Five dimensions. Plain words.
Move each slider to the sentence that sounds like your company today — no consultant required. The radar shows the shape. The plan underneath is assembled from your two weakest dimensions, and it always starts with one repeatable workflow. At the bottom, run any process you have in mind through five questions.
Core data in one system; exports by hand.
Some integrations, held together by hand.
Basic access control.
Copilots in use, no standards.
Some steps documented.
Nothing is ready to automate yet — and that is the useful finding. Fix one dimension, then come back.
Your first ninety days · built from the weakest dimensions
Weakest: Architecture and Security. Strongest: People — start where you are already strong, fix where you are weak.
- Days 0–30One workflow, chosen for its data
Put a small API in front of the one system you will automate around, and log every call. Agents need something to call and a trail to leave.
- Days 30–60The second gap, and a human gate
Two weeks: classify data (public / internal / confidential / regulated) and write a one-page AI-use policy. Until then, no customer data into public models. Put a person at the output: every result approved by someone, minutes per item, counted.
- Days 60–90Measure, then decide about the second workflow
Cost, time and first-pass per stage for the pilot — against the two-week baseline. If the numbers moved, the harness you built is reusable; if they did not, you learned it for the price of one workflow, not a program.
Does this process fit? · five questions
Think of one concrete process — invoices, onboarding, support triage — and answer for it.
Close the one gap first — usually the data or the rules — then run it with a person signing off.
01 · Rules, not a survey
How the plan is assembled
Five dimensions, scored 0–4 in plain language. The tier is the average. The plan takes the two lowest dimensions and puts their fixes first — always inside one workflow, never as a program.
02 · Where it comes from
One workflow beats a transformation
The pattern is the same in every case we have watched work: one repeatable stream, a harness built once, metrics from day one. Programs produce slides; workflows produce numbers.
03 · What it doesn't say
It will not pick your product for you
This compass says where you can start and what has to be true first. What to build, and whether it is worth it, stays a business decision — yours, with your team validated, not replaced.
The five things nobody puts in the demo. Each one is a gate.
None of these is a technology decision. They are the conditions that make automation possible — and the reasons most AI budgets turn into slides. Read them as the owner's checklist.
Data structure decides everything
A model can read a PDF; it cannot read a process that lives in six inboxes. If the facts of a process are not in one place, digital and owned by someone, you do not have an AI problem yet — you have a data problem, and it is cheaper to fix first.
Architecture: can it be triggered, is it logged
Agents need something to call and a trail to leave. A small API in front of the one system you automate around, and a log of every call, is the whole architecture requirement for the first workflow. The monolith can wait.
Security is a policy before it is a product
Public, internal, confidential, regulated — until data is classified, nobody can say what may go to which model. Add vendor terms, PII handling, and the two AI-specific risks: prompt injection and data leakage. A one-page policy, two weeks. Then tools.
People: the skill is spec-and-inspect
The job that appears is writing exact inputs and auditing outputs — closer to a security review than to using a tool. It can be taught, on a real workflow, in weeks. What cannot be taught in weeks is knowing the domain — which your team already has. Validate them; don't replace them.
Which processes fit
Same steps every week, rules mostly written down, inputs digital, a wrong output survivable with a human check, and volume that matters. Five yes-es: automate. Four: pilot with a person signing off. Three or fewer: fix the basics first — automating confusion only makes it faster.
Measure from day one, or don't start
Two weeks of baseline before the pilot, then cost, time and first-pass per stage after. Without numbers, adoption theatre and real advantage look identical — in DORA's 2025 report 90% use AI and more than 80% feel more productive, while delivery stability still falls. The difference is measurement.
Why start small, and why start at all.
The compass is a rule set. The pressure behind it is measured — on other companies, and on our own.