The Four Stages of Reviewing AI Work

Your team can produce more with AI and still take longer to get the work to a client, because somebody has to check everything that is now arriving faster.

And when a bad piece of work has already reached a client, I understand why that person starts reading every line. They aren't being difficult, they are using the control they have, and the hours you thought AI would save are going into that review.

What I have watched in service firms is a progression through four stages, marvel, panic, rubric and maturity. People start by trusting too much, then checking everything, then giving AI a written standard to check against, and eventually needing less human review as the whole system improves.

The third stage is where the work changes, but getting there takes time and the company can run slower while it learns. If you don't plan for that part, a team can stop just when it is beginning to find out what the system needs.

Let me take you through the stages, because the answer isn't to tell your reviewers to be less careful.

Marvel, when the output is enough

At first people are impressed by how much the tool can produce. A report appears quickly, it reads well, and the work gets sent on without much checking because it looks like the job is done.

This is harder to see when the AI use is unofficial. If nobody knows how a piece of work was made, they may not know which checks it needs, so they use the old process and assume the person producing it already did them.

The problem isn't that every output is wrong, it is that nobody can explain what would stop a wrong one. Ask who or what checked the last deliverable and the answer is often unclear.

That is part of the Production Gap, the difference between a tool making something impressive and a process delivering work you can depend on.

You don't have to go through a client-facing failure to learn this. You can build the checking alongside the new workflow, even though the sequence I have watched often begins without it.

Panic, when everything gets read

After something goes wrong, the team starts checking. A partner reads the reports line by line, sometimes another person reads them too because nobody is quite sure who is responsible for the final approval.

The work is faster to create and no faster to send. Senior people stay late, the same documents come around again, and the calendar starts filling with review rather than delivery.

Most firms I see are in some version of this stage. That is my observation, not a measured share of all firms using AI, and I don't think the people doing it are wasting time on purpose.

If the last wrong number reached a client, reading the next report carefully is a reasonable response. The owner has to give the reviewer a better way to check the work, not ask them to ignore what happened.

Atlassian described a similar pattern in its 2026 State of Teams research, which included more than 12,000 knowledge workers and more than 170 Fortune 1000 executives. Among those executives, 89% said AI increased the speed of work, while only 6% felt confident pointing to organization-wide returns.

Those are survey responses from a much larger-company setting, not a measured result for a service firm like yours. The part I recognize is the description of output backing up at reviews, approvals and other points that still need human judgment.

Hiring another reviewer can help, and I would not tell you that extra capacity never matters. Queueing theory explains why delays can rise sharply when the person doing the review is nearly fully booked, and adding capacity can bring that load down.

But a queueing formula doesn't prove another hire will only help a little, or buy you a fixed number of months. That depends on how work arrives, how long review takes and whether the new person can actually take over the same decisions.

The question is whether you are building review into the workflow or adding people who must keep rereading everything the workflow produces. If output keeps growing while the checking stays a full manual read, the staffing problem can keep returning.

That is the kind of moving bottleneck I have written about before. More capacity can relieve the queue, and changing the checking can reduce the amount of work entering it, those are different jobs.

Rubric, when the standard becomes part of the workflow

A rubric is a written standard for what acceptable work looks like. The next stage is having AI do much of the checking against that standard, with people handling the parts the system cannot settle.

You aren't asking a model whether a report seems good. You are checking its claims against sources, its statements about a client against the client's records, and its output against the errors the team has already seen.

In the systems I've worked with, four parts did much of that work, agents checking online facts, agents checking client data, checks for known error patterns, and guardrails for recurring language or logic mistakes.

An eval scores the result, while a guardrail can stop work from passing a boundary when a rule is broken. You need the checks to do their actual jobs, not just attach a score or a warning that nobody acts on.

The improvement comes through feedback from the people doing the work. Operations mark a piece of work as defective, implementation gets the feedback, works out the fix and ships a new version, then operations test again.

Marking a defect can take seconds. Finding and fixing the cause can take much longer, so I wouldn't turn the quick feedback step into a claim that the whole process is cheap or pays for itself immediately.

What makes it worth doing is that the same defect no longer has to be repaired by hand every time, once the fix works. The firm already spending hours reviewing bad output has a reason to invest some of that effort in changing what comes back next time.

Both sides have to take part. The person seeing the defect needs an easy way to report it, and the person receiving it needs to be able to change the system while the feedback is still useful.

A list someone opens once a month doesn't give you that response. The team needs to see that reporting a problem changes something, otherwise the report becomes another task they do while the bad work keeps arriving.

I have watched teams get slower during this stage, because a new system made mistakes and people had to check it while it was being improved. The firms that persisted and got the fixes working saw the gain after weeks to months of tuning.

That isn't a guaranteed recovery period. A slow rollout can also have the wrong tool, bad data or a workflow that isn't worth automating, so keep looking for evidence that the errors are falling and the work is improving.

If nothing changes after several attempts, investigate what is breaking instead of telling the team to be more patient. They are responding to the work in front of them, and you need to show that the next version is better.

