The Judgment Line

Most AI projects start by asking can AI do this work? That is the easier question, and answering it alone is why so many pilots stall or embarrass the people who ran them.

There is a second question, and it is the one that decides whether a deployment survives contact with a real customer, a regulator or an auditor:

May the machine do this work alone?

Technology answers the first. Accountability answers the second. When an AI output goes wrong in public, the draft was usually fine. What was missing was a person who was supposed to look at it, and a clear statement that it was their job.

The judgment line is the boundary between work a machine may finish and work a person must finish. It moves by task, industry and regulation. It does not move because the model got better.

1. 1List your processesNot tasks. A process has a trigger, a sequence, and an outcome somebody outside the process cares about. Vendor onboarding, invoice approval, claim intake, contract review, recruitment screening, month-end close. Start here rather than with a list of tasks, because tasks automated in isolation leave the process exactly as slow as it was.
2. 2Shortlist on evidence, not enthusiasmTwo patterns qualify a process, and one of them gets overlooked almost every time. The first is high volume and repetition. The second is low volume with a long cycle time, where only a handful of instances a month each take days or weeks because the work passes through several people who each have a queue.
3. 3Map the steps, and record the waitingFor each step write down what comes in, what is done, what goes out, who does it today, how long the work takes, and how long the step waits before it starts. Most process maps capture the first four and skip the last two. The last two are where the answer is.
4. 4Choose what runs each stepFour executors, and each step gets exactly one. Choose the most deterministic executor that can do the job, because determinism costs nothing at run time, gives the same answer twice, and is what makes a process explainable afterwards.
5. 5Draw the judgment lineSix questions, answered per step rather than per project, sending each step to one of three destinations. This is the part that decides not whether the machine can do the work, but whether it may do it alone.
6. 6Write it down and keep the recordThe output is one table per process: step, executor, destination, and the named person accountable. That table is both the design and the governance artifact. It is what you show a client, an auditor or your own board when they ask how AI is used here and who is answerable for it.

Compare touch time, the minutes of actual hands-on work, against elapsed time, the days from trigger to outcome.

A process that takes nine days end to end and forty minutes of hands-on work does not have a work problem. It has a waiting problem.

That distinction changes the answer completely. If the delay is waiting, then AI that halves the forty minutes saves nothing anyone can feel. What helps is removing handoffs, so that steps run when they are ready rather than when somebody gets to them. It is also why the highest-volume process is not always the one where automation returns the most, and why a low-volume process with six handoffs never surfaces if the shortlist is built on transaction counts.

Reach for AI where the input defeats a rule, not where the work is merely tedious.

Tedious and structured is a rules problem, and solving it with a model makes it slower, more expensive, harder to explain and less reliable. The invoice arrives as a PDF, so reading it is an AI step. Comparing its total to the purchase order is arithmetic, and arithmetic should never be done by a language model when a program will do it exactly and for nothing.

The executor answers what does the work. The destination answers who completes it. Never let the first answer the second.

Rules and program steps satisfy the first three questions by construction, so they finish on their own unless accountability or consequence puts a signature on them. AI steps need a confident yes on verifiability before finishing alone is worth discussing, because if checking the output takes as long as doing the work then the honest destination is that a person finishes. And for human steps the interesting question stops being whether to automate, and becomes how much can be prepared in advance.

Six steps, all four executors, and the judgment line falls in exactly one place. Every other step runs without waiting for a person, which is where the cycle time goes.

A shortlist of processes worth the effort, chosen on evidence rather than enthusiasm. A step-by-step design for each one that says what runs it and who finishes it. And a record that answers the question every client and auditor eventually asks, which is not are you using AI but where, and who is responsible for it.

The method is vendor-neutral. You can run it with a whiteboard and it will still tell you which processes to work on and where your judgment line sits. What software decides is whether the design you drew survives contact with the work, which is the problem Forge is built for. There is a longer write-up of the method on the blog.