Why does a money product need laws at all?
An owner who is asked to trust a recommendation about a reorder, a markdown or a win-back campaign is being asked to trust the arithmetic under it. If the arithmetic was produced by a language model, the owner cannot check it, and a confident wrong number is worse than no number. The three laws exist to make every figure checkable and every action reversible before it happens.
The commerce engine treats the laws as architecture rather than marketing: the analytics code is separate from the AI agents, the agents receive figures through tools, and every outbound path runs through one screen. That separation is what lets the same platform be both fast with numbers and careful with money.
Law one: the backend calculates
Revenue, margin, cover, return rate, conversion, tax and landed cost are computed by a deterministic, unit-tested analytics engine using set-based SQL against the platform's own database. Deterministic means the same inputs produce the same output every time; set-based means the calculation runs over whole tables rather than row by row, which is why it stays fast against millions of order lines.
The consequence for the owner is simple: the same number appears on every page. The dashboard's net sales, the product analyst's net sales for the same range, and the figure a chat answer cites are one figure from one calculation. There is no reconciliation meeting between two reports, because there is only one report underneath.
Law two: the AI interprets
AI agents in the engine receive figures that already exist and turn them into verdicts, narratives, drafts and recommendations. They write the Morning Brief's commentary, the Executive Report's prose, the draft email to a store manager and the answer in the chat analyst. No AI in the product can produce a number of its own; if a number appears in an AI answer, a backend tool produced it.
That is visible on the surface, not just promised. A chat answer carries source chips and a trust footer that shows exactly which tools were called and how fresh each system's data is. An Explain-this marker on a KPI card lets the analyst independently recompute the figure through the same tools. The AI's job is to read and explain, and the surface shows it doing exactly that.
Law three: a human approves
Every action that leaves the platform, whether an email to a manager, a list pushed to a marketing tool, a post published, a blog article, or ad spend, is shown in a Review & Send screen first. The screen shows what will leave, to whom, and on what evidence. Unattended sending is an explicit switch, off by default, and where it is turned on the send is still logged and later scored in Results.
This is the same discipline the Digital FTEs on ftes.ai work under: an FTE proposes, the manager approves, and autonomy is granted per class of action and can be taken back. In the engine the gate is a screen; in the workspace it is the Approvals tab. The principle is identical, which is why the two fit together.
The fourth promise is honest data
When a figure cannot be computed, the page says why. A product with no cost data shows No cost data, not a margin of zero. A metric a source does not report shows Not reported. Where only part of a catalogue carries the inputs, the coverage is stated so the reader knows what share of the picture the figure describes.
Test orders and catalogue errors are excluded from every calculation and the exclusion is disclosed. The point is that a number on a page is a claim, and a claim the engine cannot support is not shown as if it could be. Advice built on shaky data goes quiet and says so, which is a feature, not a failure.
What do the laws mean for a buyer, an owner and a CFO?
For a buyer, the laws mean a reorder proposal comes with the sell-through, the cover and the cash locked in the line, all computed the same way the dashboard computes them. For an owner, they mean nothing was sent to a customer, a manager or an ad platform without a signature, and the record of every send is kept. For a CFO, they mean the margin figure can be traced to a query, not to a prompt.
The next step is the same for all three: connect the store read-only, let the first brief be written from the store's own numbers, and read the trust footer on the first answer. The laws are easiest to believe when you have watched them hold on your own data.