The first time I watched a machine produce a nearly publishable draft, I came close to approving it. The prose was polished and the structure was logical. But when I checked three claims against the source material, I spent more time rebuilding the underlying case than the draft had saved. The speed was real; the cost had simply moved elsewhere.
What stayed with me was not the draft itself, but the frustration surrounding it. The tool had performed as promised. The review process had not.
For context, I founded Main & Machine, an AI implementation firm serving small and midsize businesses, and I also evaluate technology from the buyer’s perspective when it enters an 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. From each vantage point, I see the same mistake repeatedly: business owners invest in generation while giving too little thought to verification.
Many companies treat “human in the loop” as little more than a box to check on a vendor presentation. In practice, that approach breaks down. The economic benefit of machine-generated work depends on how quickly and reliably a qualified person can verify it within the actual decision-making process. U.S. Census Bureau data from late 2025 through early 2026 showed overall business AI adoption at roughly 17% to 20%, with additional companies preparing to follow. The next phase will not be held back because the models suddenly become less capable. It will slow where validation remains expensive, difficult to see through or risky for employees to challenge.
I once assumed that a second click amounted to human judgment. It does not. If the reviewer never encountered the underlying facts independently, a second click may still represent only one judgment.
Calculate the review process before automating the draft
Before automating any deliverable, identify who will review it, how much time that review should require and what happens when the work fails inspection. Automation can lower the cost of a first draft, but uncertain output creates a new obligation to check it. The benefits hold when verification is quick and mistakes are easy to reverse. They disappear when review requires scarce expertise, evidence reconstruction or the assumption of significant liability.
The rule is simple: Output that is cheap to create can still be expensive to use if verification costs too much.
Test that assumption this week. Choose one workflow you hope to automate and estimate both the generation time and the verification time for a difficult case. Do not rely on a polished demonstration. A restaurant owner may use software to prepare a weekly schedule, but a manager still needs to understand labor requirements and what is happening on the floor. A clinic owner may use AI to summarize intake notes, while a clinician must still review the original chart. If your team cannot complete the hard-case review within the normal workday, the process is not ready for automation.
Give reviewers an independent route to the facts
A reviewer cannot properly assess an automated recommendation when the system provides the only narrative of the case. If the evidence, framing and conclusion all arrive through the same interface, challenging the recommendation requires reconstructing facts the reviewer may never have seen directly. Under time pressure, approving the result can become the most practical option.
The principle is straightforward: Meaningful review requires an independent path to the underlying facts.
Create that path deliberately. For high-stakes work, keep a sample of raw inputs accessible alongside the machine-generated summary. Before showing the recommendation, ask the reviewer to identify the decisive evidence and describe the uncertainties. Then compare that assessment with the system’s reasoning. A franchisee can apply this approach to customer-complaint drafts. A contractor can use it when reviewing estimates. A professional-services firm can test it on client memorandums. Agreement is not enough; people who share the same framing can also share the same mistake. The important question is whether disagreement can alter the decision and remain documented.
Practice overriding the system under pressure
An override that exists only in a policy manual or as a button on a screen is not a dependable safeguard. Automation limits exposure to routine system behavior, but difficult cases arrive with less time, more uncertainty and higher consequences. Employees need practice identifying when something is wrong, making a decision and acting without the automated recommendation.
The rule here is simple: Practice saying no until the team can do it effectively.
Run a brief exercise this month. Select three recent outputs the team nearly approved. Remove the system’s confidence language and ask a designated owner to reject one output in writing, explain why and propose the next step. Then observe whether the rejection is accepted or quietly softened because the draft appeared complete. Automation leaves behind a more difficult set of cases: exceptions that are unusual, ambiguous and consequential. If employees see only clean demonstrations, they will not retain the skills required when those exceptions surface on an ordinary Tuesday afternoon.
My recommendation for this week is simple: Choose one AI-assisted workflow, document its review time and independent evidence route on a single page, and conduct one override exercise with the person responsible for the outcome. Share the process with the team before expanding the system’s reach.
I still think about that draft that appeared ready to approve. The true product was not the draft; it was the cost of checking it. Build the review process first, and the machine has a much better chance of earning its place.
The first time I watched a machine produce a nearly publishable draft, I came close to approving it. The prose was polished and the structure was logical. But when I checked three claims against the source material, I spent more time rebuilding the underlying case than the draft had saved. The speed was real; the cost had simply moved elsewhere.
What stayed with me was not the draft itself, but the frustration surrounding it. The tool had performed as promised. The review process had not.
For context, I founded Main & Machine, an AI implementation firm serving small and midsize businesses, and I also evaluate technology from the buyer’s perspective when it enters an 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. From each vantage point, I see the same mistake repeatedly: business owners invest in generation while giving too little thought to verification.
