Finding AI Use Cases: Methods That Work in Mid-Sized Companies
Plenty of companies have decided to use AI without knowing where it should go first. In a September 2026 survey by the German digital association Bitkom, 30% of companies not yet using AI said they had not identified any use cases. This guide compares the methods for finding candidates, shows the blind spot of each, and explains how to turn a raw idea into something you can evaluate.
What does finding an AI use case actually mean?
Finding an AI use case means naming a specific task that costs time, money or quality today, and where AI can contribute something conventional software cannot. The output is not a sticky note. It is a describable task with a trigger, input data, a desired result and a named owner.
Treat the search as its own phase, separate from evaluation. When teams judge ideas while still collecting them, promising candidates get killed in the first meeting and pet projects survive without scrutiny. The sequence that works is simple: search broadly, describe consistently, then evaluate and prioritise.
Good first finds in mid-sized firms are rarely glamorous. Sorting a shared inbox by product line. Drafting answers to technical questions from manuals and past tickets. Preparing quote drafts from bills of materials and price lists. Checking inspection reports for missing fields. These work well precisely because the trigger and the expected output are clear.
It also helps to separate a use case from a tool. "We are introducing a chat assistant" is a tool choice. "Every technical enquiry to the inside sales team arrives with a draft reply that cites the product documentation" is a use case. Only the second one can be measured later.
How does the search work? Four building blocks
Every method combines four building blocks: a source of problems, a check against what AI is good at, a quick look at the data, and a standard way of writing it down. Methods differ mainly in where the problems come from: observation, conversation, system data or outside inspiration.
1. A source of problems. This produces the raw list: waiting times, retyping between systems, questions answered every week, errors caught too late. The closer the source sits to daily operations, the more reliable the finds.
2. A capability check. Not every problem is an AI problem. Google's People + AI Guidebook advises looking for the overlap between real user needs and AI strengths. Among the strengths it lists natural language understanding, prediction of future events, recognising an entire class of entities and detecting rare events that change over time. Among the situations where AI is probably not better, it lists cases where predictability matters, where only static or limited information is provided, where costly errors must be minimised and where complete transparency is required.
3. A data check. At this stage a rough answer is enough. Does the input exist in digital form? Is it in the ERP system, a shared mailbox, a network drive, or only in someone's head? Who is allowed to use it?
4. A brief. Every find is written up in the same format. That is what makes ideas from different departments comparable. The fields are listed in the step-by-step section below.
Which methods exist, and when does each one fit?
Six methods cover the ground: process walks, pain-point interviews, data trail analysis, an outside-in capability catalogue, an open idea channel and a facilitated use case workshop. None is sufficient on its own. The table shows what each method is good at and where it systematically misses things.
| Method | Good at finding | Typical blind spot |
|---|---|---|
| Process walk | Retyping, waiting, duplicated work | Only sees what happens that day |
| Pain-point interviews | Frustration, error costs, tacit knowledge | Loud problems crowd out quiet, costly ones |
| Data trails | Volumes, frequencies, real lead times | Only what systems record |
| Capability catalogue | Ideas outside current habits | A solution looking for a problem |
| Open idea channel | Breadth, participation, buy-in | Many half-formed entries |
| Use case workshop | Alignment, shared language, a fast list | Depends on preparation and who is in the room |
Process walk
You follow one case from arrival to completion, for instance a customer enquiry until the quote goes out, and note every step, system and hand-over. AI candidates show up wherever people copy information out of unstructured sources, hunt for it across systems or rewrite it for someone else.
The blind spot: a walk shows the typical case. Seasonal peaks, rare exceptions and problems tied to particular customers stay hidden, so pair it with interviews or data.
Pain-point interviews
Short, structured conversations that ask for specific episodes rather than opinions. When did you last spend ages looking for something? Which question do you answer again every week? Which mistake did a customer spot before you did? Concrete stories produce better finds than asking where AI might help.
The blind spot: interviews amplify what annoys people, not necessarily what costs money. A rare but irritating task gets mentioned more often than an unremarkable one that eats an hour every day. Cross-check with frequencies.
