The documented process is fiction
Every organization has a process on paper and a different one in practice. Automating the paper version produces a system nobody uses and everybody works around.
Home / Method
A³ — Audit, Architect, Automate.
Most AI projects fail before any model is involved, because nobody measured the process first. A³ puts the measuring back at the front, and refuses to skip it — on a single prompt pack and on an organization-wide platform alike.
01 · Audit
Find where the hours go
Map the process as it is actually performed, which is rarely how it is documented. Separate what repeats from what requires judgment, and count the waiting — queues and handoffs usually cost more than the work itself.
02 · Architect
Design the system
Decide what AI handles, what stays human, and exactly where the handoffs sit. Choose the model layer, the data path, and the behaviour on failure — because a system that fails silently is worse than the manual process it replaced.
03 · Automate
Build, test, hand over
Build the prompts, workflow or platform. Test it against real inputs and real edge cases. Hand it over with documentation your team can operate from, without the person who built it in the room.
Step one
Why the audit is not optional.
The expensive step is rarely the obvious one
Teams point at the task they dislike. The audit usually finds the cost sitting somewhere quieter — a handoff, a re-entry, an approval that waits three days.
Some things should stay manual
Judgment, escalation, and anything with a person’s wellbeing attached. Naming those before building is what stops an automation from becoming a liability.
Two scales
The same three moves, at any size.
A³ is deliberately scale-free. The difference between a prompt pack and a platform is the blast radius, not the sequence.
One vertical, one task
A fashion label’s campaign workflow. A clinic’s intake paperwork. A producer’s lyric drafting loop.
- Audit — watch the task performed, collect real inputs and real rejected outputs
- Architect — design the prompt chain, the constraints, and the guard against the failure mode observed
- Automate — ship the pack with worked examples and adaptation notes
Measured in days.
One organization, many processes
An agency’s citizen enquiry line. A company’s entire support function. An NGO’s reporting cycle.
- Audit — map processes across departments, quantify volume, find the shared bottleneck
- Architect — system design, model layer, data path, governance, monitoring and escalation
- Automate — build in checkpoints, test against live patterns, hand over with a runbook
Measured in weeks, with milestones.
Principles
What holds across every build.
Model-agnostic
Built on whichever model you choose, with the model layer isolated so it can be swapped without a rebuild.
Documented, not magic
Every decision written down. A system nobody but the builder understands is a system with a countdown on it.
Human in the loop by design
Escalation paths and review points are architected in from the start, not bolted on after the first bad output.
Handover is the deliverable
The engagement is finished when your team runs it without me — not when the build works on my machine.
Next step
Bring the process. I’ll bring the audit.
If the honest answer is that your problem does not need AI, that is what you will hear — and it costs you one conversation to find out.