A practical test for AI in institutions: would you accept it from a junior analyst?
AI is most useful where its output can be checked quickly and its mistakes are cheap to correct. That rules in more than sceptics think, and less than vendors claim.
- Author
- Data & Intelligent Solutions PracticeGayduo Advisory
- Published
- Reading time
- 1 min
- Topic
- AI & automation

Leadership teams in ministries, banks and development organizations are being asked the same question: what are we doing about AI? The pressure produces pilots, and pilots produce demonstrations. Far fewer produce something that runs in daily operations a year later.
We use a simple test when deciding where AI belongs: would you accept this work from a capable junior analyst, provided someone experienced reviewed it? If the answer is yes, AI may help. If the work requires judgement that you would not delegate to a junior colleague, it should not be delegated to a model either.
Where the test says yes
Reading standard documents — invoices, applications, forms — and extracting fields for a person to confirm.
Finding relevant passages across thousands of reports, with the source cited so it can be checked.
Drafting routine correspondence from structured information for a person to approve.
Sorting incoming requests so they reach the right desk faster.
Where it says no
Decisions about eligibility, sanctions, credit or legal position should not be made by a model, however good the demonstration. These are areas where errors are expensive, hard to detect and, in public institutions, carry obligations of explanation that current systems cannot meet.
What makes it hold up
Every AI component we deploy has an evaluation set drawn from the client's own documents, a measured accuracy before go-live, a log of what it produced, and a person accountable for reviewing exceptions. Those controls are what allow an institution to explain its use of AI to an auditor, a board or a court.


