Should an AI testing agent say PASS, FAIL, or I DON’T KNOW?
One challenge we faced while building an AI testing agent was deciding what it should do when the evidence is incomplete.
A binary PASS or FAIL result looks simple, but it can create false confidence.
For example, a page may display an error message that is expected in one workflow but is a real defect in another. If the agent does not have enough context, forcing a binary result can create a false positive.
We decided to use three verdicts:
PASS: The expected behaviour was verified.
FAIL: The agent found enough evidence to confirm a defect.
INCONCLUSIVE: The evidence was not strong enough to make a reliable decision.
I would rather receive an honest INCONCLUSIVE result than a confident but incorrect bug report.
How does your team handle uncertain automated test results?
Replies
@tayyabakmal Great point! Forcing a binary 'pass/fail' on an AI agent often leads to forced hallucinations or over-confidence when things get ambiguous. Embracing 'I don't know' as a first-class citizen—or a trigger for human-in-the-loop—is essential for building real trust in agentic testing.
350+ E-Commerce Tools Database
I've also seen false negatives spiral into a token-burning loop which also makes the agent push wrong and possibly destructive decisions/actions. Just the other day, I had Claude running in the background and it decided to delete the contents of the database it was working on because something wouldn't run the way it expected. In every way, it was doing the wrong thing and never stopped to ask a question or reconsider that drastic move.
It then decided to break out of the project folder and find a replacement database which I had stored in a different folder, and it restored it. So that one false negative, where everything was working but it got stuck made it try some wild and very time consuming avenues that weren't needed.
I think having a system way out to catch ambiguous or unsure outcomes would definitely help BUT you want to balance that with the confidence to make a concrete binary decision when appropriate.