The same eight steps for every problem
A repeatable loop means you can predict what happens next, check our work at every step, and run the result without us. Most of what we build turned out to be less of an AI problem than it first looked, and the loop is how we find that out early.
- You
State the problem, messily
One paragraph, in your own words. The unpolished version is the honest one.
- Us
Right-size the tool
Cheapest approach that could work: a script before a spreadsheet macro before a workflow before a model. We say what we recommend and why.
- Us, with you
Interrogate until it is sharp
Who is affected, what happens today step by step, how often, what “fixed” looks like as a number, what data is involved, what is out of scope.
- Both
Write it down
A short brief you can read and correct. Decisions are logged as we make them.
- Us
Confirm what we can build with
Your Microsoft 365 licenses, your hardware, your vendors, your rules about PHI. Design starts from what you actually have.
- Us, you approve
Choose the simplest thing that works
Then apply the AI-or-not test, below. If plain automation wins, we say so and use it.
- Us
Build it and deploy it for real
On a real host, scheduled, with failure alerts. Nothing is finished until the runbook exists.
- Us, you edit
Write it up so it can be repeated
The runbook, the result in before-and-after numbers, and what we would skip next time.
Every one of these turned out to be less of an AI problem than it looked. Credential tracking is set arithmetic against a folder listing. Invoice validation is table geometry and OCR. The one place a model earns its place is reading a handwritten date, and even there it only proposes.