Reviewer fatigue
Approval rates drift towards 100% as queue volume grows. Reviewers click through diffs they have not read. Your HITL control exists on the architecture diagram and nowhere else: compliance theatre with a named owner.
Your engineers design the agent side, we design the human side of decisions.
Decides when an agent must stop and request human approval.
Routes each hold to the one reviewer qualified to own it.
Forces that reviewer to reason, so approvals stand up in an audit.
Call Checkpoint Copilot at a decision point inside your own agent code, pass five observable dimensions, and get back auto_execute, log_and_proceed, or hold_for_human. Placement stops being a judgement call somebody makes once and forgets.
Hard to reverse and high consequence are driving this toward human review.
curl -X POST https://jrfinntscssvhalrnxkp.supabase.co/functions/v1/checkpoint-score \
-H "Content-Type: application/json" \
-H "x-api-key: bhv_your_api_key" \
-d '{
"decision": "Adjust user credit limit",
"reversibility": 4,
"consequenceMagnitude": 5,
"confidence": 2,
"precedent": 3,
"regulatoryExposure": 4
}'
# {
# "recommendation": "hold_for_human",
# "rationale": "..."
# }The operational problem
Approval rates drift towards 100% as queue volume grows. Reviewers click through diffs they have not read. Your HITL control exists on the architecture diagram and nowhere else: compliance theatre with a named owner.
Legal, Security or Eng leadership cannot bound the blast radius, so write access stays revoked. The agent ships in read-only mode, the business case stalls and the pilot never graduates.
One unpredictable failure and operators revert to manual work. Adoption decays, the agent sits idle and the organisation concludes the technology failed, when the handoff design did.
Deadline
From 9 December 2026, EU product liability law treats software, including AI, as a product. See what's changing →
Engagement model
The entry point. Fixed scope, fixed timeline.
Founding audits
First audit 50% off
Priced by company size.
3 founding spots. 3 remaining.
Two weeks, fixed scope. You leave with an oversight risk map of every agent-to-human handoff.
Claim a founding spotDeliverable: Risk & Intervention Architecture Map. Every decision point ranked by consequence, with the recommended gate, reviewer and fallback.
next →Sprint-based co-design inside your codebase.
Delivered as PRs in your repo, reviewed by your engineers, owned by your team.
next →The enforcement layer for the architecture in production.
POST /v1/checkpoint
{ "agent": "ledger-agent",
"action": "journal.post" }
→ { "route": "hold_for_human",
"reviewer": "financial_controller",
"reason_id": "chk_8f21" }end state20-minute technical scoping call
Why now
In the UK first, with the US and EU close behind.
UK · Data law
Since February 2026, UK data law lets organisations automate more decisions, as long as people can reach meaningful human intervention. Review only counts if the reviewer can genuinely change the outcome.
UK · Financial services
Senior managers stay accountable for what their firm's AI decides, and Parliament has asked the FCA to set out what good assurance looks like.
UK and US · Insurance
Insurers are writing AI out of standard business policies, while the London market builds dedicated AI cover. Getting covered increasingly depends on showing control.
US · State law
States such as Texas now have AI laws in force, and US carriers have filed thousands of AI exclusions on everyday business policies.
EU · Liability and regulation
From 9 December 2026, EU product liability law treats software, including AI, as a product. The AI Act requires human oversight to be effective.