Human in control: draw the approval map before the workflow

Ask an automation team whether a human is in the loop and the answer is always yes. Ask who can inspect the evidence behind tomorrow morning's outbound customer email, stop it, and put their name on the decision, and the room gets quiet. That silence is the gap between presence and control, and it is no longer a philosophical point. The EU AI Act's human-oversight article, Article 14, applies to high-risk systems from August 2026, and it goes well past asking whether a person is present: it requires that oversight persons can properly understand the system, watch for over-reliance, interpret its output, override or reverse it, and interrupt it “through a ‘stop’ button or a similar procedure.”

I build automation for a living and I want more of it in the world, which is exactly why I start every engagement with the map this article describes. Automation inside explicit boundaries moves faster, because nobody has to fear it.

What control actually requires

Presence fails quietly. NIST's generative AI profile names the mechanism: automation bias, where humans “may over-rely” on generative AI systems or unjustifiably perceive their output as higher quality. A reviewer who cannot see the evidence, lacks time to read it, or has no authority to halt the action has become a witness with a title, and their approval is a formality the system performs on itself.

Control is three testable capabilities. The owner can inspect: the evidence behind the action is assembled for them, readable in minutes. The owner can stop: halting or reversing the action is a button they actually have. And the owner decides: the approval carries their name and their reason, in a log. Every one of those is testable before launch, which is what separates a control from a diagram.

The regulators already drew the floor

The design pattern is now written into law and policy for the highest-consequence classes, and it is worth reading as a floor plan even where it does not bind you. The EU's Article 14 requirements are capability requirements, listed above. On the US federal side, OMB's Memorandum M-25-21 (April 2025, replacing the prior administration's memo) sets minimum practices for high-impact federal AI. The list includes human oversight, intervention, and accountability, plus timely human review and appeals for affected individuals. The impact assessment's risk acceptance must be “supported by a signature from the individual accepting the risk.” A named person signs. Non-compliant high-impact use must be safely discontinued. The OECD principles, updated in 2024 and adhered to by 47 countries and jurisdictions, carry accountability as a standalone principle.

Scope honesty: those instruments bind high-risk systems in the EU and high-impact use in US federal agencies, and most private automation sits outside both today. Read them anyway. When public reviewers this demanding converge on named owners, inspectable evidence, override, and stop, treat the convergence as a floor plan for trust.

The approval map

Here is the map I draw before any workflow gets built. Classify every action the system can take as reversible or consequential. Reversible actions (drafts, internal analysis, staging changes) run free; review after the fact is cheap and speed is the point. Consequential actions get a gate, and the classes are stable across almost every organization I have worked with:

For each gated class, the map records four things: the named decision owner (a person, never a team name), the evidence packet that owner sees (what the system wants to do, why, based on what, with what confidence), the fallback when the owner is absent, and the review date for the class itself, meaning the date the class definition and its threshold get re-examined. If a class has no nameable owner, that automation is not ready to run.

Walked through once, hypothetically. A support agent wants to issue a $4,000 refund, above the $500 auto-clear threshold, so the spending gate fires. The owner (the support director, by name) receives a packet: the refund amount, the customer thread that justifies it, the policy clause the agent matched, and the agent's confidence. The director approves, rejects, sends it back for revision, or lets it expire, and the log records the name, the state, and the reason. Thirty seconds of reading, one decision, a permanent record. That is the whole machine.

Four decision states, logged

Every gated decision lands in exactly one state: approve, reject, revise (send it back with instructions), or expire (nobody decided in time, and the action does not happen). The log records the state, the author, and the reason, which turns governance from a meeting into a record. Expiry deserves emphasis because it encodes the fail-closed default: an unanswered approval is a no. And approvals themselves age: a standing approval granted in March is a standing risk by September, so the map gives every delegation and every blanket approval an expiry date that forces a re-decision.

Test the map like software

A map nobody has exercised is a diagram. Before launch, run three drills. An escalation test: the agent hits a boundary; does the right owner get the packet, promptly? An absence test: the owner is out; does the fallback engage or does work silently pile up? And a bypass probe: try to route a consequential action around its gate; the attempt should fail and be logged. This is the same discipline as the forbidden-action tests in agent evaluation, applied to the human side of the boundary. Then put the map itself on a review cadence, because scope creep is how a bounded agent becomes an unbounded one without any single decision being wrong.

Keep the map from becoming bureaucracy

Excessive approval destroys the value automation creates, and I treat approval sprawl as a defect in the map, never as prudence. The guards: reversible actions never queue. Thresholds scale with consequence, so routine actions clear automatically. Delegation is time-boxed and expires. Evidence packets stay short enough to read in minutes, because an owner who rubber-stamps under volume is automation bias wearing a lanyard. And some fail-safes belong to the machine: US federal guidance itself notes that an automated safety action can be the right first responder, with the oversight and accountability process determined separately. The map is the explicit statement of where judgment lives, and everything outside it runs free.

How I run it

My own method runs this way because I built it that way: staged pipelines with gates, human-readable review files at each gate, and an editorial loop where the system drafts and a person decides what ships. I spent a decade as a state web accessibility officer inside government-grade process, and the lesson that survived is the one in this article: explicit boundaries are what let ambitious work move fast without breaking public trust. The HALO method is that lesson applied to AI orchestration, and I state its maturity honestly: a documented method with working components, improving with every engagement.

Draw the map before the workflow. It is one working session, it fits on one page, and it is the difference between automation your organization controls and automation it merely contains.

Questions worth asking next

Sources

1. Regulation (EU) 2024/1689 (EU AI Act), Article 14, Human Oversight; text via the AI Act Explorer. https://artificialintelligenceact.eu/article/14/
2. OMB, Memorandum M-25-21, "Accelerating Federal Use of AI through Innovation, Governance, and Public Trust," April 3, 2025. https://www.whitehouse.gov/wp-content/uploads/2025/02/M-25-21-Accelerating-Federal-Use-of-AI-through-Innovation-Governance-and-Public-Trust.pdf
3. OECD, AI Principles (adopted 2019, updated May 2024). https://oecd.ai/en/ai-principles
4. NIST, "Generative Artificial Intelligence Profile" (NIST AI 600-1), July 2024. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf