Human-in-the-loop: why the best automation keeps a person in charge
The automations that survive contact with reality share one design choice: at every point where a mistake would be expensive, embarrassing or irreversible, a person says yes.
The fastest way to lose trust in automation is one bad email sent to one important client. The fastest way to keep it is making sure that email couldn't have sent itself.
What does human-in-the-loop mean in practice?
The system does the work; a person makes the calls. An inbox automation reads, sorts, chases and drafts, but the draft sits in a queue until someone clicks send. An invoice system extracts and checks, but anything unusual routes to a person instead of posting. The human isn't doing the work anymore; they're doing the judging, which is the part they were always uniquely good at.
Which automated tasks still need a human in the loop?
- Anything outward-facing. Emails, quotes, letters: reputation rides on them.
- Anything financial. Payments, pricing, credit decisions.
- Anything irreversible. Deletions, filings, contract commitments.
- Anything the model flags as uncertain. Good systems know when they don't know, and say so instead of guessing.
Internal, reversible, low-stakes work (sorting, summarising, drafting, matching) can and should run free. Putting approvals on everything is how you build a system nobody uses. This is also the design answer to will AI replace my staff: it removes the free-running slice of a role and keeps a person on the judgement calls, rather than removing the person.
Doesn't keeping a human in the loop slow automation down?
"Doesn't the human slow it down?" Barely, and it's the wrong comparison. Approving a drafted email takes ten seconds; writing it took four minutes. The automation still removed 95% of the work. What the loop costs in seconds it repays in the one intervention a month that prevents a real mistake, and in something less measurable: your team trusts a system they supervise, and adopts it. Automation people don't trust gets quietly worked around, and dies.
Does UK and EU law expect a human in the loop?
Since February 2026, UK GDPR, as amended by the Data (Use and Access) Act 2025, allows more significant decisions to be made solely by machine, but only with safeguards, including a way for the person affected to get a human to step in and to contest the decision. The EU AI Act requires high-risk AI systems to be designed so that people can oversee them effectively. Systems designed with human decision points aren't just safer: they're already shaped like the compliance environment they'll live in. More on that in GDPR and AI.
How does M22 help with this?
Keeping a person on anything that matters is not an optional extra at M22 Consultancy, it is one of the principles every M22 build follows, alongside working software over slideware and numbers over adjectives. When M22 designs a workflow automation, an agent or a Company Brain, M22 decides where the approval points sit before a line of code is written, so the system M22 hands over has a person approving anything that matters. The AI Audit starts at £1,500, a fixed fee agreed before day one. Book a thirty-minute call to talk through where the approval points should sit in your business, or see who owns what gets built.