Manual Review
Manual review is the step where a human decides a case an automated system would not. In identity verification it catches what sits between clear approve and clear decline — and its volume is the most honest measure of how well the automation underneath it is working.
| What triggers it | A score in the uncertain band, a policy rule, or a regulatory requirement |
| Typical decisions | Approve, decline, or request additional documentation |
| Cost per case | Analyst time, measured in minutes |
| Cost to the customer | Hours to days of waiting, and a meaningful share abandon |
| Quality risk | Inconsistency between reviewers on identical cases |
| Compliance angle | Override rationale must be documented and defensible to an examiner |
| Healthy rate | Low single-digit percentage for consumer onboarding |
| What a high rate means | Thresholds nobody trusts, or capture quality too poor to decide on |
How it works
A case reaches review one of three ways, and they need separating because they mean different things.
Score-band routing sends cases the model found truly ambiguous. This is manual review working as designed — a human resolving what the system could not.
Policy routing sends categories to a human regardless of score: high-value applications, certain jurisdictions, politically exposed persons. The system may be entirely confident; the institution has decided a person should look anyway.
Failure routing is the one that matters operationally. Cases arrive because the image was too poor to assess, the document type was unrecognized, or a check timed out. These are not judgment calls. They are capture and coverage problems being absorbed by an expensive human process, and they are usually the largest share of the queue in a struggling flow.
The reviewer sees the document images, the extracted data, the check results, and whatever risk signals fired — then approves, declines, or asks for more. The quality question is consistency: two reviewers given the same case should reach the same conclusion, and without clear criteria and regular calibration they frequently do not.
Why it matters for identity verification
Manual review rate is the single most diagnostic number in a verification program, and it is worth reading rather than just minimizing.
A high rate is not evidence of caution. It usually means thresholds are set where nobody believes the automated decision, or that capture quality is poor enough that a large share of images cannot be assessed at all. Both are fixable upstream, and neither is fixed by adding reviewers.
The customer cost is larger than the analyst cost and rarely measured. An applicant told to wait while their identity is reviewed frequently does not come back, and the ones who abandon are disproportionately legitimate — a fraudster is patient because the payoff justifies it. So a review queue filters out good customers more efficiently than bad ones.
There is a compliance dimension too. Examiners look at override rates and at whether overrides were consistent and documented. High or inconsistent overrides suggest a program whose own thresholds are not trusted. Configurable thresholds with recorded rationale is what makes that defensible, and Microblink’s verification workflow is built to keep the band narrow by resolving capture problems before they become review cases.
Manual review vs step-up verification
| Manual review | Step-up verification | |
|---|---|---|
| Who resolves it | An analyst, later | The user, immediately |
| Customer experience | Wait, with no visible progress | One additional check, then a decision |
| Time to resolution | Hours to days | Seconds |
| Abandonment | High | Low |
| Cost | Analyst time per case | Marginal |
| Best suited to | Genuine edge cases and policy-mandated review | Uncertainty the user can resolve themselves |
A large share of what sits in review queues is not an edge case — it is uncertainty the applicant could clear in seconds by recapturing a document or completing a liveness check. Routing that to a human rather than back to the user is the commonest avoidable cost in a verification flow.
What it can’t do
It cannot make humans better than the evidence. A reviewer looking at a blurred, glare-covered image is guessing more carefully than the model did. Where the information is not in the image, the human adds process rather than accuracy.
It cannot outperform automation on document authenticity at scale. Reviewers are good at contextual judgment and worse than a trained model at detecting subtle forgery across thousands of document types. Manual review is a judgment layer, not a superior detection layer.
It cannot stay consistent without calibration. Reviewer agreement drifts with fatigue, tenure and caseload. Programs that do not periodically test reviewers against the same cases usually do not know how inconsistent they have become.
It does not scale. Volume doubles and the queue doubles. Any strategy that relies on manual review absorbing growth is deferring a problem rather than solving it.
Frequently asked questions
What is a good manual review rate for identity verification?
Low single digits for consumer onboarding. Anything substantially higher usually points at capture quality or threshold settings rather than at truly ambiguous customers, and is better fixed upstream than staffed for.
Why do identity verifications go to manual review?
Three reasons worth separating: the model was truly uncertain, policy requires human review for that category, or the submission could not be assessed at all because of image quality or an unrecognized document. The third is usually the largest and the most fixable.
Does manual review improve accuracy?
For contextual judgment, yes. For detecting document forgery across many document types, trained models generally outperform human reviewers. Manual review adds judgment, not better forensics.
How does manual review affect conversion?
Substantially, and asymmetrically. Applicants asked to wait abandon at high rates, and those who abandon skew legitimate — fraudsters wait because the payoff justifies it. Reducing review volume usually improves both conversion and fraud outcomes.
Related reading
- Document capture — the upstream cause of most review volume
- Identity document verification — the automated decision review sits behind
- Liveness detection — a step-up check that resolves cases without a queue
- Compliance officer — who answers for override rates at examination