Data trail analysis
Systems record what actually happens: ticketing tools, shared mailboxes, ERP documents, change histories. A simple count of which ticket categories pile up, which requests stay open longest and which documents get corrected most often replaces impressions with numbers. For deeper analysis, process mining reconstructs workflows from event logs; the IEEE Task Force on Process Mining published a manifesto setting out its principles in 2012.
The blind spot: data only shows what is captured. Phone calls, desk-side agreements and knowledge kept in spreadsheets are missing, and data rarely tells you why something is slow.
Outside-in capability catalogue
This reverses the direction. You start from known AI capabilities, such as classifying text, extracting data from documents, answering from a knowledge base or inspecting images, and ask each department where that capability would change a task. It is useful when a team has little sense of what current AI can do. Ready-made example lists are in the overviews of AI use cases by department and AI use cases by industry; they show what others typically tackle, whereas this guide is about uncovering the right cases inside your own operation.
The blind spot is the obvious one: a solution looking for a problem. Every catalogue find must be tied to a real task with a cost attached, or it is dropped.
Open idea channel
A form, an email address or an intranet channel where anyone can submit suggestions. It builds participation and surfaces ideas from corners a central team would miss. It only works if every submission gets a clear answer about what happens next.
The blind spot: most entries are one-liners without data or frequency. Budget time to follow up on the promising ones and turn them into briefs.
Facilitated use case workshop
A workshop compresses several methods into a few hours: prepared process maps and data extracts are discussed with the people who do the work, and finds are captured as briefs on the spot. Formats, agendas and outputs of such workshops are covered in a separate guide on workshop formats; this overview only places the workshop among the other methods.
The blind spot: without preparation, a workshop becomes a brainstorm that mostly records the ideas of the most senior person present. A process walk, interviews or a data extract beforehand make it far more productive.
Which combination makes sense?
A practical mix is one observational method (process walk or data trails), one conversational method (interviews or a workshop) and a pass against the capability catalogue. The blind spots largely cancel out. A find that appears independently in two methods deserves extra attention.
What do you need in place before you start?
Less technology than most people expect. You need access to the people who run the process, a rough map of systems and data sources, a goal to measure finds against, and someone who carries the results forward. A finished data strategy is not a prerequisite.
People and mandate. Someone has to own the search, have time for it, and have backing from senior management. Departments only answer honestly when it is clear the aim is to take load off them, not to cut jobs.
A goal. A rough goal filters the finds. Is it faster quotes, fewer repeat questions in service, or keeping knowledge that leaves with retiring experts? The guide on AI strategy for mid-sized companies shows how to derive such goals from business priorities.
A systems map. A list of the main systems with a contact person is enough: ERP, CRM, document storage, ticketing, shared mailboxes. You need it to judge the data situation for each find.
Employee representation. Where employees are represented by a works council or similar body, involve them early. In Germany, for example, works councils have a co-determination right over technical systems designed to monitor employees' behaviour or performance under Section 87(1) no. 6 of the Works Constitution Act. Treat this as orientation, not legal advice.
From find to evaluation: the steps
The process runs in a fixed order: set the frame, search with at least two methods, write a brief for every find, set aside the obvious misfits, and hand the remaining list to evaluation. The brief is the step that matters most, because it decides whether later prioritisation is fair.
- Set the frame. Agree the goal, the departments involved and the timeline. Decide who receives the results and what happens next.
- Search. Use at least one observational and one conversational method from the table above. Collect finds without filtering.
- Write briefs. Describe every find in the same format (fields below).
- Set aside misfits. Mark finds with no visible benefit, no digital data or an obvious legal conflict. Do not delete them; circumstances change.
- Hand over. The cleaned list goes into evaluation and prioritisation, where benefit, feasibility and risk are weighted consistently. That is a separate phase with its own methods. Checking a candidate's data, systems and error tolerance is covered in the guide to the AI feasibility assessment; the commercial side is covered in how to calculate AI ROI.
A use case brief should contain at least:
- Task and trigger: What starts the task? An email, an order, a sensor reading?
- Current process: Who does what, in which systems, how often?
- Input data: What data does the task need, where does it live, in what format?
- Desired output: A draft, a classification, a check result, an answer?
