We build AI workflow tools for companies, and we turn down more than we accept. The reason isn't a lack of appetite — it's that most 'we should automate this' moments are actually 'we should clean this process up' moments, and an AI tool bolted onto a messy process just fails faster.
After enough of those conversations, we settled on a three-part filter. If all three are true, the workflow is worth automating. If they're not, the fix is cheaper elsewhere.
1. The process is predictable, not creative
Triage, routing, matching, extraction, reconciliation, classification — these are judgement-like tasks with a right answer that a senior person would produce ninety percent of the time. Creative synthesis, stakeholder politics, and 'figure out what the user actually meant' do not pass this filter.
2. The volume justifies the build
The rule of thumb we use: if a full-time person couldn't spend most of their week on it, the tool will never pay for the mistakes it makes. Low-volume workflows are better served by a checklist, a form, or a single senior person with a good spreadsheet.
3. The process is documented badly
The best automation projects are the ones where the process only exists inside the head of one person who is about to leave.
That's not a reason to avoid the build — it's the reason the build is worth doing. The tool encodes the process as it should exist, which means writing it down becomes the deliverable. Half the value shows up before the model is ever switched on.
What passes the filter
Support ticket triage and escalation. Invoice and document extraction. Onboarding task routing for new hires. Data cleaning between systems that were never meant to talk. Claims and application intake. These are the workflows we keep saying 'yes' to, because the payoff is measurable in hours per week and the failure modes are mistakes a human can review.
If you're not sure whether yours passes, send it over. We'll tell you straight — including if the answer is that it doesn't.