A human approval step does not create effective oversight by itself.

If the reviewer lacks time, evidence, authority, training, or a usable fallback, the approval button becomes a delay—or a rubber stamp. As AI systems prepare more decisions and take more actions, oversight must be designed as an operating capacity with demand limits, response targets, escalation coverage, and measurable quality.

The constructive goal is not to route every AI output to a person. It is to reserve qualified human attention for the decisions where judgment, accountability, or consequences require it—and to make sure that attention is actually available.

What changed

On October 15, 2026, the Financial Stability Board published its final Sound Practices for Responsible Adoption of Artificial Intelligence. The report sets out 12 practices for financial institutions, including governance, materiality assessment, data quality, explainability, performance management, human oversight, cyber resilience, and third-party risk.

Documented facts: The FSB says human oversight should be appropriate to the materiality, risk, autonomy, complexity, and explainability of each AI use case. Its release describes the practices as a toolkit that institutions can tailor and integrate into existing frameworks. They are sound practices, not a new law that specifies staffing ratios or review times.

Independent reporting: Reuters reported that the final guidance strengthened recommendations concerning agentic AI after consultation feedback and growing institutional interest. Reuters also reported that the FSB called for more intensive human oversight in higher-risk uses and clear, risk-based triggers for when an agent must seek approval.

Interpretation: Bridge Road’s operating lesson extends beyond finance: once AI volume exceeds the review capacity assigned to it, “human in the loop” can fail even when the workflow technically contains an approval step. That is an operational inference, not a claim made directly by the FSB.

Unresolved questions: The guidance does not prescribe reviewer-to-agent ratios, maximum queue age, shift coverage, evidence-screen design, override limits, compensation for review work, or the point at which an overloaded process should reduce or suspend automation. Each organization must define those controls for its own risks and volumes.

Why oversight fails in normal operations

An AI pilot may generate ten review items a week. A production workflow may generate hundreds during a promotion, month-end close, production disruption, supplier delay, cyber incident, or seasonal surge. If review demand grows faster than qualified capacity, queues age and business pressure rises.

That pressure creates predictable failure modes: approvers click through alerts, reviewers focus on easy cases while old exceptions accumulate, the same person prepares and approves a change, critical decisions wait behind low-risk items, or teams bypass the system to keep orders and work moving.

Bridge Road’s AI governance practice treats approval as one control in a larger operating design. Its operations practice defines ownership, staffing, service levels, escalation, and continuity so the control can function under real workload.

Classify the work before assigning reviewers

Start with risk-based routing rather than sending every decision to one shared inbox. For each action type, document:

• the business process, system of record, data classes, and affected parties;
• the financial, safety, legal, customer, production, and reputational consequences;
• whether the action is reversible, partially reversible, or effectively irreversible;
• the AI’s permitted autonomy and the deterministic checks already applied;
• the evidence a reviewer needs to make an informed decision;
• the qualifications and separation of duties required of the reviewer;
• the maximum response time and maximum queue age;
• the escalation path if the reviewer is unavailable or uncertain; and
• the conditions that pause the workflow or return it to manual processing.

Low-risk, well-validated work may use sampling or post-action review. Material price changes, safety-related work, unusual refunds, supplier changes, sensitive communications, production releases, and irreversible actions may require explicit pre-action approval. The model should not decide which control tier applies to its own action.

Plan review capacity like any other constrained operation

Estimate the expected arrival rate, review time, skill mix, peak volume, coverage hours, and rework rate for each queue. Then test the plan against abnormal conditions rather than average demand.

A practical capacity record should include:

• items expected per hour, day, and peak event;
• median and high-percentile review time by risk tier;
• number of qualified primary and backup reviewers;
• coverage for absence, turnover, nights, weekends, and holidays;
• maximum work-in-process and oldest acceptable item;
• time reserved for investigation, not only approval clicks;
• training and recertification requirements;
• escalation capacity for complex or disputed cases; and
• the safe processing rate when AI or integration systems are unavailable.

Capacity should include the exception tail. A workflow that handles ordinary cases quickly but overwhelms specialists with ambiguous cases may shift work rather than remove it.

