AI saves money only when review stays cheap
A polished machine draft can hide the real expense: checking whether it’s right. The first time I nearly approved one, three claims sent me back to the source material. Reconstructing the evidence took longer than the draft had saved. The tool had worked; the review process hadn’t. I’d seen a machine draft something that looked finished, and mistaken clean writing for reliable work.
I founded Main & Machine, an AI implementation firm serving small and mid-size businesses, and I also buy technology for an operating company. I teach management and entrepreneurship at Arizona State University’s W.P. Carey School of Business. I have a financial relationship with Main & Machine. From those roles, I keep seeing owners focus on what AI produces while giving too little attention to who can verify it.
That gap matters as adoption grows. U.S. Census Bureau data put business AI use at 17% to 20% from late 2025 through early 2026, with more firms planning to adopt it. The next constraint may be the work around the models: checking output can stay slow, opaque or costly even when generation gets cheaper. A second person clicking “approve” doesn’t guarantee independent judgment if that person sees only the same machine-framed account.
Count the review before automating
Before automating a deliverable, decide who will inspect it, how much time that inspection should take and what happens if it fails. Generation lowers the cost of a first draft; uncertain results create a bill for review. If verification is quick and mistakes are easy to reverse, the savings can hold. If review requires scarce expertise, rebuilding the evidence or accepting liability, the cost has simply moved downstream.
Try pricing one workflow this week. Estimate generation and review time on a difficult case, not a polished demonstration. A restaurant owner may use AI to draft a schedule, but a manager still needs to check labor rules and conditions on the floor. A clinic may summarize intake notes, but a clinician needs access to the original chart. If that review won’t fit into the workday, the process isn’t ready.
Let reviewers reach the evidence themselves
A reviewer can’t meaningfully check an automated recommendation if the system supplies the only account of the case. When the evidence, framing and conclusion come through one interface, challenging the answer means reconstructing facts the reviewer may never have seen. Under time pressure, approval can become the easiest option.
Give reviewers an independent route to the facts. For consequential work, keep a sample of raw inputs available alongside the machine’s summary. Ask the reviewer to identify the decisive evidence and uncertainty before showing the recommendation, then compare the reasoning. A franchisee could use this for customer-complaint drafts; a contractor, for estimate language; a professional-services firm, for client memos. Agreement is weak evidence when both people share the same framing. The useful test is whether disagreement can change the decision and be recorded.
Practice rejecting the answer
An override policy or button won’t help much if nobody has practiced using it. The cases that remain after routine work is automated may be rare, ambiguous and consequential. People need to recognize when an answer feels wrong and know what to do next.
Run a short drill this month. Choose three recent outputs the team nearly sent, hide the system’s confidence language, and ask a named owner to reject one in writing and propose another step. Then see whether the rejection stands or gets softened because the draft looked complete. If the team only encounters clean demonstrations, it may not be ready for the difficult exceptions that show up during an ordinary workday.
Start with one AI-assisted workflow. Put the expected review time and the independent evidence route on one page, then run an override drill with the person accountable for the outcome. The key economic question isn’t how fast the machine writes. It’s whether someone can check the work reliably at a reasonable cost—and say no when the evidence calls for it.
Christopher Myers, CEO of B:Side Capital and B:Side Fund and professor at W.P. Carey School of Business, wrote this opinion piece; Kara McIntyre edited it. It was published Nov. 1, 2026. Myers’s author page includes a X account, LinkedIn profile, B:Side Capital website and book listing.
