1. Why a map is needed before projects
Many companies start with AI by trying a tool out of curiosity, and only afterwards look for a problem to solve. The result is often a pilot that nobody uses a few months later. A map by department reverses the order: first you list the processes where lost time is evident, the volume of text or documents is high and the result can be checked by a person. Then you decide which deserve a test. This work does not require deep technical skills, but it does require that whoever knows the process sits at the same table as whoever knows the technology.
The map also helps stop requests that make no sense. A use involving personal data of customers or employees, or producing decisions with effects on people, needs a stricter analysis of its benefit. A use that simply summarises internal notes can start almost immediately. Separating these two cases is the first concrete gain from mapping, even before a project is launched.
2. How to read a use case: data, people, decision
Every possible use of AI can be described with three questions. Which data enters the system, and whose is it? Who reads or uses the result, and what do they do next? Is an error discovered before it has effects, or not? An assistant that proposes a draft reply to an email has a very different profile from one that suggests refusing a request. The same technology, with two different answers to these questions, leads to two different assessments.
This reading should be done together with the process owner, not only with IT. Whoever knows the work knows where the exceptions lie, which documents are incomplete and which errors are costly. If the case is still unclear after this discussion, the right step is to narrow it, not widen it. A narrow case, with a single type of document and a single group of users, is easier to assess and to measure.
3. The starting picture, department by department
The table shows the most frequent case for each department. It is not a catalogue of functions to buy: every row should be verified with someone who really knows the process, because volumes, document quality and internal rules vary greatly from one company to another. A case that is useful in one firm may be worthless in another if the data is scattered or managed with different tools. Use it as a starting point for conversations with the head of each department.
The most useful piece of information is the last column. It indicates where an error hurts and therefore how much supervision is needed. In departments that deal with customers or suppliers, wrong information quickly becomes a commitment or an offer. In internal uses, such as searching procedures, an error usually surfaces before anyone suffers the consequences, and it is easier to correct. Whoever reads the table should ask this question for every row.
| Department | Typical use | Main caution |
|---|---|---|
| Purchasing | Comparing offers and checking clauses in supplier contracts | Errors on financial terms and deadlines |
| Administration and accounting | Reading and classifying invoices and documents | Incorrect entries and tax data |
| Customer service | Draft replies and search in internal documentation | Unverified answers and personal data |
| Human resources | Answers on policies and onboarding support | Decisions about people and candidate data |
| Sales | Standard quotes and summaries of requests | Unauthorised prices and terms |
| IT helpdesk | First-line support and search in procedures | Uncontrolled access and technical actions |
| Legal office | Comparing contract versions and searching for clauses | Opinions produced without professional review |
| Production and maintenance | Searching manuals and fault logs | Incorrect safety instructions |
4. Support functions: purchasing, administration, human resources
In purchasing and administration, the value of AI lies mainly in reading and comparing documents: offers, contracts, invoices, orders. The system can extract data and flag differences, but recording and the final choice remain with a person. The most concrete risk is wrong information passing unnoticed, for example a deadline or an amount misread from a scanned document. For this reason the sample check must be designed before the first error, not after it.
Human resources is a special case. Answers on internal policies and onboarding support are relatively simple uses, provided the system responds only with the company's official documents. When, however, the system assesses candidates or suggests decisions about people, sensitive personal data and stricter rules come into play. Some practices are prohibited by the AI Regulation, and others could fall under high-risk uses: check with an adviser before starting.
5. Customers and sales: customer service, sales, helpdesk
In customer service and sales, AI is more useful as an internal assistant than as a direct interlocutor. An operator receives a draft reply, checks it against up-to-date documents and sends it in their own name. Anyone who answers the customer directly without review takes on a risk that no technology removes: a promise on price or timing that the company cannot keep. The rules on what the system may say, and on what must always go through a person, should be written before launch.
The first-line IT helpdesk is a frequent case because requests are repetitive. Resetting a password, explaining a procedure or finding a manual are suitable tasks. Technical actions such as granting access or changing a configuration should never be carried out on automatic suggestion without human review. The limit must be defined in writing, because that is exactly where an error becomes an incident.
6. Production and maintenance: the most different case
In production, AI can help search for information in manuals, procedures and fault histories, and help sort technicians' reports. The starting data is often scattered: documents in different formats, hand-filled logs, sensors that do not talk to each other. Before discussing models, you need to know whether this data exists, is up to date and has an owner. In many organisations the first useful project is precisely tidying up this material.
