How should a public institution choose its first AI project?
· BEXTIO
A public institution’s first AI project has disproportionate importance.
If it is too broad, too risky or poorly measured, it can consume significant effort without producing useful evidence.
If it is too trivial, it teaches very little.
The right first project sits between those extremes.
BEXTIO recommends starting with a problem that matters enough to users but is narrow enough to be tested within a few weeks.
The objective is not to “deploy AI.” It is to verify whether a specific use case improves a specific process.
Seven questions help filter candidates.
First: is the problem clearly described? “We want AI” is not a problem. “Our teams spend several hours finding the right procedure in a fragmented corpus” is.
Second: can the current situation be measured?
Without a baseline, it will be difficult to demonstrate improvement.
Third: do the required data or documents exist, and can they be used legally and operationally within the proposed scope?
Fourth: are the users identified?
A project without committed users quickly becomes a demonstration disconnected from real work.
Fifth: is the risk manageable?
For a first project, it is often preferable to avoid high autonomy or systems that directly influence significant decisions affecting people.
Sixth: can clear human validation be maintained?
Early experiments should make it easy to compare AI output with the judgment of a competent employee.
Seventh: is a decision planned at the end?
A pilot should lead to a conclusion: stop, redesign, extend, scale gradually or prepare deployment.
A simple scoring framework can assess each candidate according to value, feasibility, data readiness, risk, adoption, measurability and time to evidence.
In many cases, document search, case-file summarization, meeting preparation or support for an internal workflow can be good starting points because they allow the institution to test usefulness without transferring too much responsibility to technology too early.
The first project should primarily build capability: learning how to scope, test, document, govern and decide.