Give reviewers evidence and real authority

The approval interface should present the proposed action, affected records, authoritative source data, deterministic validation results, relevant policy, confidence or uncertainty information where meaningful, prior comparable decisions, and expected downstream effects. Do not force the reviewer to reconstruct the case across several systems.

Reviewers need authority to approve, reject, edit, narrow the scope, request evidence, escalate, pause the agent, revoke a credential, or route the work to the normal manual process. Record the decision, reason code, edits, timestamps, evidence viewed, and downstream transaction identifiers.

Use least privilege for both the AI and the reviewer. A person authorized to review a product description may not be authorized to change price or inventory. A maintenance planner may approve a routine work order but not bypass a safety interlock. Separate preparation, approval, and execution where the consequences warrant it.

Monitor the oversight system, not just the model

Track operational measures that show whether human review remains effective:

• queue volume, arrival rate, completion rate, and oldest item;
• review time by risk tier and reviewer;
• approval, rejection, edit, escalation, and timeout rates;
• decisions completed without required evidence;
• overrides of deterministic checks or policy limits;
• repeat exceptions and disagreement between reviewers;
• post-approval defects, reversals, rework, and customer or operational impact;
• reviewer workload, concentration, absence coverage, and training status; and
• the proportion of actions sampled by an independent second line.

A sudden approval-rate increase may mean the model improved, but it may also indicate reviewer fatigue or weak evidence. A falling review time is not automatically good. Interpret the measures together and sample decisions for quality.

Use load-shedding rules before the queue becomes unsafe

Define deterministic thresholds that reduce automation when oversight capacity is constrained. The workflow might lower batch size, stop accepting high-risk actions, shift from execution to recommendation-only mode, require a second reviewer, extend customer promises, or return selected work to the established manual process.

Escalation should occur before the service target is missed. Name the person who can add capacity, change the risk tier, suspend a workflow, or accept a temporary exception. Time-limit every exception and require a documented reason, compensating controls, and review date.

During an incident, preserve the AI inputs and outputs, model and configuration versions, validation results, reviewer actions, credentials used, downstream writes, and affected business records. The incident owner should be able to stop new work without destroying the evidence needed to understand completed work.

Apply the capacity model to operating workflows

Manufacturing: Route routine, validated maintenance recommendations to planners during staffed hours. Safety-related work, production releases, and unexpected equipment conditions go to qualified specialists. If the queue exceeds its limit, the agent stops releasing new work and technicians continue with the established CMMS process. Bridge Road’s industrial automation practice can map those boundaries.

Industrial distribution: Standard quotes within margin, credit, availability, and substitution rules may receive streamlined review. Unusual quantities, export restrictions, custom items, price overrides, or new accounts route to sales operations, finance, or compliance with defined response times.

eCommerce: Low-risk catalog enrichment can use sampled review, while regulated claims, price changes, refunds, customer messages, and high-traffic product changes require stronger approval. The commerce workflow should preserve prior versions and support rapid reversal.

Business systems: Reporting and case preparation may use post-action sampling, but changes to master data, access, payments, integrations, or records of truth require qualified reviewers and correlation IDs linking approval to execution. Bridge Road’s systems practice defines those handoffs.

A practical implementation sequence

1. Inventory every AI workflow that depends on human review or exception handling.
2. Assign a materiality and autonomy tier to each action type.
3. Define approval triggers, reviewer qualifications, evidence, and separation of duties.
4. Measure actual arrival rates, review times, exception volume, and peak demand.
5. Staff primary, backup, and escalation coverage with explicit service targets.
6. Set queue limits, aging alerts, load-shedding rules, and manual fallback.
7. Test overload, absence, conflicting evidence, partial failure, and incident scenarios.
8. Review quality and capacity together before increasing AI autonomy or volume.

Human oversight should enable adoption, not create permanent friction. When the review system has qualified capacity, good evidence, real authority, and a tested fallback, the business can automate more confidently because it knows when people must intervene and whether they can do so in time.

Schedule a conversation to turn a “human in the loop” requirement into a staffed, measurable operating control.