- Automate or assist: Should AI take over the task or support a person?
- Cost of errors: What happens if the output is wrong, and who notices?
- Owner: Who in the business owns the process?
For the automate-or-assist field, the People + AI Guidebook offers a useful rule of thumb: automate tasks that are difficult, unpleasant, boring or repetitive; augment tasks people enjoy, where personal responsibility for the outcome matters or where the stakes are high. For a first deployment, assisting is often the more cautious route: a draft that a person reviews and approves.
The cost-of-errors field is routinely underrated. A misrouted ticket costs one forward. A wrong price in a quote costs margin or the order. Both can be good use cases, but they need very different controls.
Limits and risks of the search
The most common failures are starting from the technology instead of the problem, lists that never had a data check, and finds that touch legally sensitive areas without anyone noticing. The search itself can also raise expectations the company later fails to meet.
Technology first. Start with a tool and look for uses, and you will mostly find demos. Test every find by asking whether it would still be a problem worth solving without AI.
Missing data. Many finds fail later because the data exists only on paper, in personal inboxes or not at all. In the Bitkom survey, 40% of companies without AI cited missing data as a barrier. The rough data check in the brief exists to surface this early.
Legal limits. Some application areas count as high-risk under the EU AI Act. Annex III includes "Employment, workers' management and access to self-employment", covering for example systems that filter job applications or monitor performance and behaviour. Such finds are not banned, but they carry substantial obligations. Flag them in the brief and get legal advice; this note is not legal advice.
Broken promises. Ask staff for ideas and never report back, and you will get fewer ideas next time. Communicate which finds move forward and why others are parked.
Too many finds. A long list feels productive but simply moves the work into evaluation. Fewer finds with complete briefs beat dozens of one-liners.
Acceptance. Acceptance stays an issue after launch: among companies already using AI, 42% named lack of employee acceptance as a barrier, according to Bitkom. Involving the people who do the work during the search gives you a better starting point.
How do you get started?
Start small and specific. Pick one area with a lot of recurring text and document work, such as inside sales or technical service, and combine a process walk with a handful of short interviews there. Capture every find as a brief and pass the list to a structured evaluation.
A pragmatic four-week start could look like this:
- Week 1: Agree goal and scope, build the systems map, inform employee representatives, choose interviewees.
- Week 2: Walk one typical case end to end, interview the people who handle it, run a simple count on the ticketing tool or shared mailbox.
- Week 3: Check finds against the capability catalogue, write briefs, mark obvious misfits.
- Week 4: Present results, hand the list to evaluation, report back to everyone involved.
If you would rather run the search as one facilitated session, the page on the AI workshop describes such a format. To place the finds in a wider target picture, see the guide on AI strategy for mid-sized companies. Typical target areas for first finds are outlined under AI agents for business functions, for example quotation and technical service.
When does this approach not fit? If you already have a clearly scoped use case with a budget and an owner, a broad search is a detour. Spend the time on a feasibility check of that one case instead.
Frequently asked questions
How many use cases should a first search produce?
There is no magic number. Aim for a list long enough to give you a real choice but short enough that every entry gets a complete brief, because half-described ideas cannot be compared fairly.
Do we need an AI strategy before we start looking?
No, but a rough goal helps you filter. Without one, the list fills up with ideas that are technically interesting and commercially irrelevant.
Is rolling out a company-wide chat assistant a use case?
It is a tool decision rather than a use case. A use case names a specific task with a trigger, an output and a benefit, such as routing incoming enquiries by product line.
Who should take part in the search?
People who run the process every day, someone who knows the systems and data, and a person who decides on budget and priorities. Without the operational view, the finds stay superficial.
Sources
- Google PAIR: People + AI Guidebook – User Needs + Defining Success pair.withgoogle.com
- Bitkom: Majority of German companies use AI for the first time (14 Sept 2026, German) bitkom.org
- EU AI Act, Annex III: High-risk AI systems referred to in Article 6(2) artificialintelligenceact.eu
- German Works Constitution Act, Section 87 (German) gesetze-im-internet.de
- IEEE Task Force on Process Mining: Process Mining Manifesto tf-pm.org