they find online, because the business environment was never explained. Almost always well meant, and rarely undone.
  • Expectations that do not materialise. After a demo, management expects the back office to free up a day a week. What actually happens: a few tasks go faster and time is added for checking. Promise the second one and nobody is disappointed.
  • The rollout as a one-off event. A kick-off with no follow-up is an afternoon of enthusiasm. Put three moments in the diary: the start, a review, and the decision on expanding.

Copilot, ChatGPT, or an assistant on your own data

The question “which one should we have” comes too early. More useful is which work you want to change; the choice usually follows from that. An assistant running inside the office programs you already use makes sense when your people spend the day in email and documents. A separate chat window on a business subscription suits you better when you mainly want to write, summarise and look things up. An assistant on your own data is worth it when the value sits in your documents rather than in the model’s general knowledge. What the packages can do right now, and in which subscription, changes regularly: check with your supplier.

Here is the caveat again. We are not a reseller and earn nothing on your licences. That lets us give advice a supplier rarely gives: start with fewer licences than you think you need. You can always add more, and paying for unused access is the most common waste on this subject.

When not to start — and what guidance costs

If the work bothering you is mainly retyping between two systems, an assistant is a detour. Then you build an integration: more expensive once, and after that it runs itself. An assistant makes manual work faster; it does not remove it.

If nobody knows which data may go in and there is no time to find out, postpone the rollout. A week of waiting is cheaper than an incident. And if all you want is a few people writing email and summaries faster, you do not need us: a subscription, an agreement on what must not go in, and an afternoon of practice together will get you further than a project.

If the answer is yes, these are our rates. They do not depend on which licences you buy.

  • The AI scan: free, five minutes online. You get a score, a few concrete opportunities and a price indication.
  • A core session or training half-day (4 hours) at your location: €596 excluding VAT and travel, a full day €1,192. Our advisory and guidance rate is €149 per hour, travel €0.45 per kilometre.
  • Building — integrations, an assistant on your own data, automation: €110 per hour for straightforward work, €165 for complex development, drawn each month in advance from a budget you set.
  • Outside us: the licences themselves from your own supplier, plus your own people’s hours. Freeing up twelve people for a half-day is often the largest item.

Frequently asked questions

Do you sell Copilot or ChatGPT licences?

No. We are not a reseller and earn nothing on the number of licences you buy. You get those from your own IT supplier or directly. We only guide the rollout, which is exactly why we dare to say you can start with fewer licences.

How many licences do we need to start?

Fewer than you think. Start with eight to twelve people who use it daily, each with a reason and an agreement attached. After six to eight weeks you will know whether expanding makes sense. Giving everyone a licence at once is the most expensive way to find out nobody had a reason.

Can we put client data into Copilot or ChatGPT?

That depends on the product, the subscription and the settings in your environment, and it changes regularly. Ask your supplier what happens to your input and where answers are stored, and put that answer in writing. Then make a short list of what is and is not allowed, and add it to your GDPR records.

How long does rolling out Copilot take?

Count on an hour for the agreements, a half-day of practising on real work with the first group, and then six to eight weeks before you can judge what is running differently. The wider rollout follows after that. If it takes much longer, there is usually no owner, or no concrete piece of work was chosen.

What if nobody uses it after the rollout?

That is the normal outcome of a rollout without agreements, and it can be repaired. Usually one of three things is missing: a concrete task, clarity on which data may go in, or someone explaining what still has to be checked. We call that adoption support: looking over shoulders in the first weeks and recording what works.

Is an assistant the same as process automation?

No, and the difference decides what you should pick. An assistant helps a person with work that differs each time: writing, summarising, looking things up. Automation takes over work that is the same every time, such as processing incoming order confirmations. For repeat work an integration is more reliable and cheaper in the long run.

Do we have to do something because of the EU AI Act?

Organisations using AI systems must ensure sufficient AI literacy among their staff. The Act prescribes no specific course, no number of hours and no certificate. At the very least, record who received which explanation and what your usage rules are. For what exactly applies to your organisation, take legal advice.

What we do for this

AI scan & strategy

Most AI projects don't fail in the build, but in the choice of what to build. That's why we don't start with a tool, but with your

AI guidance & training

A system nobody uses delivers nothing. That's why we don't leave your team alone with a tool: we train on the shop floor, with you

Further reading

6 min read
GDPR and AI: what can you do with customer data?

The GDPR does not prohibit AI, but it does set requirements for purpose, legal basis and processors. Most mistakes at SM

6 min read
Why AI projects in SMEs fail

Projects rarely fail on the technology. They fail on messy data, a pilot that never ends and nobody who owns it.

8 min read
Connecting AI to your existing software: Exact, AFAS, Microsoft 365 and the rest

AI only becomes useful once it can reach your own data. What an integration is, the four ways to get at your data, and w

Other topics

AI implementation for small business workflow automation for small business AI consultant for small business automating quotes automating invoice processing automating customer contact having an AI agent built AI training for employees AI policy for business

First find out whether it pays off for you

Take the free AI scan: five minutes online, no sales call. Want to set it up properly straight away, book a core session — a half-day on your own shop floor (€596 excluding VAT and travel). We will not sell you licences; we make sure the ones you already have are good for something.

Book an intro call Take the free AI scan