ELLE01EFFE02Services03Events04Blog05Company06Careers07
← All articles
AI & processesSeptember 29, 20264 min read

The first AI use case that pays back

Most AI pilots stall because the first process was the wrong one. Five criteria, a simple score and a path from pilot to production in weeks.

Written with AI and reviewed by the Logol team.

Start with the process, not the model

Many companies begin their work on AI with a tool. A licence is bought, a chatbot is opened, a few people experiment. Six months later nobody can say what changed. The problem is rarely the technology. It is the choice of the first use case.

The first process you automate with AI does two jobs. It has to deliver a result you can measure. And it has to teach your organisation how AI work is done: who owns it, how data is prepared, how people check the output. Pick well and the second and third use cases come faster. Pick badly and the whole topic is set aside as “not ready yet”.

A good first candidate is usually an unremarkable routine process. Incoming invoices, supplier emails, contract intake, time entries, bank reconciliations, document classification. Processes that run every day, follow known rules and quietly consume hours.

Five criteria that separate good candidates from bad ones

  • Volume. The process runs often enough for small gains to add up. A task done hundreds of times a month beats one done three times a year, however painful.
  • Rules. Experienced staff can explain how they decide. If the answer is “it depends, I just know”, the process has to be described before it can be automated.
  • Data available. The inputs exist in digital form and you can reach them: emails, PDFs, records in your ERP or DMS. If the data sits on paper or in someone's head, start there.
  • Measurable outcome. You can state today's baseline: time per case, error rate, turnaround, backlog. Without a baseline there is no payback, only opinions.
  • Human in the loop. A person can review the AI's proposal before it takes effect, at a reasonable effort. AI proposes, people approve. This keeps risk low while trust is built.

Notice what is not on the list: strategic visibility, the most senior sponsor, the most advanced model. Those matter later. For the first use case, operational fit matters most.

A simple scoring method

Draw up a shortlist of five to ten candidate processes. Ask the people who run them, not only their managers. Then score each one on the five criteria.

  1. Rate volume, rules, data, measurability and review effort, each from 1 (weak) to 5 (strong).
  2. Treat data and measurability as gates. If either scores 1 or 2, set the candidate aside, whatever its total.
  3. Add up the scores and rank the remaining candidates.
  4. Test the top two with one question: if this worked perfectly, would anyone outside the team notice?

The method is deliberately rough. Its value is not precision. It forces a conversation about facts: how many cases, which systems, who checks. That conversation alone often eliminates many of the ideas under discussion, and the ones that remain are far easier to defend in front of a CFO.

Before anything is built, record the baseline for the selected process. Measure a few weeks of the current process: volume, time spent, errors, delays. This is the number your pilot will be judged against.

From pilot to production in weeks, not quarters

A pilot that runs for nine months is not a pilot. It is a project without a decision. Keep the first cycle short and explicit.

  • Weeks 1 to 2: define. Describe the process step by step: decision rules, exceptions, the review point. Agree on success criteria with the process owner.
  • Weeks 3 to 5: build. Connect the real data sources, configure the model and the workflow, and design the review screen so that checking a proposal takes seconds, not minutes.
  • Weeks 6 to 8: run in parallel. A small group works with the AI on live cases while the old process continues. Compare against the baseline. Fix what the edge cases reveal.
  • Decision. Go, adjust or stop. All three are valid outcomes. Drifting is not.

Timelines depend on your systems and your data, so read these as an order of magnitude, not a promise. The principle holds: real data from day one, a named owner, a fixed date for the decision.

Production adds three things a pilot can skip. Access rights that mirror your existing roles. Logging, so that every AI proposal and every human decision can be traced. And a short operating guide: who monitors quality, who handles exceptions, who approves changes. If personal data is involved, assess the processing against the Swiss FADP and, where you operate in or serve the EU, the GDPR and the EU AI Act. For most back-office use cases with human review this is manageable, but it should be documented, not assumed.

What the first use case really buys you

The hours saved in the first process matter. The bigger return is what your organisation learns: how to describe a process precisely, how to prepare data, how to design human review, how to measure. That capability is what makes the second use case cheaper and faster than the first.

This is the kind of work our AI integration team does with companies: mapping where time is lost, scoring the candidates and building the first agent into the Microsoft stack you already run, with Copilot, Azure AI Foundry or Power Platform, and models from OpenAI or Anthropic where they fit. The method works with or without us. Start small, measure honestly, and let the results decide what comes next.

Would you like to see it applied to your processes?

Thirty minutes with our team: we start from your situation, not from a presentation.