The order book data model: what an AI needs to read from it
An AI assistant is only as good as the structure of the order book it reads. This explainer sets out the entities, keys and status fields that matter.
KEY TAKEAWAYS Summary by the editors
- A wholesale order book data model links accounts, products, orders, delivery windows, prices and statuses, and an AI agent can only answer reliably if these links are explicit.
- Global trade item numbers (GTINs) carried in EAN and UPC barcodes identify products across supply chains, which makes them a natural key for linking order lines to stock and sell-out data.
- Clear status fields (confirmed, split, cancelled, shipped, invoiced) matter more to AI reliability than extra descriptive fields.
- The Le New Black summary of State of Fashion 2026 expects retailers to want structured attributes, consistent wording and clear categorisation from brands.
- Before connecting an AI agent, test it on a read-only copy and check answers against known orders.
An AI that reads a wholesale order book needs a consistent model of who ordered what, at what price, for which delivery window and in what status. If those relationships are stored as free text or scattered across systems, the AI will guess, and guesses in order data are costly.
What is an order book in wholesale?
The order book is the record of all confirmed and open orders from retail customers for a season or period. It includes the account, the ordered styles and sizes, quantities, prices, terms and delivery dates. It drives production commitments, financial forecasts and customer service, and reps use it to see what an account has already bought.
Seen from an AI's point of view, the order book is a set of tables with business meaning attached. A human rep knows that a line with quantity zero and a cancellation code means the retailer dropped a style, and that a line with a changed delivery date is probably a negotiated move. The AI knows none of this unless the meaning is stored in the data or in documentation it can read.
Fashion adds complications that other industries do not have. A single style comes in several colours and sizes, often as size runs or prepacks, and orders may be placed for several delivery windows in the same season. A model that treats a style as a single item will misread quantities and values.
Which entities should the data model contain?
| Entity | Key fields | Why the AI needs it |
|---|---|---|
| Account | Account ID, legal entity, ship-to addresses, segment, currency, terms | Answers who is ordering and where goods go |
| Product / SKU | Style, colour, size, GTIN, season, category | Links order lines to inventory and sell-out |
| Order header | Order ID, date, channel, rep, status | Shows origin and lifecycle |
| Order line | Quantity, unit price, discount, delivery window, line status | Allows partial shipments and cancellations to be read correctly |
| Price list | Price list ID, validity dates, currency, rules | Explains why a price was applied |
| Delivery / shipment | Ship date, carrier, tracking, quantities | Closes the loop to what arrived |
Two further structures are worth modelling explicitly: the assortment (which styles were offered to which account, under which conditions) and the order change history. Assortment data lets an AI answer questions about what an account did not buy, which is often more valuable to a rep than what it did. Change history lets it explain how an order evolved.
Why do keys and identifiers matter?
An AI joining order lines to stock, returns or retail sell-out must be able to identify the same product in each system. GS1 describes EAN and UPC barcodes as the longest-established and most widely used of its barcodes, printed on virtually every consumer product. Where a GTIN is carried on the SKU record, matching across systems is straightforward. Where products are matched by descriptive names, errors creep in, especially when colour names differ by system.
The same applies to accounts. If one retailer appears under three slightly different names, an AI asked about total orders from that retailer will give an incomplete answer unless a master account ID exists.
The Le New Black summary of the State of Fashion 2026 expects retailers to look for structured attributes, consistent wording and clear categorisation in product data supplied by brands. The same discipline helps internal AI use: consistent attribute names, such as always using one definition of 'colour' and one list of size scales, reduce silent errors.
Which status fields prevent wrong answers?
Many incorrect AI answers about orders come from ambiguous status. Is an order 'open' because it is unconfirmed, or because it awaits shipment? Does 'cancelled' apply to a line or the whole order? Define a small, mutually exclusive set of statuses at both header and line level, with timestamps for each change. A line-level history lets the AI answer questions such as when a delivery date moved and who changed it.
- Use controlled lists, not free text, for status, reason codes and channel.
- Record the original and the current delivery window.
- Store price source (list, special, manual override) on each line.
- Keep currency and tax treatment explicit.
- Retain change history for audit and for the AI to explain changes.
Prices deserve the same treatment. An AI asked why an order total differs from the price list needs to see whether a discount, a special price or a manual override was applied, and who approved it. Without this, it will either guess or report a discrepancy with no explanation, and neither helps a rep who must answer a buyer.
How should an AI agent access the order book?
The safest approach is a read-only interface exposing curated views, not raw tables. Views can pre-join accounts, products and orders and apply access rules so that a rep sees only their accounts. Writing back, for example creating a draft order, should be a separate, supervised step.
McKinsey's State of AI survey (2026) reports that 40 percent of respondents at organisations with more than 1 billion dollars in revenue say they are scaling AI agents, while the share among smaller organisations is 22 percent. Fashion-specific figures are not given, but the numbers show that agent use is spreading and that smaller companies in particular need simple, safe data interfaces.
Access control is part of the data model, not an afterthought. Reps should see only their own accounts, regional managers their region and finance everything relevant to credit. If an AI assistant runs under a single privileged service account, it can leak information across those boundaries. Pass the user's identity through to the data layer so that existing permissions apply.
How do you test that the model works?
- Pick 20 real questions reps ask, such as what an account ordered last season and what is still unshipped.
- Answer them manually from the source systems.
- Ask the AI and compare, noting every difference and its cause.
- Fix the data or the view, not the prompt, wherever the error is structural.
- Repeat after every change to the order system.
Keep a log of every question the AI answered wrongly and classify the cause: ambiguous status, duplicate account, missing key, stale data or a misunderstanding of the question. Over a few weeks the log shows which fixes matter most, and it becomes a regression set that can be re-run whenever the order system or the AI tool changes.
Frequently asked questions
What is an order book data model?
It is the structure that links accounts, products, order headers, order lines, prices, delivery windows and statuses so that orders can be read and analysed consistently.
Why does an AI need GTINs or other product identifiers?
Identifiers let the AI match order lines to stock, shipments and sell-out data without relying on descriptive names, which vary between systems.
Should an AI agent be allowed to change orders?
Start with read-only access to curated views. Allow drafts or suggested changes only with human approval and a full audit trail.
How can I test whether my order data is AI-ready?
Use a set of real questions with known answers, ask the AI, and investigate each mismatch. Most errors trace back to ambiguous statuses, duplicate accounts or inconsistent product keys.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