The risk is high when the system suggests safety procedures or interventions on machinery. A wrong instruction in that context can cause physical harm. For this reason every suggestion must point back to the original approved document, and no change to a plant should depend solely on generated text. The use makes most sense as support for research, with a technician remaining responsible for the decision.
7. Cross-cutting risks: personal data, controls and accountability
Almost every use case touches personal data, even when it seems to be just text. The GDPR requires knowing which data enters the system, on what legal basis, how long it is kept and who can see it. If the supplier keeps data to improve its own models, this must be clarified in writing. Where data is stored and who can lawfully access it are separate questions, and both should be assessed without taking an answer for granted.
The second risk is accountability. If a result generated by a system is used for a decision, you need to know who takes the decision and who answers for any error. A simple rule is that every use has a named owner, a defined human check and a record of what was done. Without these three elements the case is not ready, whatever the quality of the model. Check with an adviser which obligations arise from the AI Regulation for your case.
8. The series' in-depth guides and where to start
The more specific guides in the series answer the questions that emerge from the map. You do not need to read them all in sequence. Choose the department where lost time is most evident and start there. Each in-depth guide follows the same logic: the process, the data, human control and the first possible test. The questions we cover are these, presented as topics:
- Purchasing: how to compare offers and supplier contracts without missing important clauses
- Administration and accounting: how to check documents before they are recorded
- Customer service: how to build an internal assistant for operators
- Human resources: how to answer questions about policies and onboarding
- Legal offices: what can be delegated and what remains with a professional
- Sales: how to prepare quotes and replies to requests
- IT helpdesk: how to automate first-line support
- Production and maintenance: what is realistic to do with manuals and machine data
- Assessment sheet: how to compare use cases before deciding
9. How to start: one case, measured
The first step this week is to choose a single use case and describe it in one sentence: who uses it, on which document, to achieve which result. If the sentence cannot be written simply, the case is too broad. Narrow it until a manager can describe it in a few minutes, and read the in-depth guide for the relevant department before filling in any sheet.
Then set the criterion for deciding whether it works. It may be the processing time of a document, the number of corrections needed or the rate of errors found by checking. The threshold depends on the process, so it should be defined together with those who carry it out every day. A test without a criterion always ends up saying that everything went well, and in that case you have learned nothing.
Frequently Asked Questions
Which department should we start with?
It is usually best to start in an area where the work is repetitive, the material is already in text form and an error is caught before it affects people. Often these are the internal helpdesk, search across procedures or the reading of administrative documents. What matters is not management's enthusiasm for a technology, but how measurable the problem is and whether a process owner is willing to check the results for a period.
Do we need a separate AI project for each department?
No. Many use cases can share the same tools and the same control rules. It is better to define first a common policy on data, responsibilities and approvals, and then assess each case individually. This prevents each department from buying different tools without knowing what goes into the systems. Where a case requires special rules, such as processing health data or employee data, a separate assessment is needed.
Can AI make decisions about people, for example in selection?
It should not do so without careful assessment. Uses that affect candidates or employees involve personal data and may fall under the high-risk systems set out in the AI Regulation. Some practices are prohibited altogether. The minimum is to keep a documented human decision, explain the role of the tool and check with an employment adviser and your data protection officer before starting.
How much does a use case map cost?
The cost depends on the number of departments involved, the availability of process owners and whether the work is done with internal resources or an external supplier. It makes no sense to ask for a price before deciding the scope. Ask the supplier for an initial limited piece of work covering a few processes, with a written result, and evaluate that result before extending the work.
Does company data end up in suppliers' models?
It depends on the service and the contract. Some suppliers do not use customer data to train their own models; others do if the default setting is not changed. Do not take anything for granted: ask in writing where data is stored, for how long, who can access it and whether it is used for training. Where data is stored and who can lawfully access it are different questions, and both should be assessed.
What is the difference between generative AI and traditional automation?
Traditional automation follows fixed rules: if one thing happens, do another. Generative AI produces text, summaries or classifications from natural language, and so it can make mistakes in less predictable ways. Human control therefore matters more. Many effective projects combine the two: rules for certain steps and AI support where a document needs interpreting. Choose the tool according to the task, not according to fashion.