Six systems, one decision, three grains deep

A channel with no history in it removes the one thing most decisions rest on: what normally happens. There was no seasonality to compare against, no conversion rate to call good or bad, and no history telling anyone which products would carry the launch.

What there was, immediately, was six systems. Advertising in one console. Price and Buy Box in another. Stock split between a freight forwarder and the marketplace. Fees buried in settlement files. Reviews and returns in a third place, and repeat purchase nowhere at all. Each one answered honestly and each one answered differently.

We built one model underneath all six, keyed so that every source landed on the same day and the same SKU. The spine was unit economics — not ordered product sales, but what was actually left after the real fees Amazon charged, freight, cost of goods and the advertising attributed to that unit.

Then the same model was made to answer at three grains. At business level: is this channel paying for itself. At collection level: which line is carrying it and which is being carried. At SKU level: the specific decision — reprice, restock, stop advertising, fix the listing, or stop making it.

That last grain is where a launch is won or lost. A launch catalogue is a hypothesis about which products the market wants, and the sooner you can retire the wrong half of it, the further the cash goes.

This is the part worth being precise about, because the category is full of claims that do not survive contact with a finance team. The model does not decide anything, and it does not find anything either. Detection is deterministic and versioned: a rule either fires or it does not, and the same inputs produce the same output every time. That is what makes a finding auditable.

What the AI does is the work no team has the hours for — reading every system every day, joining them, and writing the finding in language a person can act on. It correlates across domains that no single console can see at once, and it flags when a source has gone stale rather than quietly reporting yesterday as though it were today.

And when something fires that the numbers alone cannot explain, it does not guess. It asks. The answer goes into an append-only log, and the same question is never asked twice.

Decisions that had been made inside one system started being made across all six. The diagram below traces a decision of that shape, in the order the questions get answered: a shortfall at business level, narrowed to a collection, then read three different ways at SKU level before the call.

None of the three lenses on their own would have produced that decision. That is the whole argument for joining them.

The engagement ran its course and the contract closed. The model outlived it — the same structure now runs underneath larger accounts, which is what happens when the thing you build is a method rather than a report.