Let each person build it first.Then fold it into the company's systems.

Things inside a company change at different speeds. ERP, membership, and reporting systems get a major overhaul every few years — and they should stay stable. The questions people need answered every day change week to week.

The traditional approach makes the slow system chase the fast questions: bring in consultants to gather requirements, hand them to IT, and wait two or three months when it's fast, half a year when it's slow. By the time the system is ready, the question has usually moved on. Whatever can't keep up falls back on people: export the reports, open Excel, piece them together one sheet at a time.

AI tools that can do the work themselves, like Claude Code, fill in the fast layer. They run on each person's own computer, read their files, work with the data the company's existing systems already produce, and handle the cleanup, calculations, and code on their own. A method can change the same day. The fast layer goes to these tools; the slow systems stay stable. The people doing the work every day know best which steps take longest and where mistakes creep in — so the fast layer is theirs to build.

Letting each person build it doesn't mean handing out the tool and walking away. In every engagement we keep three points in view.

The exercises come from their own work. Before we start, we list out the needs in each person's daily work, one by one, and build the exercises backward from them. Practice uses the company's own report structures. What people do in the session is what they'll do back at their desks.

AI learns their work first. By the end of the first session, every computer has a workspace holding that person's projects, responsibilities, regular reports, and data rules. AI reads it every time it opens — no explaining from scratch.

It only counts when the whole chain is connected. A workflow is done when it runs end to end, from data in to result out. It works from the data the company's existing systems already export, under the permissions each person already has — no separate integration project. The existing systems stay as they are; data comes out of them as before, and AI takes over the cleanup, calculations, and charts.

The first thing each person's workflow saves is their own time: reports that took hours a week to piece together, routine tasks repeated every day, now run by AI.

Then comes the company layer.

Methods are written down and stay with the company. Once a workflow runs end to end, it is written up as instructions for AI: which data, what logic, what output format. After that it reruns with one sentence, and colleagues can use it as is. The method stays in the company's hands in black and white — it doesn't disappear when its author takes leave, changes roles, or moves on.

Connections grow on their own. As more gets built, connections headquarters could never have mapped in advance start to appear: one person's tracking tool turns out to be the first half of another person's report; the numbers two departments each compile turn out to feed one into the other. Lay the results side by side, and departments propose connecting them themselves — until everyone is looking at the same numbers.

Methods that prove useful are folded in two directions: what everyone can use is shared across the company; what has stabilized and runs every day is handed to IT to build into the systems. What IT receives at that point is a method people already use every day — no need to gather requirements from scratch.

Three things: people who build for themselves, the methods they've written down, and the connections between departments. When the questions change again, they can keep adapting on their own.