In short: start adopting AI from a single process with a measurable effect, not from buying a tool. Pick a task your team repeats every day, measure how long it takes today, introduce AI in that one place, and compare the result. Only once it works should you scale to other areas. The most common mistake is the reverse order - tool first, then hunting for a use.
Why not start with a tool?
Because a tool alone doesn’t change how people work. By the BCG 10-20-70 rule, in an AI transformation about 70% of the value comes from people and processes, 20% from technology and data, and only 10% from the algorithms themselves. In other words: access to the best model is just 10% of success. The rest is understanding where AI genuinely helps and teaching people how to use it.
That doesn’t mean “don’t adopt” - by 2025 already 78% of companies use AI in at least one function. It means “start from the right end”. Companies that buy licences “because everyone has them” usually end up with a tool nobody uses. Those that start from a concrete problem see an effect within weeks.
How to choose the first process?
Look for a task that meets three conditions:
- Repeatable - it happens daily or many times a day, so the savings compound.
- Measurable - you know how much time or money it costs today, so you’ll see the difference.
- Low risk - a mistake won’t cost you a client or a legal problem.
Good starting candidates: handling repetitive enquiries, drafting first versions of offers and documents, tidying up data, triaging incoming tickets. These are places where AI saves time without taking over the decision.
In a manufacturing company, add to that list: describing and searching technical documentation, an initial triage of quality complaints, preparing supplier enquiries, summarising shift reports. The common thread is the same - lots of repetitive work with text and data, where a human still makes the final call.
What not to pick as your first process
Just as important as a good choice is avoiding a bad one. At the start, skip tasks where a mistake is costly or hard to catch: binding financial decisions, legal communication, anything that goes straight to the client without a check. Skip processes that happen rarely too - the savings from them won’t compound. And avoid tasks nobody measures today: without a baseline you can’t prove the effect. The first process should be boring, frequent and countable - not spectacular.
What to expect from the first rollout?
Realistically - tens of percent of time saved on the chosen task, not a revolution across the whole company. The numbers from research are consistent and concrete here:
- An NBER study found that customer-support workers supported by AI resolve 14% more issues per hour, and novices up to 34% more.
- In a controlled experiment, GitHub found developers with an AI assistant completed a task 55.8% faster than the group without one.
- McKinsey, analysing customer support, recorded 14% more issues resolved per hour, 9% shorter handling time and 25% fewer escalations to a manager.
Those are numbers from one process. The sum of such improvements across several places produces an effect you can see in the results. How to convert that effect into money, we show in the post how much adopting AI costs.
How to measure whether it worked?
Before you introduce AI, record the starting point: how many hours a week the chosen task takes and how many people it involves. After the rollout, measure the same thing. The difference - in hours times rate - is your return from that one process.
This simple measurement does two things. First, it gives you the proof you need to decide on scaling to more areas. Second, it protects the budget: if there’s no effect, you find out within weeks and at a small cost, not after a six-month project.
Who should lead the first process?
Not the IT department, but the person who lives that process every day - the shift manager, the head of sales, whoever prepares the offers. They know where the time really leaks and whether the new way of working makes sense. The role of IT or an outside partner is supporting: choose the tool, wire it in safely, take care of the data. The “this works, it stays” decision belongs to the process owner - because they’re the one who keeps the effect alive once the project ends.
What does the first week look like?
In practice the first step is smaller than it seems. Start of the week: pick one task and record the “before” - how many hours, how many people, how many errors. Mid-week: wire the tool into that specific task and show the team how to use it on their own examples. End of the week: first real use, supervised. After two or three weeks you have the “after” number and a decision: scale, adjust, or drop it. That’s enough to turn “it seems to help” into proof.
What if our data isn’t tidy?
That’s the most common excuse and, at the same time, a myth. You pick the first process so that it doesn’t require a big data clean-up - which is why the start is usually work with text (offers, documents, enquiries), not analysis of your whole database. Tidying data comes later, once you already have proof it’s worth it. Waiting for “perfect data” is the most expensive way to never start.
When to scale to more processes?
The signal is simple: when the “after” number is clearly better than the “before”, and the team uses the tool without being reminded. Then you take the next process from the map and repeat the same cycle - measure, deploy, compare. You don’t scale by buying licences “just in case”, but by adding proven improvements. That’s how you grow an effect visible in the whole company’s results, not just one team’s.
The most common mistakes at the start
- Starting with a 12-month strategy. Before you finish it, the tools will have changed. Start with a single working week.
- Rolling out without training the team. A tool without people who can use it is a cost, not an investment.
- Measuring “impressions” instead of time. Without a “before” number you can’t prove the “after”.
- Putting AI where 100% certainty matters - with no human in the loop. AI speeds up work; it doesn’t remove responsibility.
- Spreading across too many processes at once. One well-chosen process, carried through to the end, teaches the company more than five half-finished ones.
Key takeaways
- Start with a process, not a tool - 70% of the value sits in people and processes.
- Pick a task that is repeatable, measurable and low-risk.
- Measure the “before”, deploy in one place, compare the result, then scale.
- Expect tens of percent of savings on a single task, not a revolution across the whole company.
- First results are realistic within weeks if the scope is narrow.