1. Why choosing the first case matters more than it seems
The first AI project is not just about producing a result. It teaches the organisation how to assess, govern and measure this technology. A first case that is too ambitious, with scattered data and several departments involved, tends to fail for non-technical reasons, and the failure weighs on staff confidence for a long time. A small, useful and measurable first case builds competence and makes later choices simpler.
That is why the selection should be made with a method, not out of enthusiasm or because of the latest news. The method in this guide compares two dimensions: how much value the case can bring to the company, and how realistic it is to deliver it with the resources, data and rules you have today. The best cases sit in the area where the two dimensions meet.
2. What you need before you start
Before assessing the cases you need three things. The first is a list of repetitive or text-based activities that take up staff time, gathered by talking to the people who carry them out. The second is a process owner: a person who decides how the process should work and agrees to measure the change. The third is an internal rule on which data may be used and in which tools.
Without an owner the project is orphaned: nobody can say whether the output is good, and nobody decides when to stop. Without a data rule, every trial risks exposing information that should not leave the company. If these elements are missing, it is more useful to spend a week building them than to start a technology test. The time the owner devotes to the project must also be planned from the outset, otherwise the project proceeds in fits and starts.
3. The four-step method
The method is applied with a simple worksheet and a two-hour meeting with the heads of the departments involved. No mathematical model is needed. What is needed is a structured discussion in which every claim is linked to evidence: a figure, a log, a real example of work carried out. Two competent people may be enough, provided they know the process and the data that feeds it.
Pay attention to the last step. A case chosen without a definition of success ends up being judged on the impression of whoever talks about it most convincingly. Writing down the success conditions and the stop conditions before starting avoids endless debate and allows you to say honestly whether the trial worked. The person leading the meeting should record the agreed definitions in writing and have those present approve them.
- Gather the candidates: list the activities where text or information is central, such as responding to recurring requests, preparing drafts, summarising documents or classifying requests.
- Assess the value: for each candidate ask how much time it takes, how much error or delay it causes, what effect it has on customers or revenue, and who benefits.
- Assess the feasibility: ask whether the necessary data exists and is accessible, whether the process is stable, whether someone can check the result, and whether the data risks are manageable.
- Choose a case and define success: select a candidate with high value and high feasibility, and write down what it means for the trial to have worked and what would stop it.
4. How to assess value and feasibility
The table sets out the most useful criteria. For each one there is a question to ask and a signal that indicates a good choice. Use it as a guide for discussion, not as an automatic score: the weight of each criterion depends on the company and the sector it operates in. Use qualitative judgements, such as low, medium and high, and write next to each judgement the reason that supports it.
The most underestimated criterion is human oversight. A case in which nobody can say whether the answer is right is not a good first step, because it does not produce learning. Prefer activities in which the output is easy to verify, such as a draft that a person reviews before sending, over tasks where an error goes unnoticed. An activity with simple human oversight may have a lower value, but it is often the safest choice to begin with.
| Criterion | Question to ask | Sign of a good choice |
|---|---|---|
| Frequency | How many times a month does the task recur? | The task is recurring, not occasional |
| Impact | What effect does the task have on customers, timing or quality? | The benefit is visible to those who use the process |
| Data quality | Does the necessary data exist in readable form? | The information sits in a few well-identified places |
| Data risk | Does the case involve personal or confidential data? | The risk is low or manageable with clear rules |
| Human oversight | Can someone verify the output within a reasonable time? | A competent person is available to review it |
5. Common mistakes in the choice
The first mistake is starting with the most spectacular case, such as an assistant that answers everything. A project like that involves a lot of data and many exceptions. The second mistake is choosing the case based on the technology available rather than the problem. The third is excluding less visible processes, such as administration or the handling of internal requests, which often offer the best balance between value and feasibility.
The fourth mistake is measuring only the time saved. Time must be read together with the quality of the result. A process that saves a few minutes but generates errors that are hard to detect is not a gain. A first case that saves little but works well is worth more than a large project that never closes, and it builds the trust needed for the next step.
6. How to check the choice
Once you have chosen the case, test it with three questions. The first: is the problem real, and do people experience it every week, or is it someone's idea? The second: if the trial worked perfectly, would it change anything concrete in the process? The third: if the trial failed, what would we lose and how easy would it be to stop? Each requires a clear answer, not a general impression.
A good choice passes these questions with simple answers. If the answers need lengthy explanations, the case is probably too large or poorly defined. Narrow the scope until you have a case that can be described in two sentences and measured within a few weeks. If the case does not pass this test, it is not yet ready to start, and it is worth returning to the list of candidates.
7. How to get started
This week, gather the list of candidate activities with the heads of at least two departments. Ask each one which task costs them the most time or generates the most corrections. Do not discuss technology yet: the aim is to build a list of real problems with the people who live with them every day. Also note any solutions already tried in the past and why they did not work, because they often contain valuable information.
Then apply the four steps from section three to no more than five candidates, choose a single activity and write the definition of success and of stop. Only after this choice does it make sense to compare suppliers or tools, bringing a precise question rather than a generic request. The preparatory work saves time in later phases, because it avoids evaluating solutions that do not address the real problem.
Frequently Asked Questions
How many use cases should I start together?
Usually just one, or at most two in different areas, so that attention and responsibility are not spread thin. Several cases in parallel make it hard to tell which one produced a result, and staff end up overwhelmed by change. Once the first case has shown that the method works, you can widen the scope using a selection based on the same criteria.
Must the first case save money?
Not necessarily. A first case may aim to reduce a delay, improve the consistency of responses or free up skilled people for higher-value work. The benefit should be declared in advance, so that the result can be judged at the end without interpretation.
Do I need an external supplier for the first project?
It depends on your internal skills and the type of data involved. A supplier can help assess feasibility and risks, but the process owner must remain within the company. Before entrusting the project, clarify who defines success, who checks the outputs and where the data is held during the trial.
How do I tell a feasible case from one that is not?
A feasible case has accessible data, a process that repeats under reasonably stable rules, and possible human oversight. A case that is hard to deliver depends on information scattered across too many systems, on continuous exceptions, or on decisions that nobody can justify.
What happens if the first trial does not work?
A trial that does not produce the expected result is useful if it clarifies why. It may turn out that the data is unsuitable, that the process is not stable, or that the value is lower than expected. Document what you learned, decide whether to stop or correct the scope, and choose the next case with the same discipline.