AI in business: how to choose your first use case
A company's first AI project needs to answer a simple question: which task is worth improving now, for whom and with what evidence? The answer helps determine what to build, who to involve and when to stop an idea that is not working.
It is common to start with the tool. Someone sees a demonstration, spots possibilities and wants to try everything. To turn that curiosity into a useful project, we need to name the problem, understand the routine of those facing it and define what an improvement would mean in practice.
This choice complements the discussion of how AI changes the work of someone founding a startup: building has become more accessible, and deciding what deserves to be built remains central.
Start with a task someone already needs to perform
“Use AI in customer service” is still too broad. “Help the team find the correct guidance for a recurring question” allows a more concrete conversation: where is that guidance, how much time does the search take and how do we know the answer is correct?
The same reasoning applies to preparing meetings, checking documents or organizing requests. Observe a task from beginning to end. Talk to the people performing it and record its frequency, difficulties and rework. Part of the problem may lie in outdated information, unclear responsibilities or too many steps.
When discussing MVPs in a class, I emphasized the importance of validating the pain point and the demand. That question remains relevant inside a company: does someone need this result, and will they incorporate the solution into their routine? AI masterclass, from 38:56.
Compare ideas using the same criteria
Before developing anything, put a few possibilities side by side. Five questions help compare the options using clear criteria:
| Criterion | What we need to know |
|---|---|
| Problem | Who faces the difficulty, how often and with what consequences? |
| Data | Does the necessary information exist, is it current and can the team involved use it? |
| Cost of error | What happens if the answer is wrong, incomplete or late? |
| Owner | Who knows the process, monitors the test and decides whether to continue? |
| Evidence | What observable change would justify further investment? |
An idea can appear valuable while depending on information no one can access. Another may be simple to implement but solve something that happens twice a year. Making these differences explicit avoids choosing solely on the ease of the demonstration.
Microsoft's guidance on prioritizing AI agent use cases considers business impact, technical feasibility and user interest. The useful point is to compare these dimensions before committing the team to a solution. Cloud Adoption Framework.
An example: finding answers in internal documents
Imagine a fictional company whose sales team loses time searching for delivery rules in scattered documents. There are three possible paths: reorganize the documents, create a conventional search or test an AI assistant.
The initial investigation shows that some questions arise from conflicting versions. In this scenario, correcting the sources is a necessary step for any path. An assistant consulting contradictory documents may simply make the confusion happen faster.
After that organization, the team can select recurring questions and record how they are answered today. For an AI test, the initial scope could be to locate the applicable rule and present the source passage for a salesperson to review. Commercial concessions and commitments to customers would remain with the people responsible.
The process owner would check whether the guidance found is correct, whether the source is current and how much work the review requires. If a well-organized page solves the difficulty with less effort, it may be the more appropriate choice. The goal is to improve the task; technology is one possibility for doing so.
Define what the pilot needs to teach
Before the test, record how the task works today. Use situations representative of the routine, including exceptions and questions with no available answer. Agree with the team on an evaluation period, the participants and the criteria for continuing, adjusting or stopping.
Speed is one part of the assessment. It also matters whether the result is correct, whether people can review the answer and whether they return to the solution. Time saved searching can disappear in checking or correcting errors.
Include the effort required to prepare data, keep information current and support users. A pilot that works with a developer beside every person still needs to demonstrate how it will operate day to day.
At the end, try to answer: did the task improve? In which situations? What still depends on intervention? What would it cost to maintain this improvement? These answers are more useful for the next decision than a demonstration selected because it worked.
Building and operating require clear responsibilities
AI has expanded the ability of business professionals to put ideas to work. At the same time, understanding what has been built remains necessary. In that same class, I made this caveat when discussing applications created with AI: a working demonstration does not amount to mastery of the solution. AI masterclass, from 46:18.
That is why you should involve those who understand the process and those who will be responsible for technical operations. Moving from a test to recurring use requires decisions about access, maintenance, failures and support. These responsibilities need people and available time.
A well-chosen first use case leaves both a deliverable and a lesson: the company better understands the problem, the conditions under which AI helps and what needs to change to move forward.
If your company is organizing challenges to connect business, technology and new talent, explore NewHack's corporate innovation work.