Start with your work, not with AI tools. Find out where your team's hours go, pick the two or three workflows that cost the most, and change them together with the people who do them. Measure each one before and after the change. Then do it again. That is the whole method, and it works whether your company barely uses AI today or has already tried a dozen tools.
The rest of this article explains why this order matters, what a good first workflow looks like, and what you should have in your hands after the first eight weeks.
The problem is not the technology
Almost four years into the generative AI wave, almost every company uses AI somewhere. In McKinsey's 2025 global survey of nearly 2,000 organisations, 88% use AI in at least one business function. But only 39% report any effect on EBIT, and for most of those the effect is below 5% (McKinsey). A preliminary 2025 report from MIT's Project NANDA reached a similar conclusion from a different angle: most organisations in its research saw no measurable return from generative AI, and only a small share of custom tools ever reached daily use (summary).
It is tempting to read these numbers as a verdict on the models. They are not. The models are good enough for most of the work a mid-size company does every day: reading documents, checking them against a list, drafting replies, filling systems, preparing reports. The gap sits elsewhere.
What usually happens is this. A company buys licences, or runs a pilot, or appoints an AI champion. Individuals start to use the tools, and some become much faster at their own tasks. But the work around them does not change. The same email still arrives in the same shared inbox. The same document still moves from one desk to the next. The same report still gets rebuilt by hand at the end of the month. A tool on top of an old workflow gives you an old workflow with a tool on top.
Personal productivity is real, but it does not show up in the numbers a board reads. Organisational change does, and it only happens when a workflow runs differently. McKinsey's research points the same way: of the changes it studied, redesigning workflows is the one most linked to an effect on EBIT.
Two kinds of companies, one starting point
When we talk to CEOs and owners of companies with 50 to 500 people, we meet two situations.
In the first, AI is barely in use. Someone rewrites an email with ChatGPT now and then. Leadership knows AI matters, reads about it every week, and feels a little behind. But nobody can say where it would make a difference in this company, and the honest answer to "what should we do?" is "we don't know where to start".
In the second, AI is everywhere and nowhere. People use ChatGPT, Copilot, a note-taker, a translation tool. There have been one or two pilots. There may be an AI policy. And yet, when the board asks what changed, the answer is a collection of anecdotes.
These look like different problems. They are the same problem. In both cases nobody has looked at the work itself: which tasks take the hours, where work waits, where people copy and check by hand. The first company does not need to learn what AI can do before it starts, and the second does not need another tool. Both need a map of their own work.
Start with the work: three questions for every team
The fastest way to build that map is to ask every team three questions. You do not need a technical background to ask them, and the answers are more useful than any AI strategy deck.
What takes the most time each week? Not what is most important, but what consumes hours. Teams usually know the answer immediately. In client services it is often reading and sorting email. In finance it is month-end reporting. In operations it is checking documents.
Where do you wait for someone else? Waiting time is invisible in most companies, because nobody owns it. A file sits in a queue for two days because a payslip is missing. An approval waits for a manager who is travelling. Clients notice waiting time long before they notice how fast any single step is.
What do you copy, check or chase by hand? This is where AI is strongest today. Copying details from an email into a system, checking a document against a checklist, chasing a client for a missing item: these are repetitive, rule-based and tiring. They are also the tasks people are happiest to give away.
Ask the people who do the work, not only their managers. Managers know the strategy. The people in the workflow know where the time actually goes, and they are often surprised that anyone asks.
What makes a good first workflow
The first workflow you change sets the tone for everything after it. Choose it for learning and for proof, not for ambition. A good first workflow has five properties.
| A good first workflow | A poor first workflow |
|---|---|
| One workflow, with a named owner in the team | "Use AI across the company" |
| Repeats many times a week and costs hours | Happens once a quarter |
| Has numbers before and after | Is judged on whether it "feels faster" |
| Keeps a person checking the AI output | Is fully automated from the first day |
| Can run on live work within weeks | Is a six-month programme |
The last two rows deserve a word. Full automation on day one is the most common way to lose trust. A workflow where AI prepares and a person checks is safer, faster to build, and teaches the team what the AI gets right and wrong. The check can shrink later, once you have evidence. And speed matters because momentum matters: a workflow that runs in six weeks convinces more people than a plan that promises results in six months.
