Skip to main content

Reviewing AI Work: What This Is For

By Jim Wilde, Emergent Skills · Nov 15, 2026

Producing analysis got cheap. Checking it did not.

AI tools now draft status reports, business cases, forecasts and client memos in minutes. Someone still has to decide whether each one is reliable enough to act on. That decision is where the cost has moved.

A study of four accounting firms, published in Harvard Business Review in November 2026, found that AI prepared tax returns faster than reviewers could verify them. Review time fell only modestly in some cases and did not fall at all in others. One of the authors has a financial tie to the platform studied.

The old habits of review also stop working. An experienced reviewer spots a junior colleague's mistakes by pattern: a number that looks off, a paragraph that hedges. AI-drafted work gives none of those signals. It is fluent, well organized and consistent with itself, so it reads as right. The errors sit in the gap between the document and its sources, and rereading the document does not find them.

A rule that says a person reviews everything only helps if that person knows how to check.

What organizations do about it now

Most rely on one of five approaches. Each has a use, and each leaves the same gap.

  • Trust the reviewer's experience. Experience matters, and it was built on work that people produced. It does not tell a reviewer where an AI tool tends to go wrong.
  • Add a sign-off rule. A rule makes someone accountable, which helps. It does not supply the skill. A reviewer who approves whatever arrives has still followed the rule.
  • Train people to use AI. Prompting courses teach how to get output. They say little about judging it once it arrives.
  • Have AI check AI. A second tool can catch some errors. A person still has to decide whether to act and has to answer for the result.
  • Build a large internal program. PwC is putting more than 85,000 people through a critical thinking and judgment certificate and estimates the investment at more than $100 million. Few organizations can do that, and an individual project manager cannot.

The gap is practice within reach. The first four give a professional no repeated attempts at the actual task, with feedback on what they missed. The fifth does, for the people who work at a firm that size.

What good review looks like

A professional receives a piece of AI-assisted work and, in about twenty minutes, can say whether it is reliable enough for its intended use, explain the reasoning, and name the next justified action. They got there by reviewing realistic work, making a call and seeing what they missed.

What this is

It is a practice program for people who review work and advise on business decisions: project and program managers, department managers, PMO leaders and analysts, business analysts, and consultants. An individual can take it for their own work. A team leader can buy it for a team.

The learner works through cases. Each case is a brief, a few exhibits and an AI-drafted analysis. The learner applies a five-step method (claim, weak spots, chain, sources, call), makes a decision and gets scored feedback.

What is different about it

Two calls

One is on the document: accept, fix or escalate. The other is on the business decision: proceed, proceed with conditions, test, get evidence or stop. A sound document can describe a bad proposal, and a flawed one can describe a good opportunity.

Clean cases

Some cases have nothing wrong with them. Flagging problems that are not there counts against the reviewer, because rejecting everything is as unhelpful as accepting everything.

A view before the answer

In selected cases the learner records their own view before seeing what the AI concluded.

Credit for not signing

One case gives the reviewer too little time or information. Saying so, and saying which check remains unfinished, is the right answer.

Assessment without AI help

The final cases are done unaided, so the result shows what the person can do alone.

The evidence, and where this stands

The design rests on published research. The program itself has no results yet.

  • The problem shows up among working professionals. In a field experiment run by Harvard Business School researchers with Boston Consulting Group, 758 consultants were given realistic tasks. On a task beyond the AI's ability, the consultants who used AI were 19 percentage points less likely to reach the correct answer.
  • Working conditions make it worse. A systematic review of research on over-reliance on automated advice found that workload, task complexity and time constraints increase it, and that training and accountability reduce it.
  • Practice with feedback changes decisions. In a study published in Psychological Science, 290 graduate students took an unannounced business case. Those who had already completed one training session with practice and feedback were 29% less likely to choose the weaker answer that confirmed their first belief.

None of these studies tested this program, and we have not found a published study of training people to review generative AI output at work. The program is a pilot. The first case is written and the first trials have not yet run. Those trials will measure errors caught, false alarms and review time, before and after.

Questions people raise

"We already require a human review."

That settles who is responsible. It leaves open what the reviewer does with the half hour. This gives them a method and practice at using it.

"Is this AI training?"

No. It does not teach prompting or tools. It teaches checking what the tools produce.

"Our people are experienced."

Experience helps most with the mistakes people make. AI output fails in other places, and the consultants in the Harvard and BCG experiment were working professionals.

"We have no time for training."

A case takes about half an hour. The time it is meant to save is the review and rework that follow a recommendation nobody checked.

"Does it work?"

That is unproven. The research behind the design is sound, and the pilot is set up to measure the result.

Try one case

The first case is a memo recommending that a company close one of its branches. It goes to the CFO tomorrow under your name. Read the exhibits, review the memo and decide whether you would sign it. You get a score and an explanation of what you missed.

Sources

Share