"Human-in-the-loop" appears on almost every AI vendor's slide deck. It usually means one of two very different things, and procurement teams rarely ask which one they are buying.
Key takeaways
- Human-in-the-loop means a person reviews and approves every material output before it takes effect. Human-as-fallback means a person is available if something goes wrong, but nothing waits for them by default.
- The two patterns cost differently, fail differently, and satisfy different governance requirements — they are not interchangeable, whatever the pitch deck implies.
- Regulated and public-sector work almost always needs the former for anything consequential. Most vendors quote pricing that assumes the latter.
- Ask a vendor one question: "what happens if no human looks at this for 48 hours?" The answer tells you which pattern you are actually buying.
Two patterns, one label
In a true human-in-the-loop system, the AI proposes and a person disposes. Nothing material happens — an invoice isn't sent, a decision isn't recorded, a document isn't published — until someone with authority has looked at it and said yes. The AI's job is to make that review fast and well-informed, not to replace the decision.
In a human-as-fallback system, the AI acts by default. A person is in the org chart somewhere, notionally accountable, and can intervene if they notice a problem — usually after the fact, usually because something else flagged it. The system does not wait for them. It is faster and cheaper to run, right up until it is wrong in a way nobody caught in time.
Both patterns get called "human-in-the-loop" in sales conversations, because the phrase sounds reassuring and nobody wants to be the vendor who admits their system runs unsupervised most of the time.
Why the distinction is a cost decision, not just a governance one
True human-in-the-loop is slower and more expensive to run because someone's time is on the critical path for every material action. That is a real, ongoing cost, and it should be priced as one. Human-as-fallback is cheaper precisely because that cost has been deferred, not removed — it shows up later, as incident response, remediation, or reputational damage, usually at a worse exchange rate than if you had paid for review upfront.
We have seen procurement teams compare two vendor quotes as if they were pricing the same service, when one has built in review time and the other has priced the unsupervised version and is quietly hoping the fallback case never triggers.
Where each pattern actually belongs
Human-as-fallback is a legitimate, sensible choice for low-consequence, high-volume, easily reversible actions — drafting an internal summary, tagging a support ticket, suggesting (not sending) a reply. Human-in-the-loop is the only defensible choice for anything that moves money, commits the organisation externally, or affects a citizen's or client's legal position. Government and regulated-industry work sits almost entirely in the second category, which is why "we use human-in-the-loop AI" needs a follow-up question, not a tick in a box.
The question that cuts through the pitch
Ask any vendor: what happens if no human looks at this for 48 hours? If the honest answer is "nothing changes, it just keeps waiting," you are buying human-in-the-loop. If the honest answer is "it keeps running and someone finds out eventually," you are buying human-as-fallback, no matter what the slide says. Both can be the right architecture. Only one of them should be sold as the other.
Outcome-priced from day one
See what this would cost at Effektiv pace.
Pick a project that finished or stalled. Show us a quote you've received or an invoice you've paid. We'll price the same scope on outcomes, not hours.