Troubleshooting when AI output doesn't look right

Troubleshooting when AI output

doesn't look right

The most common reasons AI scores, briefs, or recommendations look off, and how to fix each one.

Start with the brief, not the score

When something looks wrong, the temptation is to stare at the score. The score is the symptom. The brief (the 3-column view) is

where the diagnosis lives. The brief shows what the AI actually saw and how it weighed it.

Open the candidate's profile, read column 1 (Why They Match) first, then column 3 (Things to Consider). Most of the time, the

explanation for a surprising score is in one of those two columns. The rest of this article assumes you have done that and the

brief still does not explain the output.

Symptom: scores are uniformly low across the pool

The most likely cause is Core Criteria that are too narrow or too specific.

A criterion like "exactly five years of experience in Series B fintech in the GCC region" will produce low scores across the board,

because almost no candidate matches it precisely. The AI is doing its job; the criteria are unrealistic.

The fix is to widen the criterion. Replace exact ranges with thresholds ("five plus years" instead of "exactly five"), replace narrow

industries with broader categories ("financial services" instead of "Series B fintech"), and split single criteria into two

("experience in finance" weighted High, "GCC region experience" weighted Medium).

A useful sanity check: aim for a top ten candidates with Resume Match in the 70 to 95 range. If your highest-ranked candidate is

below 70, the criteria are too narrow. If your lowest-ranked candidate is above 80, the criteria are too loose.

Symptom: scores are uniformly high across the pool

The mirror image. Core Criteria are too vague, too few, or weighted too generously. The AI cannot differentiate candidates

because the criteria do not differentiate them.

Two fixes. The first is to add criteria. A job with two criteria will produce a pool that all scores the same; a job with six well-written

criteria will produce a meaningful ranking. The second is to recheck weights. If every criterion is weighted High, the High weight

loses its meaning.

Symptom: one candidate's score is wildly off from what you expected

Three things to check, in order.

Did the resume parse correctly? Open the candidate profile and look at the parsed fields next to the original CV. If the parsing is

wrong (a job title appearing as a skill, an Arabic CV that parsed left-to-right, dates that got mangled), the AI is scoring against

bad input. For Arabic CVs and CVs that use non-standard layouts (graphic-heavy designs, two-column layouts with images),

parsing can occasionally fail. The fix is to click "Reparse CV" in the profile actions, which runs the parser again and usually

resolves the issue.

Are the criteria actually being matched? Read column 1 of the brief carefully. If the AI thinks the candidate matches on a criterion

that they obviously don't, the criterion is probably worded ambiguously. Edit the criterion to be more specific.

Did a knockout question fire? Knockout questions on a Skills Assessment auto-flag candidates as disqualified regardless of their

Resume Match. A candidate with a strong CV and a low overall ranking is sometimes the result of a knockout that wasn't

intended to be one. Check the assessment results on the profile.

Symptom: the recommendation says Advance but the brief says otherwise

The recommendation is a synthesis. It weighs both scores, the brief, and any assessment results. Occasionally it surfaces an

Advance on a candidate whose Things to Consider column is long, because the bonuses or the score weight outweigh the

concerns in the AI's calculation.

If this happens often enough to be noticeable, the Core Criteria probably do not include things you actually care about. Add

criteria for the gaps you keep seeing in column 3. The AI cannot weigh what it was not told to look at.

You can also suppress the recommendation on a specific candidate (covered in the previous article). That is a fix for the

individual case, not the pattern.

Symptom: the AI Skill Match is much lower than the Resume Match

This is one of the more useful disagreements between the two scores. It usually means the candidate's CV reads well for the role

but the actual underlying experience does not align as strongly. AI-assisted resume writing produces this pattern often.

This is not a problem to fix. It is the dual scoring system doing exactly what it was built to do: catch candidates whose document

is stronger than their experience. Read the 3-column brief, especially column 3, and decide whether the assessment results (if

you have them) or an interview can resolve the gap.

The reverse case (Skill Match much higher than Resume Match) is the more interesting one. It usually means the candidate is

being underestimated by their own CV. Worth a closer look.

[Illustration: A two-column "if/then" table. Left column lists six symptoms in Forest. Right column lists the first action to take for

each, in shorter text. Alternate row backgrounds in Signal White and Signal Mint. Sage Gray border. Caption at the bottom: "Start

here. Most issues resolve in one step."]

When to contact support

Most AI issues are configuration issues, and configuration issues are faster to fix yourself than to escalate. Contact support

when:

The same parsing failure happens repeatedly on candidates from a specific source.

A score changes without any change to Core Criteria, candidates, or assessments.

The Assistant returns inconsistent answers to the same question on the same tab within the same session.

Anything that looks like the AI is doing something you cannot explain by reading the brief and the criteria.

The full path is in Contacting Talinty support.