Most automation projects that disappoint were aimed at the wrong steps. The technology worked; it was pointed at work that either did not need automating or should never have been automated in the first place.
Before writing anything, we sort every step of a process into one of four buckets. It takes an afternoon and it changes what gets built.
1. Automate outright
These steps are repetitive, rule-governed, and have a clear right answer that anyone doing the job would agree on.
Copying a purchase order number from an email into a finance system. Filing a signed document in the right folder. Sending a reminder three days after a quote goes out. Checking whether a form has all its required fields.
The test is simple: if two competent people did this step, would they produce the same result? If yes, a machine can do it, and the value is straightforwardly the time it takes multiplied by how often it happens.
2. Augment with AI, keep the human
These steps need judgement, but most of the effort is preparation rather than judgement.
Marking a piece of student work is the clearest example. Reading the submission, checking it against each rubric criterion, and drafting feedback is hours of work. Deciding whether the argument is actually any good takes a few minutes and requires an educator.
Automating the preparation and leaving the decision is where AI earns its keep. The person still owns the outcome; they just stop doing the typing.
The risk here is subtle. If the draft is good enough most of the time, people stop reading it carefully. Design for that: show confidence, highlight what the model was unsure about, and make the review step feel like reviewing rather than rubber-stamping.
3. Leave alone
Some steps are rare, or the cost of getting them wrong is high, or they are the part of the job people actually enjoy.
A step that runs four times a year is not worth automating even if it is tedious. The build cost will not come back, and infrequently-used automation rots quietly — nobody notices it broke until the fifth time.
Steps involving difficult conversations, negotiation or genuine discretion belong to people. Automating the message that tells a customer their shipment is delayed is fine. Automating the conversation about compensation is not.
4. Delete
The most valuable outcome of mapping a process is discovering steps that exist for no reason.
A report nobody reads. A double-entry that a system integration made redundant three years ago. An approval that was added after an incident in 2019 and now just adds two days. A spreadsheet maintained in parallel with the system of record because someone once did not trust the system.
Automating a pointless step makes it permanent. It becomes invisible, and now it is in code, so removing it needs a developer. Always ask what would break if a step simply stopped — surprisingly often, nothing would.
Sequencing what is left
Once the steps are sorted, order the automation work by volume × effort, not by how interesting it is.
The most technically satisfying part of a process is rarely the most valuable. A well-built document extraction step that runs forty times a day beats an elegant agent that runs weekly.
Start with one high-volume step, get it live, and let the team use it for a fortnight before building the next one. You will learn things in that fortnight that no amount of planning would have surfaced — usually about the exceptions nobody mentioned during discovery, because to them it is just “what you do when the form is wrong”.
The short version
Map the process. Sort every step into automate, augment, leave, or delete. Delete first, because it is free. Then automate the highest-volume rule-governed steps. Then augment the judgement-heavy ones, with the human genuinely in the loop.
If you would like help doing that on one of your own processes, tell us what it is — the mapping is the part we do first anyway.