Automate high-volume, low-judgement, reversible tasks first, and keep low-volume, high-judgement, irreversible ones human. The workflow audit sorts every task on three properties — how often it happens, how much judgement it needs, and how bad it is if it is wrong — and the combination tells you what to automate, what to assist, and what to leave alone. Most teams automate the wrong things because they choose by how impressive the tool looks rather than how the work behaves.
I have run this audit on my own work and on client teams, and the same pattern shows up every time: the tasks people want to automate first are rarely the tasks that should be automated first. This is the method I use to stop that from happening.
The three axes
Frequency. High-frequency tasks are worth automating because even a small time saving compounds over a month. Low-frequency tasks are usually not, because the setup cost never pays back. This is the axis that decides whether the effort is worth it at all.
Judgement. Low-judgement tasks have a checkable right answer; high-judgement tasks have a defensible one that depends on context. Judgement is the axis that decides whether the model can do the task or only assist it. A task can be frequent and still high-judgement — which means it is a candidate for assistance, not automation.
Reversibility. Reversible tasks are cheap to undo, so a wrong first pass costs little. Irreversible tasks — a live campaign, a public post, a client commitment — carry a cost per error that no efficiency gain justifies. Reversibility is the axis that decides how much human review has to sit on top, regardless of the other two.
The four quadrants
Run every task through the three axes and four shapes fall out.
- High frequency, low judgement, reversible. Automate first. Data pulls, formatting, pacing calculations, variant generation, transcription, translation drafts. These are the tasks where AI should be running the loop with light review, and they are where the compounding time saving lives.
- High frequency, high judgement, reversible. Assist, do not automate. Creative concepting, plan drafting, research synthesis. The model produces options and the person chooses. Full automation would remove the judgement that is the point of the task.
- Low frequency, high judgement, irreversible. Keep human, with AI as an input. A major strategy call, a client relationship decision, a brand positioning change. These are the tasks where the cost of being wrong is highest and the volume is too low to justify a system.
- Low frequency, low judgement, reversible. Automate if it is trivial, otherwise leave it. A one-off data cleanup, a formatting job. The saving is small, so the bar is "does this take five minutes or a week?"
The reversibility test
If you only apply one axis, apply reversibility, because it is the one teams underweight. The instinct is to judge a task by its complexity — "this is simple, let's automate it" — when the real question is "what happens if the model gets it wrong?" A simple task that posts to a live account is riskier than a complex task whose output a person reviews before it goes anywhere.
The test is: if the model's first pass is wrong, who catches it, and what does it cost? If the answer is "a reviewer catches it before anything happens," the task is reversible and a good automation candidate. If the answer is "the customer catches it," the task needs a human gate no matter how easy it looks. Apply this test before the frequency and judgement axes, because it can override both.
The audit method
List every task in the workflow and give each one a score on the three axes: frequency as high or low, judgement as high or low, reversibility as easy or costly. Then sort them into the four quadrants. Do it with the people who actually do the work, not in a leadership meeting, because the judgement and reversibility ratings are only honest if the operators make them.
The output should be a short, ordered list: the reversible low-judgement tasks to automate first, the high-judgement tasks to assist, and the irreversible tasks to leave alone. That list is your automation roadmap, and it will usually be much shorter than the list of things a vendor would sell you.
Sequencing the work
Automate one task at a time, end to end, and measure the result before moving on. The 75–80% principle is useful here as a heuristic for how much of a task AI can carry, not a target to hit. Start with the task that is most frequent and most reversible, prove the loop, and let the evidence earn the next one.
The sequencing also protects the team. A team that automates ten tasks at once has ten new review surfaces and no idea which one is failing. One at a time means every change is visible and every problem has a cause.
The audit, worked
A media team I know ran this audit and found something counterintuitive: the task everyone wanted to automate first — pulling the weekly performance deck — was already low-value busywork, but the task that would have saved the most time was the brief intake, which nobody had considered. The brief was frequent, low-judgement in its mechanical parts, and reversible, because a drafted brief gets corrected before any work starts. Drafting it became the first automation, and it unblocked the whole workflow behind it.
That is the usual result. The best automation candidate is often not the task people complain about most, but the one that sits at the front of a slow chain and is safe to get wrong. Mapping the workflow first, rather than listing tasks in isolation, is what surfaces it. The task that unblocks three others beats the task that saves ten minutes on its own.
Assist versus automate
The distinction between assistance and automation is worth stating plainly, because teams blur it. Automation means the task runs without a person in the loop and the output ships by default. Assistance means the model produces a first pass and a person decides. Both use the same model; they differ in who holds the final say.
The audit tells you which one a task wants. Low-judgement and reversible: automate. High-judgement, or irreversible, or both: assist at most. A task that is high-judgement but reversible is a good candidate for assistance, because the person catches the errors cheaply and the model still saves most of the first-pass time. Getting this wrong in the direction of automation is how teams ship work nobody actually approved.
Keeping the audit alive
The audit is a snapshot, and tasks change. A task that was high-judgement last quarter may become low-judgement once the context layer is good and the rules are written. A task that was reversible may become irreversible if it feeds a live account. Re-run the audit each quarter, and move one or two tasks as their properties shift.
The cadence matters more than the initial result. A one-time audit gives you a plan that ages; a quarterly audit gives you an operating model that keeps tuning itself. Pair it with the measurement frame, so each move is justified by a metric rather than a hunch.
What not to automate
Do not automate the task whose value is the judgement, even if the model can produce something plausible. Do not automate anything irreversible without a human gate. Do not automate a task you do not understand well enough to specify the output, because vague specifications produce confident wrong answers. And do not automate a task just because a tool demo made it look easy.
FAQ
What if a task is high frequency but also high judgement?
Assist it rather than automate it. Let the model produce the first pass and the options, and keep a person choosing. You capture most of the speed without giving up the judgement that makes the task worth doing.
Should we automate reporting first?
The assembly part, yes — the pulls and the formatting. The interpretation should stay human. Splitting the task this way is the difference between automating the busywork and automating the thinking.
How do we handle tasks that are irreversible but frequent?
With a human gate on every instance until the error rate is proven near zero, then a sampled gate. Never remove the gate entirely for something you cannot undo; the efficiency is not worth the tail risk.
How many tasks should we automate in a quarter?
One or two done properly beats five half-done. Each automation creates a review surface and a maintenance obligation, so the count is limited by how many you can actually run and measure.
The short version
Audit every task on frequency, judgement, and reversibility, then automate the high-frequency, low-judgement, reversible ones first. Assist the high-judgement work, keep the irreversible decisions human, and sequence the changes one at a time. Reversibility is the axis to apply first, because it decides how much review a task needs. The Execution guide covers the rest, and the for team leads track works the audit through a live workflow.
If you have run this audit, I would like to hear which task you expected to automate and then decided to keep human.
Cheers, Chandler