Two Ways an Organization Ships Software

Every organization builds software on a spectrum between two operational models.

The first discovers as it builds. A product manager shapes a solution, writes a brief, and hands it off to engineering. The requirements are not complete; they are merely a starting point. When an engineer hits an unresolved question, the work stops. A cycle back to product is triggered, a decision is made, and work resumes. The application is designed one halt at a time.

The second resolves before it builds. Mature organizations run their discovery cycles upfront while changes are still cheap, drawing a hard line between discovery and execution before a single ticket is handed off. The requirements that reach engineering are not the beginning of discovery—they are the end of it. Execution proceeds mechanically, because the discoveries have already been made.

The cost-of-change curve is unforgiving. A requirement gap closed upstream costs pennies. That same gap discovered during execution interrupts built work, forces a reconciliation, and unwinds decisions made downstream of it. Both models eventually ship. But one pays a massive, invisible tax for the same destination—in a form diffuse enough that it rarely appears as a line item anyone can point to.

This was true long before AI. The interesting question is what AI does to it.

The AI Expectation, and the Reality

The prevailing expectation is that AI compresses this curve—that the tooling itself supplies the speed, regardless of how the work is organized.

That expectation is wrong in an important way. AI is operationally neutral. It amplifies the process you already have.

Point AI-assisted execution at the first model, and it will run prompt-by-prompt until it hits a gap the requirements never closed. With no firm line between discovery and execution, it halts exactly as a human team would, only quicker. The organization does not get its rework eliminated; it gets its rework accelerated.

Point the exact same AI at the second model, and the work runs autonomously. But the speed did not come from the AI. It came from the completeness of the discovery the AI was handed. The tooling is not the variable. The maturity of the process is.

The Experiment That Made It Visible

Flux Strategy’s Vendor Risk Assessment Tool was created earlier this year as an openly documented experiment. The first build was executed the first way—one prompt at a time, with decisions made in the moment of building. It worked, but it did not scale. Deciding everything at build time is the steep end of the curve by another name.

So it was delivered a second time, from scratch, the second way deliberately. Every decision had already been made and recorded; the discipline this time was to reconstruct those decisions completely, in advance, into a strict hierarchy—product wiki, then epics, then tickets—and run execution against that plan with minimal supervision.

The result proved the model from the inside. Where the upfront discovery was complete, an entire module of feature work ran flawlessly, the AI applying known patterns against a known structure. But where the upfront discovery was incomplete, execution stopped. Three times, the work halted because the reconstructed plan had failed to specify a foundational precondition.

These halts were not failures of the AI. They were the measurement. Execution was fast to the exact degree the upfront discovery was complete, and it stalled at exactly the seams where it was not. The honest part is that the discipline which would have caught those gaps was extracted after the third halt, not before the first—the methodology learned the cost-of-change curve the expensive way, which is the only way most organizations ever learn it.

The Bottom Line

Adopting AI does not move an organization from an immature process to a mature one. It executes whichever process the organization already has. Immature discovery plus AI yields faster halts. Mature discovery plus AI yields the near-autonomous execution everyone was promised.

Pattern Investment changes how projects accumulate cost. The Compounding Curve changes how cost translates to output. The Process Multiplier names the variable that decides whether AI delivers either: the maturity of the discovery, fixed before execution begins.

The tooling is converging, and will soon be identical for everyone. The process was always the answer—AI has simply made the difference impossible to hide.