4 Questions Advisors Should Ask About ‘Human in the Loop’ AI

Rachel Schwab, director of product at FP Alpha and a former product manager for advice technology at Vanguard, urged advisors to press AI vendors for specifics when platforms claim a ‘human in the loop.’ She noted some providers use the phrase without describing how models were trained or validated.

Schwab described three distinct uses of the term. The first refers to training: domain experts such as accountants, lawyers or estate planners label documents or outcomes so the model learns from professional judgment rather than raw data. The second refers to post‑training review, where experts correct outputs and rank responses; those corrections can be fed back into the system to improve it. The third refers to an operational approval gate, where an advisor or other human approves an AI action before it runs, such as sending client emails, booking meetings or moving money.

The difference matters for client outcomes. An AI tax‑planning tool that flags a Roth conversion and drafts an email could miss a client’s projected tax bracket, trigger an income‑related monthly adjustment amount, or misapply state tax treatment. An advisor who only approves the draft might not detect those errors if the underlying model was not trained or validated by experts.

Schwab recommends advisors ask four questions when evaluating AI tools: who trained the model and what were their credentials, what data did the trainers use, at what points were humans involved in development and refinement, and which parts of the training had no human input.

Schwab warned, “Vendors can market a model as ‘human in the loop’ even if the only human action is the advisor’s final approval.” She added that the phrase is too vague on its own because providers can imply any of the three meanings without committing to details about training, data quality or ongoing oversight.

Schwab advised firm leaders to require documentation of trainer credentials, datasets used, and the timing and scope of human involvement, along with clear notes on where automated processes operate without human checks. She said advisory firms and individual advisors may face accountability for client outcomes while lacking visibility into how a tool reached its recommendation, and that a final approval step works only if advisors can identify and correct model errors.