Human-in-the-loop: where approval belongs

Approval belongs wherever an error would be expensive, irreversible, or visible to a client — and nowhere else. That single rule resolves most automation design arguments: full autonomy for internal, reversible, checkable steps; a human checkpoint before anything leaves the building or touches money. The craft is in placing the checkpoint precisely, making it cheap to operate, and knowing when it has earned the right to come out.

Why not just automate end-to-end?

Because error rates compound and error costs are asymmetric. A pipeline of five AI steps, each right 95% of the time, is right end-to-end only around 77% of the time — and the failures are silent, confident, and fluent. Meanwhile the cost of a wrong output varies by orders of magnitude: a badly classified internal ticket costs minutes; a hallucinated figure in a client proposal can cost the client.

Full autonomy everywhere is the maximalist position currently fashionable at the loud end of the market; zero autonomy is the sceptic's. Both waste money. The productive position — consistent with everything in AI Automation for B2B: what actually works — is that autonomy is a property you assign per step, based on the price of being wrong at that step.

Where exactly does the checkpoint go?

Sort every step in the workflow by two questions: can the output reach the outside world (a client, a prospect, a regulator, a ledger), and can a mistake be cheaply undone?

• Internal and reversible — enrichment, deduplication, transcription, tagging, internal drafts: automate fully. Reviewing these is theatre; spot-check instead.
• Internal but consequential — lead scoring that decides who gets called, forecasts that steer hiring: automate, but expose the reasoning and audit monthly.
• External or irreversible — anything sent, published, invoiced, refunded, filed, or deleted: a human approves, every time, until the system has earned otherwise.

The commonest placement error is approval at the wrong altitude: a human reviewing every enriched data row (waste), while an AI-drafted proposal goes out unread because "the template was approved once" (exposure).

How do you make approval cheap enough to survive?

An expensive checkpoint gets bypassed within a month, so design it like a production step. The mechanism: when the system produces an output bound for the outside world, then it lands in a queue, not an inbox — one place, batched, reviewed once or twice a day at scheduled slots. When the reviewer opens the queue, then each item shows the output plus the evidence behind it, so approval takes seconds rather than requiring re-research. When an item is fine, then one click sends it; when it needs work, then the edit itself is captured and fed back into prompts and templates, so the same correction is not made twice. When the system's own confidence is low, or the input falls outside known patterns, then the item is flagged for closer attention rather than buried mid-queue.

Done properly, a human approves fifty outbound items in twenty minutes with genuine attention. Done badly — ad hoc notifications, no evidence attached — the same fifty items consume two hours, and by week six nobody is looking.

When can the human come out of the loop?

When the checkpoint has data behind it. Track the edit rate: if a step has run a few hundred times with approval, and the reviewer has changed materially nothing for months, the full review can step down to sampling — every tenth item, then spot-checks, with automatic re-escalation the moment edit rates rise or inputs shift. Autonomy is earned by track record, not granted by optimism. This is also the honest answer to the "agents will run everything" claims I push against in The hype curve and the boring middle where money is: perhaps, eventually — show me the edit-rate curve first.

Some steps should never graduate. Pricing, legal commitments, anything with regulatory weight — the checkpoint is permanent there, and in compliance-heavy verticals the "external" category is simply larger: what an accountancy practice or a recruitment agency may send unreviewed differs sharply from a design studio, a variation mapped in Growth Systems by Industry.

Where does this matter most for revenue?

Outbound. AI-drafted sales emails are the highest-volume external output most small firms will ever automate, and the failure modes are specific and well-documented — I catalogue them in AI-written sales emails: where they fail. A reply queue with human approval is the difference between a system that books meetings and one that quietly burns your domain reputation at scale.

Autonomy where errors are cheap. Approval where they are not. Data to move the boundary. That is the whole doctrine.

---

Next step: the Growth System Audit — £450, seven days, credited against any build — maps where your growth system leaks and what to build first.