Eight systems in sixty days: what shipped and what broke

Eight production launches went live across June and July 2026: a communications queue, campaign pages, competitor intelligence, recall outreach, an opportunity engine, a conversations inbox, disposition tracking and a controlled vendor-activation batch.

The list looks like a software story. It was mostly a sequencing story.

The warehouse made weeks possible

None of the eight projects started by negotiating a new data feed. The DMS, CRM, website activity and useful outside sources were already landing in one warehouse. Customer and vehicle records already had stable keys. Store scope and operating definitions already existed.

That changed the first question from "How do we get the data?" to "What is the smallest useful action?"

The communications queue came first because later systems needed one governed path for outreach. Campaign pages came next because leads needed a controlled route into the CRM. Recall and opportunity workflows reused the same identity, suppression and receipt patterns. The inbox and disposition tools closed the loop.

What broke first

Fast builds expose assumptions. The common failures were not dramatic:

• a source arrived late and made a correct report incomplete;
• a duplicate record created the possibility of duplicate action;
• a store mapping changed without warning;
• a user interpreted a label differently from the builder;
• a successful batch created more follow-up than the receiving team expected;
• a retry risked repeating an action that had already succeeded.

Those are operating failures, not demo failures. They are also why every production tool needs tripwires, idempotency, capacity limits and receipts.

The controls that mattered

Source health. Every job checked freshness and expected volume before producing an output.

Masked previews. A manager could inspect representative records before approving a customer-facing batch.

Frequency caps. Contact rules were enforced by the system across workflows, not remembered by one campaign manager.

Duplicate guards. A stable action key prevented retries from sending twice.

Receipts. Every send, export and writeback produced a record that could be reconciled to the request.

Small batches. The first vendor-activation run was deliberately limited. The objective was to prove the entire loop, not to impress anyone with volume.

Adoption was a product requirement

A technically correct tool that adds a second inbox is not finished. A recommendation nobody can explain is not finished. A launch that creates work without identifying its owner is not finished.

Each system had to answer four questions for the person using it:

1. Why am I seeing this?
2. What should I do next?
3. What would stop me from doing it?
4. Where will the outcome be recorded?

What the sixty days proved

Modern software can be produced quickly. That is no longer the interesting claim. The harder claim is that the store can ship repeatedly without creating eight new silos.

The order made that possible: foundation, governed actions, customer-facing tools, then autonomous work. Every launch either reused the warehouse or improved it. Every failure became a control the next launch inherited.

Shipping is easy now. Not breaking trust is the craft. The best sign of progress is not the number of tools launched. It is that the ninth one starts with fewer unknowns than the first.