AI Project: Define measurable objectives before the pilot
A practical sheet for choosing an AI use case, measuring current work and evaluating a trial with clear criteria.
An artificial intelligence project becomes evaluable when it starts from precise work. “We want to use AI” describes a tool; “We want to reduce the time needed to find the correct procedure” describes a problem. The difference allows you to choose relevant data, people and checks, before discussing which model to adopt.
This guide offers a worksheet for operations managers and digital teams. The method and examples are application proposals, not results obtained from EL-AI customers. The goal is to arrive at a reasoned decision: start a trial, clarify the process or postpone automation.
Describe the task from the point of view of the person carrying it out
Choose an activity with a recognizable beginning and end. For example: an operator receives a question about a product, searches for information in a defined set of documents and prepares an answer. “Improve customer service” is too broad for a first attempt; «find the relevant configuration procedure» instead allows you to observe a concrete result.
Write down who receives the output, what they need to do with it, and what information is needed. Add exceptions: incomplete application, missing document, out of catalog product. Often these conditions explain the work better than the ideal case shown in a presentation.
Measure the current process before testing
Collect a small set of representative cases without selecting only the easiest ones. For each one, write down the search time, the control time, the number of steps between people and the final outcome. There is no need to start with a complex dashboard: a shared table can be enough to make observations comparable.
A hypothetical example: today a search takes twelve minutes, of which four are to verify that the document is up to date. If the AI produces an answer in a few seconds but requires fifteen minutes of corrections, the speed of generation does not represent an improvement in the task. What counts is the work needed to obtain a usable result.
Prepare a one-page decision sheet
- Task: which activity is assisted and which remains entrusted to the person.
- User: who uses the result and with what skills.
- Input: necessary documents, fields and updates.
- Acceptable output: required content, format, sources and level of completeness.
- Major error: which error makes the result unusable.
- Responsible: who decides whether the test can be extended.
The tab should also contain a stopping criterion. If the starting information is systematically contradictory, continuing to change the prompt can hide the problem. In that case the first useful action is to sort out the sources or clarify who decides which version to use.
Compare realistic alternatives
Next to the AI solution consider a better search, a clearer form or an automatic rule. The operational question is which intervention solves the problem with sustainable complexity. Rule-based processing may be suitable for a stable format; an assistant can be useful when the request comes in natural language and requires consultation.
Include maintenance, human control and error handling in your comparison. Decide who will update the document collection and how a deterioration will be recognized. A test that only works thanks to the continuous attention of those who built it has not yet proven to be ready for daily work.
Close the pilot with a readable decision
Prepare three possible outcomes: extend, fix and retry, stop. Link each to concrete observations. A good final report reports which cases were tried, where the system was useful, where it failed, and which conditions were not explored. Avoid percentages without a denominator or judgments such as "it seems smart".
The NIST AI Risk Management Framework provides a voluntary framework for integrating trust considerations into the design, use, and evaluation of AI systems. It does not constitute a certification of the individual project: it is referred to here as a reference framework, while the proposed sheet is an operational exercise.
To prepare a comparison with EL-AI, bring a real activity, some usable examples, and how the result is verified today. You can book an appointment starting from this tab. For the next step, read how to build a test for an AI assistant.
Content prepared with AI assistance; sources consulted on September 19, 2026. Numerical examples are hypothetical and do not represent product performance EL-AI.
Illustrative AI-generated cover; it does not depict actual EL-AI people, premises or installations.