There’s a tension at the heart of AI automation. The value comes from letting software do work without people. The risk comes from exactly the same thing. Swing too far toward caution and every action waits for approval, people rubber-stamp a flood of requests and the automation saves nothing. Swing too far the other way and an agent emails the wrong customer, issues a refund it shouldn’t or updates a record incorrectly — and trust collapses.
Human-in-the-loop (HITL) design is how you resolve this tension. Done well, it lets you start safely, measure performance and expand autonomy as it is earned.
Principle 1: Tier actions by risk
Not all actions deserve the same scrutiny. Classify every action an automation can take:
- Tier 0 — Read and analyse. Looking things up, summarising, classifying. No side effects. Fully automatic.
- Tier 1 — Reversible internal changes. Adding tags, creating drafts, updating internal notes, routing tickets. Automatic, with logging and easy undo.
- Tier 2 — Consequential or customer-visible. Sending emails to customers, changing orders, updating financial records within limits. Approval required until performance is proven, then automatic within thresholds.
- Tier 3 — High-stakes or irreversible. Payments above limits, deletions, legal commitments, anything affecting many records at once. Always require human approval.
This tiering should be explicit and documented, and it should be enforced in code — not only described in a prompt.
Principle 2: Use thresholds, not blanket rules
“All refunds need approval” is safe but slow. “Refunds up to $50 for customers with no previous refunds this year are automatic; others need approval” captures most volume automatically while keeping humans on the riskier cases. Good thresholds combine:
- Amount or impact (money, number of records, audience size).
- Confidence (the model’s or a classifier’s certainty, or agreement between methods).
- Context (customer value, account status, regulatory sensitivity).
- Novelty (cases unlike anything seen before go to humans).
Keep thresholds in configuration that operations leads can adjust without a code deploy.
Principle 3: Make approval fast
If approving takes as long as doing the task yourself, the automation adds no value. An approval request should contain everything needed to decide in seconds:
- What the agent wants to do, in one sentence.
- Why — the evidence it used, with links to sources.
- What it will look like — the actual email, the actual record change.
- Risk flags — anything unusual about this case.
- One-click actions: approve, edit and approve, reject with a reason, or take over.
Deliver approvals where people already work — Slack, Teams, email, the helpdesk — rather than in yet another dashboard. Batch low-risk approvals so a reviewer can clear twenty in a minute.
Principle 4: Capture the reason for every rejection
Each rejection or edit is a gift: it tells you exactly where the automation falls short. Capture a short reason (structured where possible — “wrong customer”, “policy misapplied”, “tone”) and feed it into your evaluation set and improvement backlog. Over time, rejection reasons show which categories are ready for more autonomy and which need work.
Principle 5: Start in shadow mode
Before an automation acts at all, run it in shadow mode: it processes real inputs and records what it would do, while people continue as normal. Then compare. Shadow mode answers critical questions without any risk:
- How often does the agent agree with experienced staff?
- Where does it disagree, and who was right?
- What does it cost per item?
- What inputs does it encounter that nobody anticipated?
Two to four weeks of shadow data is usually enough to decide what can go live and at what tier.
Principle 6: Graduate autonomy with evidence
Autonomy should be earned per category, based on measured performance. A typical progression:
- Shadow — propose only, compare with humans.
- Assisted — propose with one-click approval for every action.
- Supervised — act automatically within thresholds; sample-based review of completed actions.
- Autonomous — act automatically; monitor metrics and exceptions.
Define the criteria for moving between stages in advance — for example, “98% approval-without-edit over 500 cases in this category” — so the decision is objective rather than based on enthusiasm or anxiety. Be equally willing to move a category back when metrics drop.
Principle 7: Design the escalation path
When an automation can’t or shouldn’t handle something, the hand-off matters. A good escalation:
- Goes to the right person or queue, not a generic inbox.
- Includes the full context and the agent’s partial work, so nobody starts from scratch.
- States clearly why it was escalated.
- Has an SLA, so escalations don’t become a new backlog.
For customer-facing automation, the customer should experience the escalation as seamless: no repeating information, and a clear expectation of when a person will respond.
Principle 8: Watch for automation complacency
When an automation is right 99% of the time, reviewers naturally start approving without reading. That’s when the 1% slips through. Counter-measures include:
- Keeping review volumes manageable so attention stays high.
- Highlighting what’s unusual about each case.
- Occasionally inserting known test cases to check that review is genuine.
- Moving truly reliable categories to sample-based review instead of pretending to review everything.
Principle 9: Keep an off switch
Every automation needs a simple, well-known way to pause it — per category or entirely — that doesn’t require an engineer. Incidents happen: an upstream system sends bad data, a model update changes behaviour, a policy changes overnight. The ability to pause quickly limits damage and makes people more comfortable granting autonomy in the first place.
Principle 10: Log for accountability
Record who or what made each decision: the agent’s proposal, the reviewer’s approval or edit, the final action and the timestamp. This audit trail supports compliance, helps resolve disputes and is invaluable when investigating what went wrong. Regulators increasingly expect meaningful human oversight of AI systems; good logs are how you demonstrate it.
Measuring whether oversight is working
Oversight has costs and benefits, and both should be measured. Track, per category:
- Approval rate without edits — the clearest signal that a category is ready for more autonomy.
- Edit distance — how much reviewers change drafts. Small, cosmetic edits point to tone issues; large ones point to misunderstanding.
- Time to decision — how long items wait for approval. If this grows, the queue is too large or the requests lack context.
- Post-approval incidents — errors that got through review. A non-zero number here means reviews are not as careful as they look.
- Escalation accuracy — whether escalated items genuinely needed a person.
Review these monthly with the process owner. They turn conversations about trust from opinion into evidence, and they make it obvious when to loosen thresholds — or tighten them.
A note on regulation
Several regulatory frameworks now expect meaningful human oversight of AI systems, particularly where decisions affect people’s rights, finances, employment or access to services. “Meaningful” is the key word: a person who approves hundreds of items an hour without the information to judge them is not providing oversight in any real sense. The design principles above — risk tiers, informative approval requests, manageable volumes, logging and an off switch — are good engineering, and they also produce the evidence that oversight is real. If your use case touches regulated decisions, involve your legal or compliance team early so thresholds and logging match what they need to demonstrate.
Bringing it together
Effective human-in-the-loop design isn’t a brake on automation; it’s the mechanism that makes automation trustworthy enough to expand. Tier actions by risk, use thresholds, make approvals fast and informative, learn from every rejection, start in shadow mode and graduate autonomy with evidence.
These principles are built into every agentic automation and business operations agent we deliver.