Connect a candidate record to the intended SKU or manufacturer part number.
Product scope
Product systemP4Confidence spectrum
AI confidence scores · definition
How AI Confidence Scores Work for Product Data
An AI confidence score is a signal expressing how certain a system is about a classification, extraction, match or other automated decision.
In product-data workflows, it can help prioritize which records move automatically and which require review.
A confidence score is useful only when its meaning is defined. A number such as 0.92 does not automatically mean “92% factually correct.”
See confidence in the workflow →Confidence spectrumIllustrative signal
Example0.92Task context required
Product match
Source match
Extraction
Classification
The same number can represent different signals. The task, evidence and decision must travel with it.
01 Signal02 Meaning03 Limits04 Source05 Extract06 Gate07 Review08 Example09 FAQ10 Related
02 · Task-scoped meaning
One score · several questions
The meaning depends on the model and the task
A score may describe confidence in a match, an extracted value, a source or a category assignment. The interface should name the task instead of presenting a generic number.
Assess whether a candidate page or document is relevant to the identified product.
Evidence scopeSeparate explicit specifications from values that require interpretation.
Field scopeEstimate confidence in assigning the product to the target taxonomy.
Taxonomy scope
03 · Interpretation boundary
Confidence ≠ correctness
Confidence does not replace source evidence, validation or ground truth
A system can be confident and still be wrong when the input is ambiguous, the source is incorrect or the learned pattern is misleading.
- Attached to one task
- Interpreted with evidence
- Compared inside a calibrated workflow
≠
- Not source proof
- Not factual validation
- Not a claim of ground truth
A displayed score of 0.92 must not be read as 92% factual accuracy unless a validated calibration explicitly supports that interpretation.
04 · Source match signal
Candidate source
How well does a candidate source match the known product?
Source confidence should remain connected to the identity signals and context that produced it.
Candidate sourceReviewable
- Manufacturer reference
- Matched
- Brand context
- Consistent
- Technical context
- Relevant
- Exposed score logic
- ENRIVAQ evaluates how well available evidence matches the target product and field context. Exact scoring logic remains internal.
05 · Extraction signal
Explicit vs ambiguous
A stated specification is not the same as an inferred one
Extraction confidence should describe how clearly a value is expressed in the available evidence.
“EAN 8716106986118”
attribute EANvalue 8716106986118unit —
Extract→
“large fan assembly”
attribute unclearvalue not statedunit —
Illustrative evidence states only; no production confidence value is implied.
06 · Threshold calibration
Decision checkpoint
Thresholds turn a signal into a workflow decision
Thresholds must be calibrated for the task and risk. A production rule cannot be inferred from an illustrative score.
Incoming signalTask + evidence + confidenceNo universal threshold
Calibration gatePolicy + benchmark + risk
NO SINGLE PUBLIC THRESHOLD · Supported evidence → continue · Ambiguous/conflicting evidence → review · Invalid evidence → reject
07 · Human review router
Exception workflow
Confidence should route attention, not decorate the interface
Reviewers need the candidate value, its source and the reason the record was routed to them.
Human decision
Routine cases can continue through the approved workflow. Exceptions stay visible and reviewable.
08 · Field-level example
One record · three actions
Three fields, three confidence states, three actions
The label belongs to the automated task, not to the product record as a whole.
Manufacturer referenceIdentity matchAC820825Clear matchVerified referenceContinue
WeightAttribute extraction1.74 kg / 2.60 kgConflictingSources disagreeReview
CompatibilityRelationship inferenceAdditional model candidateUnsupportedInsufficient contextDo not publish
09 · FAQ
Interpretation guide
Common questions about AI confidence scores
Use the score as a task-scoped signal and keep it attached to the evidence and decision.
Not unless the score has been calibrated and that interpretation has been validated for the specific task.
Not automatically. They can remain visible as candidates and be routed to review or left unresolved.
No. Confidence is a signal; source evidence is required to understand and review the proposed value.
The threshold should be based on the task, risk and measured workflow performance. No universal number is implied here.