What “AI-assisted review” actually means — and what it doesn’t

The phrase “AI-assisted review” means different things depending on who is using it. In our vocabulary it has a narrow, specific definition, and the narrowness is intentional.

What it does: it reads a return before a partner does. It compares line items to the prior year and flags deltas that fall outside expected ranges. It identifies anomalies — things that look structurally inconsistent with how the rest of the return is put together. It surfaces what a partner would probably want to look at, in the order a partner would probably want to look at it.

What it doesn’t do: it does not make review decisions. It does not sign returns. It does not replace the judgment of the licensed human whose name goes on the document. It does not produce explanations the firm is expected to defend without the partner’s own reasoning underneath.

The technology does the scanning. The partner does the deciding. The division of labor is the point.

This boundary is not an act of modesty. It is a design decision. Partner-level review judgment is why mid-size firms exist in the first place — it’s the craft that distinguishes a partner’s signature from a bot’s. Building a system that tries to replace that judgment would be solving the wrong problem and introducing the wrong risk.

The right problem is different. Partners are a fixed resource in March. Prep capacity is not. The question is how to make partner time go further without asking partners to work more hours or review less carefully. “AI-assisted review” is our answer to that question, held to the specific scope above.

When a vendor tells you their AI reviews returns, it is worth asking which definition they mean. The difference is not marketing. It is whose name is on the return when the IRS has a question.