TL;DR: Manual review still has a role in ecommerce fraud detection, but Signifyd cites Merchant Risk Council data showing 35% of fraud teams want to reduce or eliminate it, and says merchants now screen twice as many orders digitally as manually. Human review remains useful for edge cases, but machine-led decisioning is becoming the operational default.
At a glance
What this is: This is a fraud-operations analysis arguing that manual review is becoming a secondary control as machine learning takes the lead in merchant decisioning.
Why it matters: It matters because fraud teams need to redesign thresholds, escalation paths, and review capacity around automated decisioning rather than human triage as the default.
By the numbers:
- 35% of fraud management teams cite reducing or eliminating manual fraud review as an area for improvement, according to Signifyd.
- A manual review costs $3.47 per transaction on average, according to Signifyd.
- A review-to-decline rate of less than 10% is a strong indicator to increase automation thresholds, according to Signifyd.
Context
Manual fraud review is a governance choice, not just an operational habit. Merchants use it when automated systems cannot confidently separate legitimate customers from suspicious transactions, but that fallback becomes expensive when review volume grows and thresholds are never reset.
The identity angle here is commercial rather than IAM, but the control problem is familiar: humans are being asked to approve exceptions at scale after the machine has already done most of the sorting. That is a poor fit for high-volume commerce, where speed, consistency, and measurable thresholds matter more than artisanal case handling.
Key questions
Q: How should merchants reduce manual fraud review without increasing fraud risk?
A: Merchants should first segment transactions into stable, low-risk decisions and true exceptions. Then they should automate the routine cases, keep human review for ambiguous orders, and monitor decline rates, override rates, and false positives. The goal is not zero review. It is to make manual review an exception-control function rather than a default operating model.
Q: Why does manual fraud review become expensive at scale?
A: Manual fraud review becomes expensive because each case consumes analyst time, adds delay to order fulfilment, and increases the chance of inconsistent decisions. Even a well-run team cannot match the throughput of automated scoring when volumes rise. The hidden cost is not only payroll. It is abandonment, lost conversions, and the operational drag of routing too many normal decisions to people.
Q: What are the signs that fraud review thresholds are too low?
A: A low review-then-decline rate is a strong sign that many cases sent to analysts could have been handled automatically. If most reviews end in approval or routine decline without adding new context, the team is likely overusing manual review. That creates friction without improving decision quality.
Q: When should human review still be part of fraud decisioning?
A: Human review still makes sense when transactions are unusual, high-value, or contain conflicting signals that require context a model may not capture well. It is best used as a bounded exception path, not as the main decision layer. The goal is to preserve expert judgment for edge cases, not routine approvals.
Technical breakdown
How manual fraud review works in merchant decisioning
Manual fraud review is a human adjudication layer applied after a transaction is flagged by rules, heuristics, or machine scoring. Teams usually inspect order value, shipping and billing mismatch, customer history, payment data, and verification signals before deciding to approve, decline, or escalate. The key limitation is that manual review depends on analyst judgment and available context, so its output is inherently slower and less consistent than a trained model operating against broad data patterns.
Practical implication: reserve human review for genuinely ambiguous cases and define clear escalation criteria so analysts do not become the primary decision engine.
Why machine learning changes the fraud control model
Machine learning fraud systems change the control from after-the-fact inspection to continuous scoring and automated decisioning. Instead of waiting for a reviewer to investigate, the model can process large volumes quickly, use broader network-level data, and update decisions as patterns change. That does not remove the need for governance. It shifts oversight toward model calibration, threshold tuning, exception handling, and ongoing measurement of false declines and review rates.
Practical implication: treat the model as the primary control and manage its thresholds like an operational policy, not a one-time deployment.
Thresholds, review-then-decline rates, and decision quality
The most useful metric in the article is the review-then-decline rate, which measures how often manual review ends in a decline. If that rate is low, it means many reviewed orders could have been approved or declined automatically, and the team is spending human effort on decisions the machine could already make. Threshold management is therefore a governance problem, because poor thresholds create either excess friction or excess risk.
Practical implication: review your decline-to-review ratio regularly and adjust automation thresholds when human review is not materially improving decision quality.
NHI Mgmt Group analysis
Manual review fatigue is a governance problem before it is a staffing problem. When fraud teams stay too dependent on human adjudication, the control stops scaling with transaction volume and becomes a bottleneck for both risk and revenue. The article’s central signal is that merchants are already moving toward digital screening because the exception workflow has outgrown the exception rate. Practitioner takeaway: design the operating model so humans handle edge cases, not the main line.