What a baseline reveals: a worked example
Take a mortgage and lending adviser with 170 people. This example is illustrative, but its shape is typical. New clients send their documents to a shared onboarding inbox. Here is how that workflow runs today, with the minutes per client that the team itself reported and checked against a sample of real files.
| Step | Who | Minutes per client |
|---|---|---|
| Read the email and save the documents | Onboarding | 5 |
| Check the documents against the checklist | Onboarding | 15 |
| Chase missing documents (in six of ten cases) | Onboarding | 6 |
| Retype client details into the CRM | Onboarding | 12 |
| Screen the client against sanctions lists | Compliance | 6 |
| Hand the file to an adviser | Onboarding to advice | 4 |
| Total | 48 |
At 42 new clients a week, that is more than 33 hours of work every week. None of it is advice. And the table shows something the team felt but never wrote down: the two most expensive steps are checking and retyping, and both happen before an adviser even sees the file.
In the target version, AI reads the documents, checks them against the checklist and drafts the email for anything missing. A person approves that email before it goes out. Client details flow into the CRM without retyping. Sanctions screening stays with Compliance, unchanged, because that is a judgement the company wants a person to own. In this example the workflow drops to about 13 minutes per client, including the human checks.
The point is not the number. The point is that the number exists, before anyone builds anything. Without it, every result afterwards is an opinion.
From one workflow to a way of working
One workflow is a project. A company that changes how it works runs the same cycle again and again. We use six steps. The first cycle takes about eight weeks; after that, steps four to six repeat every few weeks.
- Vision. Leadership says where AI should take the company in the next twelve months, which goals matter, and which rules apply. Which data may go into AI tools? Who approves a new tool? Settle this first, not after the first incident.
- Inventory. Talk to every team about its work, its tools and where it gets stuck. The output is a ranked list of candidate workflows, with the evidence behind each.
- Baseline. Draw the top workflows as they run today, with frequency, minutes and people per step. Check the numbers against real cases. Date it.
- Plan. Design the target version of each workflow next to the current one. Choose two or three to build first, based on hours, effort and risk.
- Build. Put them live with the people who use them, in the tools they already work in. A person checks what the AI does, and every mistake becomes a rule.
- Evaluate. Measure the same way as the baseline. Let finance confirm the result. Decide what comes next.
This cycle is how we run our own sprints with clients. You can read the details in how we work.
Who owns it
AI that belongs to everyone belongs to no one. A mid-size company needs a small number of clear roles, not a new department.
| Role | Responsibility |
|---|---|
| CEO or COO | Sets the direction and the rules; removes blockers |
| AI owner | Runs the cycle; has time and a target, such as hours freed in onboarding by a date |
| Workflow owner | One person in the team that uses the workflow; approves changes and owns the result |
| Finance | Confirms the numbers before and after |
| IT | Connects the systems and keeps access safe |
The most common mistake is an AI ambassador without a mandate or a target. Enthusiasm is not a plan. Give the role a number to move and a date, and make sure it reports to someone who can change how work is done.
Five traps to avoid
Starting with a platform decision. Choosing a tool before you know your workflows means the tool decides what you change. Decide the work first.
Boiling the ocean. A programme that touches every team at once produces a long list and no proof. Two or three workflows, done well, move a company further than twenty half-finished ones.
Measuring usage instead of work. Licences and logins tell you who opened a tool. They say nothing about whether a workflow got faster. Measure minutes, waiting time and rework.
Promising savings. Freed time is capacity, not cash. It becomes money only when something changes: less overtime, less hiring, more work handled by the same team. Say so from the start, and let finance confirm. We explain the method in how to measure AI ROI.
Forgetting the people. In the Netherlands, a works council will want to know what changes and why. Measure processes, never individuals, and say so clearly. The message that lands is simple: less manual admin, not fewer people.
What you should have after eight weeks
At the end of the first cycle you should be able to show five things: two or three workflows that run on live work, each with an owner; a clear picture of where your team's hours go; before-and-after numbers for every workflow you changed, confirmed by finance; a ranked list of what to change next; and clear rules for how your people use AI.
That is not a transformed company. It is a company that knows how to transform, with proof that the method works on its own work. The next cycle is faster, because the map already exists.
If you are not sure which workflow to pick first, read which processes to automate with AI first. If you have tried pilots before and they stalled, read why AI pilots fail.
Frequently asked questions
How long does an AI transformation take?
The first workflows can run on live work in about eight weeks, including the baseline. The broader change takes longer: you repeat the cycle on the next workflows every few weeks, and each cycle is faster than the last.
We barely use AI today. Is it too early to start?
No. You start from your work, not from the technology, so you do not need to know what AI can do first. Companies that start now can skip the pilots that went nowhere elsewhere.
Do we need developers or a data team?
Not for the first workflows. Most start with the systems you already use, such as Microsoft 365 or Google Workspace and your CRM.
We already pay for Copilot. Isn't that enough?
Licences give people access. They do not change a workflow. We explain the difference in why Copilot licences don't change how you work.
Want to see where your team's hours go? Book a 30-minute call.