A human gate is a stopping point in the process: the AI works autonomously up to it, then a named person decides whether it continues. Not a gut feeling — a place in the flow.
The term comes from process design and is the practical core of what regulation calls human oversight. It answers the question that has to be asked first of any AI agent: where does the autonomy end?
The important part is the inverted reflex. A gate is not a brake you fit because you distrust the AI. It is the condition under which you can hand anything over at all.
Where gates belong #
The most useful line runs along outward effect. Anything that leaves the company or is hard to undo belongs behind a gate:
- Sending — emails, messages, proposals
- Publishing — website, social, press
- Paying and ordering — any financial commitment
- Deleting and overwriting — anything that destroys data
- Decisions about people — applications, terminations, credit
Researching, drafting, summarising, preparing, proposing — all of that can run ahead of the gate without approval. That separation is what distinguishes a useful agent from a dangerous one.
How many gates make sense #
Too few gates are risky. Too many make automation pointless — if a person confirms every intermediate step, they have done the work themselves, only slower.
Place the gate as late as possible and as early as necessary: immediately before the first irreversible action. Everything before it runs through. One well-placed gate beats five badly placed ones.
A gate also needs three things or it is not one: a named owner, enough context to decide, and a real option to refuse. An approval button everyone clicks reflexively is theatre.
What this looks like for me #
The key figure on my assistant Ada's card is: Unilateral actions: 0. She researches, prepares, drafts and proposes meeting slots — but nothing leaves the house without my approval.
That costs me a few minutes a day and spares me the kind of mistake you cannot take back. For me it is not a compromise but the precondition for giving an agent access to my calendar and inbox in the first place.
Common questions #
Is a human gate legally required?
For high-risk systems, Art. 14 of the AI Act requires human oversight — the human gate is the practical implementation. For everything else it is not mandatory, but it is good practice as soon as a process has outward effect.
Where is the best place for the gate?
Immediately before the first irreversible action. Everything before it — researching, drafting, preparing — can run without approval. That preserves the time saved while still containing the risk.
What makes a gate ineffective?
Missing context and missing accountability. Anyone asked to approve without seeing what is at stake will click through. A gate without a named owner is a formality, not a control.
Doesn't this defeat the automation?
Only when badly placed. A single gate in the right place lets ninety per cent of the work run automatically. Five gates in the wrong places turn automation into paperwork.
Sources #
- Regulation (EU) 2024/1689 (AI Act) — Art. 14 human oversight for high-risk systems; Art. 50(4) subpara. 2 on editorial responsibility
- Anthropic — Building effective agents (sections on stop conditions and human control)
- Own practice: approval requirement for all outward actions in the AMIA agent setup
Where would your gates need to be?
Ask me — or ask Ada. She works behind a gate herself: she may prepare, but sends nothing without my approval. Free, no form.