Skip to content
Interface language: English
Audit / Signal

AI use-case register: choose priorities and stop weak ideas

Track AI uses by problem, owner, evidence and constraints, then decide whether to trial, maintain, expand or stop.

Article language: English

A woman reaches for a bowl at a long table while the man beside her holds a cloth.
In this article

An AI tool list says what a company has bought. A use-case register should explain why each use exists and whether it is worth developing. For every task, name the problem, owner, evidence and conditions for the next step. A sound decision to stop a weak idea is a result too. Otherwise the portfolio soon becomes a drawer full of projects labelled "someday" that nobody wants to open.

The register should organise the order of work. It shows what is already operating, what is a trial, what awaits missing information and what has ended. It should not reward the number of ideas submitted by itself.

Distinguish a use case from a product and an account

Describe a use case as an activity leading to a result: preparing a working summary, organising enquiries or searching approved documentation. The product name is one field, not the whole entry. This lets you compare different ways to perform the same task without confusing a process decision with enthusiasm for a tool.

Keep licences, administrators and accounts in an AI tool inventory (in Polish). Link it to the use-case register through a product or organisation identifier. A change of service plan can affect several processes, and ending one task need not mean abandoning the whole product. Make these dependencies visible before a purchase or renewal decision.

Start each idea with a recognisable problem

Ask what makes work difficult today, for whom and how you know. "We want to use AI in sales" is too broad to assess. It could be narrowed to a working summary of meeting findings, with a defined recipient and a way to check it. Only then can you establish a baseline and propose a small trial.

Add the current way of working and the consequence of making no change. If a simple improvement without AI exists, include it among the options. The register aims to improve the process, not require a particular technology. Record observations separately from expected benefits. The proposer's belief may start a hypothesis but does not replace evidence of need. When a problem description rests mostly on declarations, consider openly observing ordinary work before deciding on a change, using the guide to observing an expert at work (in Polish) as a starting point.

Prepare a minimal use-case record

The record should let another person understand the task without a separate presentation. We suggest fields for problem, process, owner, users, input, output, recipient, tool, stage and next decision. Link to sources, data restrictions and acceptance criteria. Keep detailed documentation separately if needed, so the register remains readable. For AI-assisted product descriptions, for example, name the sources of product attributes and how claims will be checked instead of recording only "content creation". The article on product descriptions created with AI (in Polish) examines the sources and checks for those claims.

NIST's AI RMF Map function (in English) emphasises describing a use case's purpose, context, value and limitations. We use it as a voluntary reference. The record layout and prioritisation method below are our operational proposal. They are neither a universal form required by NIST nor proof that a solution complies with law.

Appoint an owner able to accept the result

The owner should know the process, recognise a correct result and be able to decide on further use. Someone interested in the technology can support the trial but may not have authority to change the team's work. Separate the sponsor, configuration implementer and recipient when their roles differ. Clear roles help close a trial with a decision.

Check available time and a substitute too. A use without someone able to check the result is not ready to expand. If specialist knowledge is needed, record it as a trial condition. The guide to team training in verification (in Polish) describes checking acceptance skills on a task. Interest in the tool alone does not demonstrate readiness to check its answers.

Check prerequisites first

Before scoring an idea's appeal, establish whether the material may be used, whether there is a recipient and whether a trial can be carried out safely. Check whether access can be limited and the action stopped. A missing prerequisite should lead to clarification or a change of scope, not disappear inside an average benefit score.

The AI use policy (in Polish), provider assessment (in Polish) and integration permissions review help here. The record need not duplicate them. It should point to the current decision, its scope and the person who resolves it. If a condition is unknown, state it as a question to check and set the next step. The guide to integration permissions (in Polish) sets out what to examine before connecting accounts.

Compare potential with effort and uncertainty

Use a few explicit criteria to decide order: the problem's importance, evidence quality, ability to accept results, effort required and consequences of error. An elaborate scale is unnecessary if the team does not understand its levels. Add a short reason and source to every assessment. A question about the basis is more valuable than a seemingly precise score.

Separate the size of the expected benefit from confidence in achieving it. One idea may promise much but require costly data preparation and lack a tested acceptance method. Another may be smaller but allow a quick, credible trial. The register should show the difference. The full cost of an AI process (in Polish) explores estimating effort and review work.

Aeneia conceptual photography

Limit trials to what the team can accept

Choose only as many parallel uses as the company can properly check and maintain. Starting a product is usually easier than preparing data, criteria and a recipient's time. If every trial depends on the same person, their availability is a shared constraint. Record that dependency in the decision instead of treating each task as a separate project with no coordination cost.

The queue may include an idea waiting for material, a data decision or the result of an earlier trial. For each entry, state the unblocking condition and review date. Do not promise a start date that depends on an unconfirmed resource. A recorded queue also helps explain to proposers why a good idea is not yet ready to execute.

Make stages meaningful for decisions

Use simple states such as submitted, to clarify, ready for trial, in trial, approved within scope, paused and ended. Each state needs an entry condition and a possible next step. "In progress" without a date, owner and expected result is hard to distinguish from an abandoned initiative. A status change should point to evidence or a decision.

Approval within scope means a specific task, users, material and oversight method. Expansion to another team may require new grounds. Keep decision history so a limited pilot is not later read as approval for the entire company. AI pilot criteria (in Polish) help set transition conditions before a trial rather than add them after a favourable demonstration.

Worked exercise: three different ideas

A team proposes meeting summaries, automatic customer replies and search across internal instructions. Each gets a separate record. The summaries have a recipient and permitted material. Sending limits for automatic replies are not established. Search needs document versions to be organised. Do not treat the initiatives as equally ready simply because they use a similar product.

The first trial depends on the company's goal and available evidence. It could begin limited summaries while clarifying communication boundaries and organising knowledge in parallel. The register records these dependencies and conditions for the next decision. This illustrates a working method, not a universal winner. Another organisation may have different priorities, materials and consequences of error, and therefore make a different justified decision.

Record the outcome when ending a trial too

When ending, retain the reason, observations and conditions for a possible return. The product may have failed a quality criterion, the cost may not have been justified or the problem may have had a simpler solution. Do not turn these findings into a general verdict on the whole technology. The result concerns the particular task and conditions tested.

Check access, subscriptions, automated tasks and data copies. Closing a register entry should lead to appropriate technical actions. A plan for retaining and deleting AI data helps organise material cleanup, while an automation audit reveals dependencies that may still operate after a trial is formally closed. The guide to AI data retention (in Polish) and the automation audit guide (in Polish) cover those two lines of enquiry.

Review active uses after changes

Reassess entries when their purpose, users, sources or tool capabilities change. Record a review date and events that trigger an earlier review. An old test result has a defined scope. If a model or integration changes behaviour, use AI response regression tests before extending an earlier approval to new conditions; the regression testing guide (in Polish) explains how to compare results after a change.

Also check whether the problem still exists and the team uses the solution as approved. A lack of use can have several causes: an unnecessary process, an unsuitable result, difficult access or insufficient review time. An activity metric alone does not distinguish them. The register should lead to a discussion of the cause and a concrete maintenance decision.

How do you recognise similar ideas without erasing differences?

During review, compare the problem, input, recipient and effect of the output. Two teams may request similar summaries but use different material and need different oversight. A shared technical component does not mean an identical use decision. Link the records, identifying what can be shared and which conditions remain separate. This reuses earlier work without automatically extending approval.

If records describe the same task, combine them while retaining owner details and decision history. Choose one basis for criteria and a shared reporting method so parallel trials do not proceed without awareness of each other. Record who accepts results for each user group. If assessments differ, return to requirements and observations instead of settling the disagreement by tool popularity or the number of supporters.

Assess tasks with different consequences of error separately, such as using AI to support complaint handling. Trial findings can then feed a retrospective that changes working practices. A status of "ended" alone will not preserve lessons for the next use. The article on AI-assisted complaint handling (in Polish) and the guide to project retrospectives (in Polish) explore the task and the subsequent review.

How do you know the register helps manage work?

Ask someone outside the trial team to find the owner, scope of approval and next decision for a selected entry. Then check whether they can identify evidence of value and the main limitation. If they must reconstruct the history from private messages, complete the record. The register should shorten the path to facts, not replace it with another list.

If there are many ideas, start with a few processes actually in use and one queue for new requests. The Aeneia Audit offering describes an entry point for reviewing uses, costs and responsibilities.

Aeneia / Aeneia editorial

Editorial article. Source links are included in the text.

About Aeneia

Brands we have worked with

  • L’Oréal
  • Garnier
  • Netflix
  • Carrefour
  • Warner Bros. Discovery
  • Biedronka
  • HBO Max
  • Wolt
  • Żabka
  • COSRX
  • Embassy of China in Poland
  • Mixa
Aeneia / Initium

From knowledge to a decision

Tell us what you want to put in order. We will agree on where to start.

Contact Aeneia

We partner with and use solutions from:

  • Google
  • OpenAI
  • Microsoft
  • Stripe
  • Anthropic
  • Linear