Before you automate anything, trace the workflow
Most AI projects fail before a model is ever chosen — they start from the wrong step. Here's how we find the one that's actually worth automating.
The most common question we get isn’t “what can AI do for us.” It’s “where should we point it.” Those are different questions, and answering the second one wrong is why so many AI projects ship something nobody uses.
The org chart lies
Ask a leadership team where the bottleneck is, and you get the org-chart version: the process as it’s supposed to work. Ask the person who actually does the work, and you get something closer to the truth — a workaround, a spreadsheet nobody officially owns, a Slack thread that stands in for a missing system.
We trace workflows end to end, following the actual work rather than the process diagram:
- Where does the work enter? Not where the org chart says — where the first person to touch it actually receives it.
- Where does it wait? Queues are where the real cost hides. A four-day approval step, mostly spent waiting on a calendar, costs more than the ten minutes anyone spends reading it.
- Where does a person make a judgment call a machine could make consistently? Not “could a model touch this” — could a defined rule or a well-scoped model make this specific decision as well as the person does, most of the time, with a clear escalation path for the rest.
Why this order matters
Skip the trace and you end up automating the visible 20% of a process while the actual bottleneck — usually a wait state, not a task — goes untouched. The team ships something, adoption is low, and the conclusion drawn is “AI didn’t work here,” when the real conclusion is “we automated the wrong step.”
The trace is unglamorous. It’s also the entire reason our engagements start with an assessment before a single line of a system gets built.