Human in the loop
A design pattern that makes sure a named person makes every decision that matters, with the AI's evidence in front of them and a record of what they decided.
What it is
Human in the loop isn't a model. It's the part of the design that decides which outcomes a person must approve, sends those cases to the right person with the evidence they need, and records what they decided and why. Since 5 February 2026, UK GDPR has required safeguards for significant automated decisions, including the right to human intervention. This pattern is how those safeguards are built.
It works through a few simple parts. The AI's output carries a confidence and its reasons. Rules decide what happens next: straight through, sampled review, or mandatory review, for example anything adverse, high-value or involving a vulnerable customer. The workflow pauses, a task appears in the reviewer's queue, and the workflow resumes only when a person accepts, changes or rejects the recommendation. Every step goes into an audit log that can't be edited.
In an enterprise, the review should happen where people already work: the claims workbench, the case system or the EPR worklist, not a new screen. The design also has to stop the human step becoming a rubber stamp. Reviewers need time, the evidence and the freedom to disagree, and their overrides need to feed back into how the AI is tuned.
Use it when, and think twice when
Use it when
- A decision affects someone's money, health, rights or access to a service
- The AI is new and you need evidence before trusting it with more
- Rules or the regulator expect a person to be able to step in
- Reviewers' corrections can be used to improve the system
Think twice when
- Every case goes to review, so the AI saves nothing
- Reviewers have seconds per case and no real way to disagree
- The review happens in a separate screen nobody keeps open
- Overrides aren't recorded, so nothing is learnt from them
Reference architecture
01Channels
- AI recommendation with reasons
- Rule or threshold breach
- Customer asks for a review
02Intake
- API GatewayAPI ManagementKong, or your existing API gatewayReview API
- Amazon SQSService BusRabbitMQ or Apache KafkaReview queue
03Intelligence
- AWS LambdaAzure FunctionsRules in code on KubernetesRouting: straight through, sample or review
- Amazon BedrockMicrosoft FoundryOpen-weight models served with vLLMEvidence pack and summary
04Orchestration
- AWS Step Functions (wait for task token)Durable Functions (external events)Temporal (signals)Pause until a person decides
- Amazon EventBridge, Amazon SNSEvent Grid, Logic AppsApache Kafka; Temporal timersReminders and escalation
05Systems of record
- Existing workbench or case tool
- Amazon S3 with Object LockBlob Storage with immutable storageCeph with S3 Object LockTamper-evident audit log
06People
- Person
Reviewer: accept, change or reject - Person
Senior reviewer for escalations - Person
Quality sample review
- Person
- Cloud or platform service (AWSAzuresovereign / on-prem)
- AI agent
- A person decides
- Your existing systems
- Outside your organisation
Components, side by side
| Layer | AWS | Azure | Sovereign / on-prem |
|---|---|---|---|
| Routing rulesAgreed with compliance, tested like any other code, and changed under change control. | AWS Lambda, with rules versioned in code | Azure Functions, with rules versioned in code | Rules in code on Kubernetes |
| Pause and resumeThe workflow can wait days for a person without losing its place. | AWS Step Functions (wait for task token) | Durable Functions (wait for external event) | Temporal (signals) |
| Review queueA priority and a deadline on every task, so urgent cases aren't buried. | Amazon SQS | Service Bus | RabbitMQ or Apache Kafka |
| Review screenAmazon A2I closed to new customers on 30 July 2026. Review where people already work. | In the existing workbench, or a custom review screen | In the existing workbench, or Power Apps | In the existing workbench |
| Evidence packShow the source documents and the reasons, not just a score. | Amazon Bedrock | Microsoft Foundry | Open-weight models served with vLLM |
| Approval rightsOnly people with the right role can approve, and nobody approves their own work. | IAM Identity Center | Entra ID | Keycloak, or your existing Active Directory |
| Audit logRecord the recommendation, the evidence shown, who decided, what and why. | Amazon S3 with Object Lock, CloudTrail | Blob Storage with immutable storage, Azure Monitor | Ceph with S3 Object Lock, OpenTelemetry |
| EscalationOverdue reviews escalate automatically to a named person. | Amazon EventBridge, Amazon SNS | Event Grid, Logic Apps | Apache Kafka; Temporal timers |
| Quality reportingOverride rates by reviewer, team, case type and customer group. | Amazon Quick (formerly QuickSight) | Power BI | Grafana, or your existing BI tool |
On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →
Operations
Keeping it reliable
AI fails quietly. These three habits make failure visible before customers see it.
