Almost every organisation we speak to this year ends up at the same question. Not whether AI is going to matter for their work, because that debate is over. The question is more practical and more awkward: how and where do we start?
It is a good question, and the reason it keeps hanging around is that people look for the answer in the wrong category. They look for a tool, while the answer sits in a work process.
The starting point is the problem, not the technology
The pattern we see most often: an organisation buys licences, switches something on across the board, gives a demo and waits. A handful of enthusiasts get going, everyone else takes one look and goes back to yesterday's work. Six months later an adoption figure lands on the table that nobody is happy with, and the conclusion tends to be that AI "isn't quite there yet".
That is almost never true. What is missing is not functionality, it is a concrete assignment. A user who opens an empty chat box wondering "what shall I do with this" invents something arbitrary, is unimpressed, and does not come back. That same tool in the hands of someone who knows exactly which part of their work they want to tackle delivers a result within a week.
Handing out more licences does not fix that. Unused seats are, in practice, the biggest hidden cost in AI projects, bigger than the consumption everyone worries about.
Start with one work process
So our advice is always the same: pick one work process and leave the rest alone for now.
Not "we are going to do something with Copilot", but "we are going to tackle the preparation of quotes". Not "we are rolling out AI", but "we are taking on the recurring service desk report". A work process has properties that a tool does not: there are people who do it, it happens at a certain frequency, it costs a certain amount of time, and there is a standard for when it is good enough. That makes it measurable, and measurable is exactly what you need to make a substantiated decision in two months rather than a decision based on gut feel.
When choosing, we look at three things. The process has to occur often enough that eight weeks produce something visible. There should already be some irritation about it, because that supplies the motivation. And it should not sit right on top of the organisation's most sensitive data, because in the first round you want to learn how this works, not explain how it went wrong.
What a pilot looks like with us
Short, small and bounded. Eight weeks, eight to ten participants, a budget ceiling set in advance, and success criteria written down before anyone starts.
Week 1 is setup: access, cost control, and a half-day kick-off in which participants do not learn everything the tool can do, but how to put their own work into it. Weeks 2 through 7 are simply work, with a short check-in every two weeks. Week 8 is the evaluation and the decision.
Each participant records one line per use: what did I have it do, what came out, and how much time did it save. It does not need to become more of an administrative burden than that, and it cannot be less, because without those lines all you have at the end are opinions.
The criteria we propose as standard:
| Criterion | Target |
|---|---|
| Usage | At least 7 out of 10 participants are still actively using it in week 6 |
| Return | At least 2 use cases where the time saved is demonstrably more than an hour per run |
| Quality | At least 7 out of 10 results are usable without significant correction |
| Cost | The cost per usable result stays below the value of the time saved |
| Control | The ceiling that was set is not exceeded, and there are no unexpected spikes |
Eight weeks is not an arbitrary number. Shorter than that and you are mostly measuring enthusiasm, which fades. Longer than that and the pilot dilutes, quietly becoming the new normal without anyone ever having taken a decision.
No one-size-fits-all: four outcomes we see in practice
This is the heart of why we do not arrive with a single answer. Which tool fits only becomes clear once you know what the work actually asks for. In practice we end up at four outcomes, and they differ enormously in cost and complexity.
It works fine with the Copilot you already have. This is by far the most common outcome and at the same time the least discussed. Chat, Notebooks, Researcher and Analyst are already in the licence and cost nothing extra. Roughly four out of five use cases people come up with fit inside this. That is not a brake on ambition, it is the cheapest way to deliver on it.
The work is too big for one question: Copilot Cowork. When an assignment consists of several steps, for example going through twenty documents, pulling out the deviations, putting them in a table and summarising them, then Copilot Cowork is the right place. It is billed per executed task in Copilot Credits, so at that point cost control does become a topic.
The work comes back every week and always runs the same way: build an agent. Repeat work does not belong in an assignment that gets rephrased every time, it belongs in an agent in Copilot Studio. Every time someone writes out the same assignment again, you pay again for reasoning that was already done. From roughly ten times a month, an agent pays for itself.
There is no off-the-shelf product for it: then we build it. Sometimes the gain sits in software that does not exist yet: a portal, a connection between systems that do not talk to each other, an application around a process that fits no package. For that we have our own development pipeline, with uCompose for websites and portals and uBuild for custom applications, in which AI agents do the building under human supervision. That makes custom software viable in situations where it used to be simply too expensive.
The rule of thumb we give users, in this order:
- Can it be done in one question? Do it in Copilot Chat, that costs nothing extra.
- Is it about a fixed set of documents? Make a Notebook, also free.
- Is it research or calculation? Use Researcher or Analyst, still included.
- Is it a multi-step assignment you do once or occasionally? Then Cowork is the right fit.
- Do you run that same assignment ten times a month or more? Then it should become an agent.
- Does the tool not exist yet? Then it becomes software.
Cap the cost before the first user starts
The moment you move into consumption-based territory, the question changes from "what does a licence cost" to "what is the most this can cost". That is a setting, not a hope.
Watch the difference between two controls that are easily confused. A spending limit in the Microsoft 365 admin center is a hard boundary: when it is reached, consumption stops. A budget in Azure Cost Management only sends an alert. Set only the latter and you do not have a ceiling, you have a smoke detector. So we always set the hard limit first, and use alerts at 50, 75 and 90 percent alongside it as a signal.
What we explicitly do not advise is locking everything down out of caution. People will then move to free AI services outside your environment, where you have no visibility, no logging and no grip on what happens to company data. That ends up costing more than a consumption bill, you just do not pay for it in euros.
"It doesn't work here" is a perfectly good outcome
The best part of this approach is that a negative result is as useful as a positive one. If the criteria are not met, you have learned for a few hundred euros and eight weeks that this particular process is not suited to it right now. You move on to the next process, or you wait until the technology has moved on.
Compare that to the cost of an organisation-wide rollout that stalls after six months, including the licences that keep running and the trust you lose with your people. Starting small is not being cautious, it is simply cheaper learning.
How we pick this up together
We help organisations with exactly this step: choose one work process, set up a short pilot with ceilings and criteria, bring the people along, and at the end establish together what the next step is. Sometimes the answer is that you already had everything you needed. Sometimes it is Cowork or an agent. And sometimes we build something that did not exist yet.
Want to know which process is the best one for you to start with? Get in touch for a one-hour conversation. It will not produce a quote, but it will produce a shorter list.
