AI Can Draft in Seconds. Can You Afford to Check Its Work?

The first time I watched a machine draft something that looked finished, I almost signed off. The language was clean. The structure made sense. Then I tried to check three claims against the source material and burned more time reconstructing the case than the draft had saved. Speed had moved. The cost had not gone away. It had just changed rooms.

I remember the annoyance more than the draft. The tool had done its job. The review path had not.

Here’s my situation, so you know where I’m coming from. I founded Main & Machine, an AI implementation firm for small and mid-size businesses, and I also sit in the buyer seat when technology lands in a real operating company. (Disclosure for the editor: I have a financial relationship with Main & Machine.) I teach management and entrepreneurship at Arizona State University’s W.P. Carey School of Business. Across those chairs I keep seeing the same pattern: owners celebrate generation and underbuild verification.

Most shops treat “human in the loop” as a checkbox on a vendor slide. That move fails in practice. The economic value of generated output depends on how cheaply and reliably a competent person can verify it in the real decision. U.S. Census Bureau data from late 2025 through early 2026 put overall business AI use in the 17% to 20% range, with more firms still planning to join them. The next wave will not stall because models got worse. It will stall where checking stays slow, opaque or socially expensive.

I used to treat a second click as proof of judgment. Wrong. A second click can still leave you with one judgment if the reviewer never saw the underlying case on independent terms.

Price the review path before the draft

Before you automate a deliverable, write down who will check it, how long that check should take and what happens when the check fails. Generation reduces first-draft cost. Uncertain output creates a checking obligation. When verification is fast and error is reversible, the gain survives. When checking requires scarce expertise, reconstructing evidence or absorbing liability, the apparent speed just moves cost downstream.

The principle fits in one line: Cheap output with expensive verification is still expensive work.

Try the same pricing this week. Pick one workflow you want the machine to touch. Estimate minutes to generate and minutes to verify on a hard case. Skip the polished demo. A restaurant owner might let a tool draft the weekly schedule and still need a manager who can read labor rules and walk the floor. A clinic owner might let a tool summarize intake notes and still need a clinician who can open the raw chart. If the hard-case check cannot fit the day, the automation is not ready.

Give reviewers a route the machine did not frame

A reviewer cannot meaningfully check an automated recommendation when the system is also the only account of the underlying case. When evidence, framing and conclusion arrive through one interface, disagreement requires reconstructing information the reviewer never independently saw. Time pressure then makes rubber-stamping rational.

The principle: Review needs an independent path to the facts.

Build that path on purpose. For high-stakes work, keep a sampled set of raw inputs available beside the machine’s summary. Ask the reviewer to name the decisive evidence and the uncertainty before the recommendation appears, then compare reasons. A franchisee can do this on customer-complaint drafts. A contractor can do it on estimate language. A professional-services owner can do it on client memos. Agreement alone proves little. Shared framing can produce shared error. What matters is whether disagreement can change the action and survive in the record.

Rehearse the override under pressure

An override that exists only as a policy or a button is not a reliable safeguard. Automation reduces exposure to normal system dynamics, while the ugly cases compress time and raise uncertainty. People need practice recognizing, deciding and acting without the automated system when the answer feels wrong.

The principle: Practice the no until it is usable.

Run a short drill this month. Take three recent outputs the team almost shipped. Hide the machine’s confidence language. Ask a named owner to reject one of them with a written reason and an alternate next step. Then watch whether the rejection sticks or gets smoothed over because the draft sounded finished. Automation leaves a harder residue: The exceptions that remain are ambiguous, rare and consequential. If your team only ever sees clean demos, they will not keep the skill they need when the residue shows up on a Tuesday afternoon.

Here’s where I’d start this week: Pick one AI-assisted workflow, write the review minutes and the independent evidence route on a single page, and run one override drill with the person who actually owns the outcome. Say it to the team before you expand the tool’s reach.

The draft that looked finished still sits in my mind. The checking cost was the real product. Build that first, and the machine earns its keep.

The first time I watched a machine draft something that looked finished, I almost signed off. The language was clean. The structure made sense. Then I tried to check three claims against the source material and burned more time reconstructing the case than the draft had saved. Speed had moved. The cost had not gone away. It had just changed rooms.

I remember the annoyance more than the draft. The tool had done its job. The review path had not.

Here’s my situation, so you know where I’m coming from. I founded Main & Machine, an AI implementation firm for small and mid-size businesses, and I also sit in the buyer seat when technology lands in a real operating company. (Disclosure for the editor: I have a financial relationship with Main & Machine.) I teach management and entrepreneurship at Arizona State University’s W.P. Carey School of Business. Across those chairs I keep seeing the same pattern: owners celebrate generation and underbuild verification.