Evaluation
Measure the people too
Track how often reviewers agree with the AI, and check a sample of both. Agreement near 100% can mean people have stopped checking. Now and then, put known test cases into the queue to see whether problems are caught.
Monitoring
Watch the queue every day
Track queue length, time to decision and overdue cases. A growing queue means the thresholds are wrong or the team is too small. Report override rates by case type and customer group.
Cost
Count reviewer time
Most of the cost here is people, not cloud. Model it honestly: cases per hour, the share sent to review and the time each review takes. Lower the review share only when the evidence supports it, one case type at a time.
Security and failure risks
Rubber-stamping
Reviewers approve whatever the AI suggests because they're busy or trust it too much, so the human check exists only on paper.
How the design avoids it
Realistic review targets, evidence and reasons on screen, known test cases in the queue, and override rates watched by a team leader.
Bypassing the review
An integration or an over-privileged service account writes the AI's output straight into the system of record.
How the design avoids it
Only the workflow can commit, and only after approval, enforced by platform permissions. The AI's own identity can't write to the record.
Gaps in the audit trail
When a decision is challenged, nobody can show what the AI recommended, what the person saw or why they decided.
How the design avoids it
An append-only log of the recommendation, the evidence shown, the decision and the reason, kept for the agreed retention period.
Queues that grow until people give up
Thresholds send too much to review, the backlog grows, and the team starts approving in bulk.
How the design avoids it
Daily queue reporting, deadlines with escalation, and thresholds tuned with evidence, not guesswork.
Playbooks that use this pattern
All 79 playbooks using Human in the loop →KYC and customer onboarding
Every application read, checked and summarised before an analyst opens it. Analysts spend their time on judgement, not copying, and a person approves or declines every customer.
- Document AI
- Agents with tools
- Human in the loop
- Event-driven integration
Payment fraud and APP scam detection
Risky payments scored in milliseconds, only the highest-risk band held automatically under rules a person approved, every other alert reviewed by an analyst before any hold, and people making every decision about a customer's payment or claim.
- Predictive ML
- Agents with tools
- Human in the loop
- Event-driven integration
Complaints handling
Every complaint caught, whatever channel it came through, answered fairly within the DISP deadlines, and traced back to its root cause. A handler writes and signs every final response.
- Retrieval (RAG)
- Agents with tools
- Speech and voice
- Human in the loop
Reconciliation and exception handling
Millions of transactions matched every day, the breaks investigated before anyone opens a spreadsheet, and a person approving every posting, with the audit evidence built as the work is done.
- Predictive ML
- Agents with tools
- Human in the loop
- Event-driven integration
Credit decisioning support
Every referred application arrives with the evidence gathered, affordability worked out and the score explained in plain English. A person makes every referred decline, clear-cut cases stay with the existing decision engine and its safeguards, and fairness is checked every month.
- Predictive ML
- Document AI
- Agents with tools
- Human in the loop
Spotting vulnerable customers
Possible signs of vulnerability noticed the first time a customer mentions them, a trained person offered the right support, and outcomes checked so no group of customers is left worse off.
- Retrieval (RAG)
- Predictive ML
- Speech and voice
- Human in the loop
Claims intake and triage
Every claim read, sorted and routed within minutes. Simple claims ready for one-click approval, complex ones flagged early, and people making every decision that matters.
- Document AI
- Agents with tools
- Human in the loop
- Event-driven integration
Underwriting submission review
Every broker submission read, its risk data extracted and checked against appetite on the day it arrives, with gaps flagged and a referral note drafted. The underwriter decides, prices and speaks to the broker.
- Document AI
- Retrieval (RAG)
- Agents with tools
- Human in the loop
Damage assessment from photos
Usable photos first time, a likely repair route within hours, and reused or edited images caught before payment. An engineer or handler makes every settlement and total-loss decision.
- Computer vision
- Human in the loop
- Agents with tools
- Event-driven integration
Patterns
All eight building blocks
Document AI
Read forms, letters and PDFs into structured data.
Retrieval (RAG)
Answer questions from your own documents, with sources.
Agents with tools
AI that takes small, controlled actions across your systems.
Human in the loop
Put people on the decisions that matter, with the evidence.
Predictive ML
Score, rank and forecast from your own historical data.
Event-driven integration
Connect AI to core systems safely, through events and queues.
Computer vision
Inspect photos, video and camera feeds for defects and damage.
Speech and voice
Turn calls and conversations into text, notes and actions.
Thinking about human in the loop in your organisation?
Talk through where it fits, what it would plug into, and which decisions should stay with people.