In the franchise network's delivery data, across many projects, one week with more than 300 artifacts had roughly 95% fewer marked as carrying incorrect information after tuning, compared with that same system as first implemented.

That is a result from a tuned system with several checking layers and people feeding the improvements back into it. It isn't 95% accuracy for an AI reviewer, it isn't a comparison with a human reviewer, and it isn't a benchmark I can promise your firm.

The distinction matters because an AI checking another AI is still capable of missing the same mistake. A score that says the output is good needs its own test against errors you know are there.

A 2023 paper on ChatGPT as a factual-inconsistency evaluator found that in five of its six datasets, it could identify more than 95% of consistent summaries while performing poorly at finding all the inconsistent ones. In other words, being good at passing correct work didn't mean it was equally good at catching incorrect work.

In the 2023 HaluEval study, ChatGPT scored 58.53% accuracy on the summarization task, classifying whether a summary contained hallucinated information. The authors described that result as “barely above chance.”

Those were specific tests of models available in 2023, not a current ceiling on AI review. I use them as a reason to test the checker on wrong work, not as evidence that every model you could use today is close to a coin flip.

And a longer checklist isn't automatically a better one. Rubicon's 2025 research, which used rubrics in model training, found that indiscriminately adding more rubrics produced only marginal gains.

A separate 2025 Rubrics as Rewards study found rubrics generated with reference answers produced performance comparable to human-written rubrics in its HealthBench experiment, while generation without those references did worse. That is training research in medicine and science, not a service-firm review trial, so the business advice is my application of it.

Give the model examples of acceptable work and a real standard from someone who knows the job. I would rather spend an hour with the reviewer who catches the most errors than ask a model to invent a long list of things that sound like quality.

Maturity, when review gets smaller and better aimed

As the rubrics improve, models get stronger and the system has more history to compare against, the full human read can become less necessary for some tasks. That is the fourth stage I have watched, more quality work delivered with less review around it.

It doesn't mean nobody is responsible anymore. People still handle the work that needs judgment, the exceptions the system flags and the checks that cannot safely be left to a model.

The four review stages aren't the same numbering as the Delivery Model Ladder. Building checking into the workflow is part of the move from Stage 1, Enhanced, toward Stage 2, Augmented, on that separate ladder.

And these stages are a description of a pattern, not a scorecard for ranking your firm. There is no promised date when every task reaches maturity or a requirement to remove human review to prove you've arrived.

Sometimes a better model lets you remove a guardrail because it no longer makes the mistake that rule was catching in your tests. You still need to check that on your work rather than assume a new model release makes the rule unnecessary.

The model may cost more while leaving less work for a person. A cheaper one may still be the right choice when its mistakes are easy to catch and repair, so measure the model cost and review cost together.

That balance changes by task and by firm. The endpoint isn't the cheapest model or the smallest number of human checks, it is work that meets the standard at a cost the business can support.

And when review time does fall, decide what that time is for. The same thing that happens with faster production can happen here, the surrounding work absorbs the hours unless someone changes what the team is responsible for doing next.

There is one more reason to keep a record of the checks, it gives you an honest answer when a client asks how the work was made.

My position is that you don't need to advertise AI use simply for the sake of a label, but you must never lie about it. Follow the client's terms and any rules that apply to your work, and don't describe human review that didn't happen.

Thirteen experiments published in 2025 found people who disclosed AI use were trusted less than those who didn't, including in analyst-type work. Different wording and making disclosure voluntary or mandatory didn't prevent that effect in those experiments.

But the same research found a stronger trust penalty when the AI use was exposed rather than disclosed. I would not read that as advice to hide it, especially when your client is relying on your account of who did the work.

A 2026 study of 370 marketplace users found consumer reviews labeled AI-assisted were judged more favorably than those labeled AI-generated. That doesn't measure how your clients would react to a professional report, but it is another reason to describe the work accurately instead of treating all AI involvement as the same thing.

You only get to say someone checked it if someone did. The wording isn't a substitute for the validation.

The courts give us one historical example, not a general rule for every profession. In June 2024, the Fifth Circuit declined to adopt a special rule for AI-drafted briefs “at this time”, and reminded parties and counsel that they remained responsible for the accuracy of their filings.

That describes the court's decision then, not its entire rulebook today. Other courts and professional rules differ, and Ropes & Gray's court-order tracker is a place to look up that variation, not a reason to assume disclosure is optional in your case.

For the service-firm version, the one-page AI policy keeps the answer straightforward, client terms decide disclosure and the answer to a client's question is the truth. Build the record so you can explain who checked what without having to reconstruct it afterward.

This method needs work that repeats enough for someone to describe a useful standard. A one-off strategy engagement may still have parts you can verify, but deciding whether the recommendation is right can remain the actual service a person is being paid to provide.

It also needs someone who can improve the implementation when operations report a defect. Without that person, you have a queue of problems and a team still doing the same full review around them.

That is why human review can remain at the fourth stage, smaller and aimed at the decisions that need it. Removing every human check isn't the goal.

Pick the deliverable your firm produces most often and ask who or what checked the last one, then sit down with that reviewer and write down what they actually looked for. That gives you a place to start building the standard instead of another instruction to read faster.