You can find the AI use cases worth pursuing in a single session of half a day with the people who do the work. Walk through five to eight processes, look for heavy repetition, lots of text or documents, lots of retyping, and a clear outcome, then score each candidate on value, feasibility, and available data. You leave with a shortlist of two or three.
Most organizations are still at the starting line. According to Eurostat, 20% of EU enterprises with 10 or more employees used AI in 2025, rising to 30.4% of midsize firms and 55% of large ones. Among companies that considered AI and decided against it, 70.9% named a lack of relevant expertise as the reason.
Expertise starts with knowing where AI pays off in your own work, and you don't need a consulting engagement that runs for months to get there. Below is a format you can run yourself: who should be in the room, which processes to cover, what to look for, and what you walk away with.
Why half a day is enough to find AI use cases
An inventory of AI use cases is about the work itself. That knowledge sits with the people who enter orders, answer emails, and compile reports every day. They know which step eats time, where errors creep in, and which exceptions keep coming back. You can collect that knowledge in four hours, as long as the right people are at the table and you follow a fixed structure.
A short, focused session also protects you from a well known trap. A study by Harvard Business School and Boston Consulting Group with 758 consultants found that on tasks AI handles well, consultants using it completed 12.2% more tasks and finished 25.1% faster. On a task outside that range, consultants with AI were 19 percentage points less likely to reach a correct solution. The researchers call this boundary a jagged frontier, because tasks that look equally hard can fall on different sides of it. That's why you start from specific process steps. A broad question like "what can AI do for us" won't give you that precision.
Who to invite and what to prepare
Keep the group small. Six to eight people is enough:
- Two or three people who do the work. They know the actual steps, the exceptions, and the workarounds.
- A process owner or operations manager. They know what a step costs the business and what an improvement is worth.
- Someone from IT or application management. They know where the data lives and which systems are involved.
- A decision maker. They can say at the end which candidate moves forward.
Ahead of the session, ask for three things per process: a rough weekly volume, two or three real examples (an email, a form, a report), and the systems involved. Real examples on the table keep the conversation concrete.
| Time | Part |
|---|---|
| 30 minutes | Goal, ground rules, and the four signals |
| 90 minutes | Walk through processes, steps, and examples |
| 60 minutes | Score candidates on value, feasibility, and data |
| 30 minutes | Pick the top three and agree on the first step |
| 30 minutes | Buffer for overruns and open questions |
Which processes to walk through
Pick processes with volume, because that's where the payback is. Good places to start:
- Sorting and routing incoming emails and messages
- Entering orders, quote requests, or invoices from PDFs
- Answering customer questions using manuals or case files
- Compiling reports from several sources
- Checking documents for completeness before they move on
For each process, ask who does it, how often, how long one instance takes, and what happens when it goes wrong. Have the people who do the work talk through the steps with a real example in hand. Note every step where someone reads, retypes, summarizes, or picks from a fixed set of options. Those are your candidates.
The pattern holds across Europe. Eurostat reports that 31% of enterprises that use AI apply it to business administration, and in the Netherlands the national statistics office CBS puts that figure at 32%. These are the processes nearly every organization has.
Signals that point to a good AI use case
Four signals keep showing up in candidates that deliver:
- Heavy repetition. The same action dozens or hundreds of times a week. A small time saving per instance adds up quickly.
- Lots of text or documents. Emails, contracts, work orders, complaints, reports. Language models are good at reading, summarizing, and organizing text.
- Lots of retyping. Data that someone copies from one document or system into another.
- A clear outcome. You can check afterwards whether the result is right: the correct field, the correct category, a complete answer.
Research on work with lots of text backs this up. In an experiment published in Science, professionals using ChatGPT finished writing tasks in 40% less time while quality rose by 18%. At a customer support operation with 5,179 agents, an AI assistant helped them resolve 14% more issues per hour, and 34% more for new and less experienced agents. In both cases a person still did the work and judged the result. AI works best when people stay close to it.
Signals that AI won't add much
An inventory only becomes useful when you're willing to cut. Set a candidate aside when you see these signals:
- It rarely happens. A task that comes up four times a month doesn't justify building and maintaining a solution.
- The rules are fixed. If every step can be written as a fixed rule, conventional business process automation is faster, easier to maintain, and more predictable.
- Nobody can check the outcome. An error that only surfaces months later is a poor starting point.
- The data doesn't exist or can't be reached. Information that lives only in people's heads or on paper has to be captured first.
- It's mostly judgment and negotiation. Tasks where experienced people disagree on the right answer often sit outside what AI does reliably.
Scoring candidates on value, feasibility, and data
Give each candidate a score from 1 to 3 on three criteria. Do it together and keep it quick. Arguing over half a point gets you nowhere.
| Criterion | 1 | 2 | 3 |
|---|---|---|---|
| Value | Few hours saved, little effect on errors or lead time | Noticeable gain for one team | Large gain in hours, errors, or lead time |
| Feasibility | Many systems, many exceptions, outcome hard to check | Some exceptions, outcome partly checkable | Clearly bounded step, outcome easy to check |
| Data | Hardly any examples, or paper only | Examples exist, spread across systems | Hundreds of digital examples in one place |
Add up the scores. A candidate with 8 or 9 points is a serious first step. A 1 on data usually means something else has to happen first, such as capturing information digitally or opening up a system.
For every candidate, also write down where the person stays in the loop. Who checks the output, who decides when it's unclear, and what happens to exceptions. That agreement is part of the use case itself, and it shapes how much the team will trust the solution later.
What you walk away with
After four hours you have five concrete outputs:
- A list of every process step discussed, with its volume
- A scored list of candidates on value, feasibility, and data
- A top three, each with the role of the person in the process
- A list of rejected ideas and the reason for each
- An agreed first step with an owner and a date
That first step is almost always a small trial on real data. Take a hundred real examples, have a model perform the task, and have the people who do the work review the results. Within a few weeks you'll know whether the gain holds up. Measure beforehand how long the step takes now and how often it goes wrong. Without that baseline you can't show what the trial delivered, and the decision to continue comes down to gut feeling.
If the candidate holds up, the build work begins: connecting the AI step to the systems the data comes from and the systems the result needs to reach. In practice, that's where most of the effort goes. Isatis has spent more than 30 years building software that carries this kind of integration, from aviation maintenance (MRO) software to a subsidy administration platform and RFID across the supply chain. With 30+ engineers in Nijmegen and Sarajevo and more than 100 projects delivered, we help with practical data and AI solutions, connecting AI to your existing systems, and independent architecture advice. Our focus is business value, and we leave experiments for their own sake to others. Get in touch if you'd like a second opinion on your top three.
Frequently asked questions
What is an AI use case?
An AI use case is a specific step in a business process where AI performs or supports a task, with a measurable outcome. Examples include sorting incoming emails by topic, extracting data from PDFs, or drafting a first reply. A good use case also states who checks the result.
How long does an AI use case inventory take?
With good preparation, one session of half a day, about four hours, is enough. Beforehand, collect a volume estimate, a few real examples, and the systems involved for each process. Testing the best candidate in a small trial comes afterwards.
Who should take part in the inventory?
Six to eight people: two or three who do the work, a process owner, someone from IT, and a decision maker. The people who do the work matter most, because they know the steps, the exceptions, and the workarounds.
When is conventional automation a better fit than AI?
When every step can be described with fixed rules and the input always has the same format, automation based on fixed rules is usually the better choice. It's predictable and easy to test. AI fits better with tasks involving free text, varied documents, and an outcome a person can check quickly.
