Human in the loop means three things. Ask which one.

A Financial Planning opinion column argues that “human in the loop,” the trust signal of choice for AI vendors selling into advisory practices, now carries three distinct meanings — only one of which works as a control.

The oldest meaning describes how a model was built: human experts — accountants, lawyers, estate planners — label documents and mark what correct looks like on a tax return line item, an estate document clause, an insurance policy term, so the model absorbs professional judgment instead of guessing from raw data. The second meaning arrives after the build, when experts review outputs, correct errors, rank answers by accuracy, and the corrections get fed back in — the process known as reinforcement learning from human feedback and the one most people picture when they say a tool gets sharper with use.

The third meaning is gaining ground as AI agents begin taking independent action — sending emails, booking meetings, moving money — and the phrase now describes the advisor approving what the tool is about to do.

Taken alone, that definition is thin: it says nothing about how the model learned to generate the recommendation now awaiting a signature. A vendor can build with no expert involvement in training at all, feed the model messy or unvetted data, and still market the product as human in the loop because an advisor clicks approve before an email goes out. Financial Planning’s illustration: a tax planning tool flags a Roth conversion and drafts the client email recommending it. The careful advisor checks current and projected brackets, confirms the conversion doesn’t push the client into a higher income-related monthly adjustment amount, verifies state tax treatment, and tests the idea against the rest of the client’s balance sheet. That work is genuine, and it is the advisor’s judgment carrying the outcome, not the vendor’s.

The approval gate is the least expensive of the three meanings to put in a pitch deck, because it requires nothing of the model builder — no labeled training set, no review cycle, no feedback pipeline, just a screen where a human confirms what the agent already decided. An advisor who accepts the phrase at face value may believe they are buying the first or second definition while actually operating the third, absorbing model risk their own errors and omissions coverage was never priced for.

Providers frequently leave the term unexplained, asking advisors to bet client outcomes and their own reputations on tools they can’t describe from the inside. The practices most exposed are fast adopters that skipped a pre-AI process audit, and ambiguity about what a vendor actually built is exactly the gap that audit exists to close.

The vetting question is narrow and answerable before any contract: which loop do you mean, who labeled the training data, and what happens to an output after a human corrects it? A vendor with a real answer will have one ready. A vendor whose only human in the loop is the advisor should say so plainly, because then the advisor knows precisely what they are underwriting — and can price the click and its consequence.