Review-to-decline rate is a more useful control signal than review volume alone. A team can process many reviews and still add little value if most of them end in outcomes the model could have reached. That is why threshold governance matters: it tells you whether manual review is truly improving precision or simply adding delay. Practitioner takeaway: manage fraud review as a measurable decisioning control, not a customer service queue.
Machine-led decisioning changes where fraud governance lives. The governance burden shifts from case handling to model oversight, threshold tuning, and exception policy. That is a healthier structure for scale, but only if fraud, operations, and customer experience teams agree on what triggers human intervention. Practitioner takeaway: move governance upstream into model policy and keep manual review as a bounded control.
Manual review is still valuable, but only when the exception set is genuinely small. The article correctly preserves a role for human judgment in unusual or high-value cases, yet the operational default has changed. That means fraud programmes should stop treating manual review as the primary line of defence and start treating it as a scarce specialist resource. Practitioner takeaway: protect analyst capacity for the cases that really need context.
From our research library:
- Nearly 60% of companies reported that fraud losses were still increasing in 2025.
- Read next: Agentic AI Compliance Guide
What this signals
Manual fraud review should now be treated as an exception-control design problem. The more merchants move to machine-led decisioning, the more important it becomes to define where human intervention is allowed and what evidence justifies it. For teams running large-scale commerce, the goal is not to remove people from fraud operations, but to stop using them as the primary throughput mechanism.
Fraud governance increasingly resembles policy tuning rather than case processing. That shifts ownership toward fraud operations leaders who can balance false declines, conversion, and analyst capacity without letting any one metric dominate. Practitioner takeaway: review thresholds, escalation rules, and analyst roles together, because they form one control system.
For practitioners
- Define review as an exception workflow Set explicit criteria for when a transaction leaves automated decisioning and enters human review, such as unusual value, mismatched identity signals, or high-risk geographies.
- Track review-to-decline performance Measure how many reviewed orders end in a decline and use that ratio to decide whether the automation threshold should move higher or lower.
- Use model managers, not manual clerks Recast fraud analysts as model managers who tune thresholds, monitor drift, and investigate the cases that automation cannot resolve with confidence.
- Reduce friction in the customer journey Remove avoidable manual checks from routine orders so legitimate customers are not delayed by decisions that automated systems can make instantly.
Key takeaways
- Manual fraud review still has a valid role, but only as a bounded exception path when automated decisioning cannot confidently resolve a suspicious order.
- The operational issue is not whether humans add value, but whether they are being used where models are already accurate enough to decide.
- Merchants that keep manual review as the default risk higher cost, slower fulfilment, and weaker customer experience than teams that lead with automation.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
NIST CSF 2.0, NIST SP 800-53 Rev 5 and CIS Controls v8 set the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| NIST CSF 2.0 | PR.AA-05 — Access Permissions, Entitlements and Authorizations | Fraud decisioning depends on who or what is authorised to approve an order. |
| ID.RA-01 — Asset vulnerabilities are identified and documented | Fraud review thresholds depend on identifying weak signals and error-prone cases. | |
| Recommendation — Use PR.AA-05 to govern which identities can override automated fraud decisions. Document the fraud signals that make manual review necessary and keep them under review. | ||
| NIST SP 800-53 Rev 5 | IA-5 — Authenticator Management | Manual review often relies on customer verification and authenticator checks. |
| Recommendation — Apply IA-5 to ensure verification steps are consistent and not left to ad hoc analyst judgment. | ||
| CIS Controls v8 | CIS-5 — Account Management | Fraud operations often depend on account and identity signals that need lifecycle control. |
| Recommendation — Use account management controls to keep identity data clean enough for reliable fraud decisioning. | ||
Key terms
- Manual Fraud Review: A manual fraud review is a human-led assessment of a transaction or account to decide whether it should be approved, rejected, or escalated. It is typically used for edge cases that automated scoring cannot resolve with enough confidence, but it becomes inefficient when used for routine decisions.
- Machine-Led Decisioning: Machine-led decisioning is a fraud control model in which automated scoring and policy thresholds make the first decision on most orders. Humans still handle exceptions, but the operating logic is designed so the model carries the main throughput and analysts focus on edge cases.
- Review-to-Decline Rate: Review-to-decline rate measures how many manually reviewed transactions are ultimately declined. A low rate usually means the review queue is catching too many cases that automation could have handled, while a higher rate can indicate that human review is finding meaningful exceptions.
Deepen your knowledge
NHI Mgmt Group covers identity security, NHI governance, and agentic AI through independent research, practitioner guides, and the NHI Foundation Level course. Explore nhimg.org for resources that connect identity governance to the broader security disciplines your programme depends on.
Published by the NHIMG editorial team on July 11, 2026.
Updated on October 11, 2026.
NHI Mgmt Group, the independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org