How we work.
We watch first. Then every rule gets a fate. Then we change the system while it runs.
Every system has three versions.
What the code does. What its builders meant. What your people actually do.
Only the third version runs your business, and it is the one nobody wrote down. The real process lives in the workarounds: the spreadsheet beside the app, the step you never click on a Friday, the re-keying a clerk does because the export is wrong. Years of small adaptations, each one made for a reason.
A rebuild copies the code and the spec, because those are what the team can see. It deletes the adaptations, because nobody knew they were there. The new system passes every test and stops working on Monday morning.
So we watch first. Not interviews, which tell you what people think they do. A recording of what they actually do.
There are five ways these projects fail. The field note walks through all five.
Five steps, and what you get from each.
Every engagement runs the same five. Each one ends in something you can hold.
01
We watch before we touch.
A small recorder runs beside the old system for about thirty days. It logs which screens get used, which paths people take, where the errors are, and where the same values get typed twice. Nothing in the system changes. The data is aggregated and anonymized, and it stays yours. Two days shadowing your people fills in what the recorder cannot see.
You get: the usage map. Screens ranked by use, screens never opened, the hot paths, and the workarounds management never knew about.
02
Every rule gets a fate.
Every business rule we can find goes into a ledger with where it lives, why it exists, and a confidence mark. Then a fate: keep, change, or retire. Then a harm rating: if we get this wrong, who is hurt, how would we notice, and how fast. High harm and slow detection go first on the test list. Two reviewers decide separately, and where they disagree the recording settles it.
You get: the rule ledger, and the disagreement list with every item resolved.
03
We keep what makes it work.
Some workarounds are mistakes. Some are the process, invented by the people who run it because the system got something wrong. The ledger marks which is which, and the rebuild absorbs the second kind instead of erasing them.
You get: the list of workarounds that survive, and the ones designed away on purpose.
04
We change it while it runs.
The plan is staged, and each stage has a price, a stop point and a kill switch. Old and new run side by side until the new one has earned it. You can stop after any stage and still be better off than when you started.
You get: a staged plan your board can price, and a business that never goes dark.
05
We keep watching after launch.
The recorder stays on. Did the workarounds go away? Did the hot paths get faster? Did anyone start re-keying again? Before and after, as a number.
You get: the before-and-after, and a measurement loop instead of a demo.
Five gates, from discovery to production.
The steps are the brain. These gates are the skeleton. Each gate is a decision and a payment milestone, so you always know where the project stands, and you only pay for progress that has been shown.
01
Capture
Watch the old system run, surface the rules buried inside it, and get them out of one person’s head. Every rule goes into the Control Matrix with its source, its purpose and a confidence mark.
Gate 1, discovery: do we actually understand the old system, and how it is really used?
02
Level-Set
Agree the future-state plan, what done means, and how changes will be handled, before a line of code. Every rule in scope gets a written acceptance test.
Gate 2, design: are the plan, the definition of done and the change process all agreed?
03
Engineer
Build in increments, validated as we go. Each increment is demonstrated, and its acceptance tests are closed with you before the next one starts.
Gate 3, build: functionally complete, and accepted increment by increment?
04
Assure
Formal validation across the whole system: security, performance, and a full rehearsal of the data move, reconciled against the record. A confirmation, not a first look.
Gate 4, acceptance: validated and ready for production?
05
Release
Cutover with a rollback ready, a stabilization window, training and runbooks, so the people who run it on Monday morning are ready to run it. The recorder stays on.
Gate 5, production: live, stable and handed off?
Two records that run the whole way.
The things we are strict about don’t get a phase. They run from the first day to the last.
The Control Matrix
The rule ledger. Every rule the old system follows: where it lives, why it exists, what the recording showed, and what happens to it. Keep, change, replace or retire. A confidence mark flags the rules nobody fully understands yet, and those get a person’s attention before they can sit on the critical path.
The Acceptance Ledger
For every rule we keep or change, a test agreed in advance: given this, the system does that. You sign off increment by increment, so the final acceptance is a confirmation, not a surprise.
Change control
Agreed before any code is written. A clarification is absorbed. A small change is logged and approved quickly. A material change gets one estimate of the time it adds, and you choose: move the date, or add capacity to hold it.
You cannot accept what was never captured, and you cannot claim it was captured without proof it was accepted. The two records lock each other.
AI reads. A person decides.
We use AI to read the old code, draft the rules it finds, generate tests and write migration scripts. That is where the speed comes from. But AI tells you what the code does, never whether the business still needs it, and it will copy a bug as cheerfully as a feature. Nothing enters the Control Matrix, the Acceptance Ledger or the codebase on a model’s say-so. Where two reviewers disagree about a rule, the recording settles it, not the model. The person’s decision is what gets recorded.
Have a system that has outlived its documentation?
That is exactly the conversation we want. Tell us what you are running, who depends on it, and where it hurts.