Adding a person somewhere in an AI workflow does not automatically create human oversight. The person needs information, time, authority and a real opportunity to choose a different outcome.
The right question is not whether a system has a human in the loop. It is which decision the person controls, before what consequence and with which evidence.
Identify the actual decision and consequence
Separate prediction from decision and decision from action. A model may classify a request as suspicious; a policy may hold the account; an employee may decide whether to release it.
ICO guidance on meaningful human review says that simply supplying input to an automated system does not make a later significant decision human-determined. Human involvement must relate to the actual outcome.
Document who can be affected, what changes for them, whether the effect is reversible and how quickly an error becomes harmful.
Use risk tiers instead of one universal rule
Not every generated summary needs approval. Oversight should increase with the consequence, uncertainty, sensitivity and difficulty of reversal.
- Low risk: internal drafts, search assistance and formatting with sampling and easy correction.
- Moderate risk: customer-facing suggestions, record updates and operational routing with validation and exception review.
- High risk: decisions affecting rights, access, employment, credit, health, safety, money or irreversible external actions with mandatory qualified approval.
- Prohibited or unsuitable: uses where the business cannot provide lawful authority, adequate evidence, competent review or effective remedy.
Place review before the consequence
A reviewer who sees the result after a payment, rejection or public message has already been sent is monitoring, not approval.
Design an explicit pending state. The AI produces a proposal, the policy engine validates it, a person reviews it and only then can a separate execution service act.
NIST AI RMF notes that human roles and responsibilities in AI decision-making and oversight need to be clearly defined and differentiated across configurations from manual to autonomous.
Give reviewers evidence, not just a score
A confidence number without context encourages automation bias. Show the relevant source records, rules applied, missing information, conflicting evidence, known limitations and the proposed action.
For AI document processing, the reviewer should see the original page next to extracted fields, validation errors and calculation checks.
Hide irrelevant personal data, but do not hide the evidence needed to challenge the recommendation.
Give the human authority and safe alternatives
A reviewer must be able to approve, reject, edit, request more information, escalate and pause the workflow. If the only practical button is approve, the review is ceremonial.
Set service targets that allow review. Punishing staff for slowing automation makes independent judgement impossible.
The OECD AI principle on human-centred values calls for safeguards that preserve human agency and oversight in a manner appropriate to context.
Escalate by consequence and uncertainty
Do not rely on model confidence alone. Combine deterministic triggers: sensitive data, protected category, unusually large value, policy exception, new customer, conflicting sources, out-of-distribution input or attempted tool misuse.
Route routine exceptions to trained operators and the most consequential or novel cases to a named senior owner or specialist.
A suspected prompt injection should stop the affected tool sequence rather than ask the same model whether it feels safe.
Use dual control for irreversible actions
For high-impact actions, separate proposal, approval and execution across roles or services. The person who reviews evidence should not share an unrestricted credential with the model.
Examples include deleting data, changing permissions, releasing funds, signing a commitment, publishing sensitive content or sending a message to a regulated recipient.
Dual control is not required everywhere. Apply it where a single mistaken or compromised step could create disproportionate harm.
Measure the quality of human review
Track approval, rejection, edit and escalation rates by use case. Measure review time, downstream corrections, appeals, repeated error patterns and whether staff can explain the decision from the evidence shown.
Look for reviewers who approve nearly everything in seconds, but do not turn the metric into a quota. Investigate interface design, workload, training and incentives.
NIST's AI RMF Core includes policies that define roles and responsibilities for human-AI configurations and oversight.
Maintain override, appeal and incident paths
A person affected by a consequential outcome needs a timely route to correction. Internally, operators need a kill switch, queue pause, credential revocation and clear incident ownership.
The EU AI Act contains human-oversight requirements for high-risk systems, including the ability to understand capabilities and limitations, detect anomalies and override or stop the system where applicable. Legal scope depends on the system and jurisdiction.
Test these paths before launch. An emergency button that no one is authorised or trained to use is decoration.
How I use approval in the publishing pipeline
In the multilingual publishing workflow, AI prepares the draft, metadata and image, but it does not decide that the article is ready merely because generation finished. Mechanical checks run first: all language pages, metadata, canonical links, hreflang, image, sitemap, internal links and clean HTML.
Publication happens only after the owner explicitly authorises the article or the defined batch. A failed check stops that article, while the API update remains a separate controlled action.
The human decision is therefore placed before the public consequence and supported by verifiable evidence, not a model's self-rating.
Questions and answers
Does every AI output need human approval?
No. Match oversight to consequence, uncertainty, sensitivity and reversibility. Low-risk drafts can use sampling; consequential external actions need stronger gates.
Is a confidence score enough for review?
No. Show source evidence, rules, missing data, conflicts and the exact proposed consequence. Confidence can support, not replace, judgement.
Can the same person approve and execute?
Sometimes, for moderate reversible tasks. High-impact or irreversible actions benefit from separated roles, credentials or dual control.
What makes review meaningful?
The reviewer is competent, has enough time and evidence, can choose a different outcome and acts before the consequence.
How should a small company start?
List consequential AI-assisted decisions, define a pending state, name the reviewer, show evidence, provide reject and escalate options, and test the override.