Start from the problem, not the product
The most common strategic mistake with AI is also the most natural one. A leader sees a capable tool, asks where it could be used, and then goes looking for places to put it. That produces activity. It rarely produces value, because the organisation is now solving the question "how do we use this?" rather than "what is costing us most?"
Turn it round. Start from the processes that matter to your results, find the ones that are slow, error-prone, expensive or frustrating, and only then ask whether AI is a sensible part of the fix. Sometimes it will be. Often the honest answer is a simpler change: removing a step, fixing a data source, clarifying who decides.
Your managers can run a task-level audit of their own teams. The AI for Managers course covers that in detail. Your job is one level up: choosing which processes deserve investment and attention at all.
A simple way to sort the candidates
Look at each candidate process through four lenses. You are not scoring to three decimal places. You are sorting into "worth a pilot", "not yet" and "no".
| Question | Good sign | Warning sign |
|---|---|---|
| Does the process matter to results? | Affects customers, revenue, cost or risk directly | Nice to have, or nobody can say who relies on it |
| Is the work mostly reshaping information? | Drafting, summarising, sorting, first responses | Deep judgement, negotiation, sensitive people decisions |
| Can errors be caught before they cause harm? | A person reviews before anything leaves | Output goes straight to customers or into records |
| Can you measure the current way? | Clear time, cost, error or backlog figures | Nobody knows how it performs today |
A process with good signs in all four rows is a strong pilot candidate. Two or more warning signs usually means "not yet": fix the process or the measurement first. A warning sign in the third row on its own, where errors reach people unchecked, is a reason to redesign before anything else.
Checkpoint
Start from the processes that matter, not from the tool. A good candidate matters to results, is mostly reshaping information, lets errors be caught before harm, and can be measured today.
Pilot small, on purpose
A pilot is not a soft launch. It is an experiment with a question, a comparison and an end date. Keep it deliberately narrow: one process, one team, a few weeks, and a clear measure agreed before anyone starts.
Small pilots are cheap to stop, which is exactly why they produce honest answers. A large programme with a launch event and a senior sponsor develops its own momentum. People stop reporting bad news because stopping would be embarrassing, and a failure that should have cost a few weeks costs a year.
Measure against the current way
The only fair test is a comparison with how the work is done today, including the time people spend checking and correcting AI output. A pilot that reports only how fast the draft appears is measuring the wrong thing.
Ask for three numbers at the end: how the process performed before, how it performed during the pilot, and how many errors were found in review. Then ask the team what surprised them. The surprises are often worth more than the figures.
Be careful with pilots measured by how many people used the tool. Usage tells you people logged in. It says nothing about whether the work got faster, cheaper or better.
Kill what does not work
Every pilot should end in one of three decisions: scale it, change it, or stop it. Say at the start that stopping is a legitimate outcome, and mean it. The organisations that get real value from AI are usually the ones that stop weak ideas quickly and put the freed attention into the ones that worked.
Stopping is not failure. A pilot that shows a process is not suitable has saved you the cost of rolling it out everywhere. Write down what you learned, share it, and move to the next candidate.
โ Weak prompt
Prompt
Where should my company use AI?
Output
Consider customer service chatbots, predictive analytics, marketing content generation, process automation and AI-powered decision support.
A list of categories drawn from other organisations. None of it knows your processes, your costs or your risks.
โ Good prompt
Prompt
Here are six processes in my organisation, with who does them, how often, and the main problem with each. Sort them into worth a pilot, not yet, or no, using four questions: does it matter to results, is the work mostly reshaping information, can errors be caught before harm, and can we measure the current way. Explain each verdict in one sentence. Where the real problem looks like a broken process rather than slow work, say so.
Output
Supplier query triage: worth a pilot. High volume, reviewed before reply, current backlog is measured. Month-end reconciliation: no. The problem is three systems that disagree, which AI will not fix.
Your processes, your tests, a verdict per item, and an honest flag where the answer is not AI at all.
I lead [type of organisation]. Here are the processes I am considering for AI:
- [process, who does it, how often, main problem]
- [process, who does it, how often, main problem]
- [process, who does it, how often, main problem]
For each, answer four questions:
- Does it matter directly to customers, revenue, cost or risk?
- Is the work mostly reshaping information, or deep judgement?
- Can errors be caught before they cause harm?
- Can we measure how it performs today?
Give a verdict: worth a pilot, not yet, or no, with one sentence of reasoning.
Flag any case where the real problem is a broken process, not slow work.
Use only what I have told you. Do not invent figures.
๐ Quiz
Question 1 of 4What is the most common strategic mistake when choosing where to use AI?