First AI use case: where to start, depending on your industry

Start with a task your teams do often, that relies on information already available, and whose result is checked by a person before it is used. That is the profile of cases that go live quickly and win people over.

The challenge is not only technical. According to Insee, 71% of French companies that do not use AI say they see no use for it. The first use case is there precisely to make that use visible, in your company, on work everyone knows.

Three families of tasks almost always meet these conditions:

• Finding information: a procedure, a clause, a reference, a case history.
• Preparing a document: minutes, a first response, a summary, a file to complete.
• Checking: whether a file is complete, whether an invoice matches a purchase order, compliance with a checklist.

When mapping operations, we score each idea on two axes: the expected gain and the difficulty. Here are the questions that help assess them.

A case that ticks every box is not necessarily the most spectacular. That is deliberate: the first project has to succeed, then serve as a model.

Industry and services

The information exists, but it is scattered: manuals, breakdown history, sales references, contracts. An assistant that answers based on your documents, and cites the source of each answer, is a good starting point: the technician finds the procedure, the buyer finds the clause. France’s Directorate General for Enterprise (DGE) has published a guide on this approach, known as RAG. Preparing responses to tenders, drawing on your references, often comes next. See our industry page.

Banking and insurance

Start with document summaries, preparing underwriting or review files, sorting customer requests and preparing responses for the advisor. Don’t start with uses the AI Act classifies as high-risk, such as assessing the creditworthiness of individuals or risk assessment in life and health insurance. Involve compliance from the start: under DORA, using an AI model provider falls within the management of ICT third-party risk. Details on our banking page.

Public sector and local government

Sorting mail and requests from the public, preparing responses reviewed by a staff member, summarizing council decisions or regulatory monitoring. The public servant remains the one who decides, and members of the public who interact directly with an assistant must be told.

Social housing

Tenant relations: sorting requests, checking that files are complete, finding the lease or the procedure, preparing a response or a work order. For housing allocation, AI only checks that the file is complete; the decision belongs to the allocation committee.

Consulting firms and service companies

Meeting minutes, preparing proposals based on past engagements, searching existing deliverables. When the engagement management tool is a spreadsheet, rebuilding it as an AI-ready tool often becomes the second project.

Retail

Logging area managers’ store visits in the CRM, preparing product sheets, cross-referencing sales and inventory to suggest restocking, customer service responses prepared for approval.

Once the case is chosen, the following sequence limits bad surprises:

• Describe the task as it is done today: who does it, with which tools, how long it takes, which errors recur.
• Gather a set of real examples: requests, files, representative documents, including the difficult cases.
• Configure the tool in the environment chosen with the IT department, following the rules of the AI policy.
• Test with the volunteer team on real cases, keeping systematic review.
• Measure on the indicator chosen at the outset, and collect user feedback.
• Decide: extend, adjust or stop. A reasoned stop is also a result.

An example from the customer service team of a distribution company. The task: answering questions about order status. The examples: a sample of real requests, with the responses the team considered good. The configuration: the assistant reads the request, finds the order in the ERP and prepares the response in the company’s tone. The test: volunteer advisors review each draft and note what they had to correct. The measurement: time to first response and the share of drafts sent without major changes. The decision is based on these facts, not on an impression.

The first case also sets habits: review, error reporting, sharing good practices. They will serve for every case that follows.

Some projects are useful, but rarely good first cases:

• High-risk uses within the meaning of the AI Act (Annex III): they carry stricter obligations, such as human oversight, logs and informing the people concerned. Better to tackle them once governance is in place.
• Projects that require data you don’t have: predictive maintenance, for example, assumes sensors and a usable history.
• AI that answers outsiders on its own: a customer or member of the public who interacts directly with an assistant must be told, and an unreviewed error commits the company.
• Cases with no indicator: if no one knows how long the task takes today, no one will be able to say whether AI helped.
• Cases with no volunteers: imposing the tool on a team that sees no point in it dooms the test.

In every case, keep the rule: AI prepares, a person decides. And check with your DPO as soon as personal data is involved.

Choosing the first use case is at the heart of the mapping step in our method: interviews by department, ranking of cases by gain and difficulty, choice of tools with your IT department. Hands-on sessions then let your teams leave with the AI agent that solves their problem, and the AI Factory connects AI to your software if the case requires it.