Where do you start with AI in your organisation? Five steps

Why an AI strategy often delivers nothing

Many organisations start with a strategy, a workshop or an exploration. Out comes a list of opportunities, and then little happens. That is rarely down to the people. It comes down to three things:

• Themes instead of problems. “Improve customer contact with AI” is a theme. “Our customer service gets 4,000 calls a month and we cannot keep up” is a problem.
• No owner. If nobody owns the problem and can decide, it stays an idea.
• Nothing to measure. Without knowing in advance what has to improve, you cannot tell afterwards whether it worked.

That is why this guide starts with the problem.

Step 1: find the problem

Look for where time or money leaks away. Four places where you almost always find something:

• Reading and writing work: quotes, invoices, applications, reports, emails that all look much the same.
• Customer questions: the same questions, by phone or email, including outside office hours.
• Retyping: data that people copy from one system into another.
• A package that no longer fits: software that is too expensive, too old or dictates how you work.

A practical tip: have a few people keep track for a week of which tasks they repeat and how much time that takes. The list you end up with is more concrete than any workshop.

Every problem fits one of three routes: a new AI tool, replacing old software or self-hosted AI when the location of the data is a requirement.

How you know you are done: you can describe the problem in one sentence, with a number in it.

Step 2: who is the owner?

Do not start without someone who owns the problem and can make decisions about it. The owner is usually the manager of the process: head of customer service, head of administration, operations director. They know what is good enough, can free up people and decide whether something gets built.

How you know you are done: there is a name next to the problem, and that person wants it solved.

Step 3: run the numbers

Put on one page what the problem costs now, what it returns if it improves and what the solution costs. An example with round numbers:

	Per year
Manual work: 10 hours a week, at €50 an hour, 48 weeks	€24,000
If AI takes over 70%, with a person involved when in doubt	€16,800 saved
Building the application, one-off (example)	€30,000
Payback period	About 1.8 years

This is an example. In your situation, quality, speed and availability count too, as well as the cost of maintenance and hosting. Read more in AI implementation costs and cost and return up front.

How you know you are done: the numbers fit on one page and the owner recognises them.

Step 4: test small and fast

A working prototype on your own data says more than a plan. Build a first version in a short time, let the people who do the work use it and see whether the numbers hold.

That is how our start sprint works: six days, from €8,000, able to start within two weeks. At Kwakkel BV, after six days there was a prototype processing administrative steps on their own records, with the business case on one page. With that, Kwakkel can decide for itself whether and when the build starts.

How you know you are done: you have seen it work on real data, and you know what the full version costs.

Step 5: build, roll out, train

If the numbers hold, you keep building in short sprints, with a demo after every sprint. Three things make the difference between a system that runs and a system that gets used:

• The team that works with it gets trained on it, on the job and on the system itself. See AI training for your team.
• Compliance is part of the design: risk class, logging and oversight. See compliance and security.
• A person stays involved where it matters. A good example is the AI voice we built for Medux: 4,000 calls a month, 80% handled by AI, and difficult or urgent questions passed straight to a member of staff.

How you know you are done: the team uses it every day, and you measure whether the problem from step 1 is getting smaller.

When AI is not the answer

Sometimes the answer is a better way of working, a standard package or simple automation without AI. That is fine. If AI is not the solution, we say so. And if the business case does not add up, we do not build.

What AI Heroes does here