Omar, the owner of Cedar & Co., is offered an AI tool that promises to recommend how much stock to reorder each month from its own sales history. The offer sounds useful, and the risk is not obvious from a sales pitch alone. Before any business adopts a new AI tool, whether it recommends stock quantities, drafts replies, or reads specifications, someone needs to check a short, specific list of things: who owns the data it is fed, what happens if its recommendation is wrong, and whether a person reviews its output before it changes real stock, a real order or a real commitment.
This lesson mainly teaches the decision flow for evaluating a new AI tool before it is trusted with any of that. Without a check like this, a business adopts a tool on the strength of a demonstration, and finds out what it actually does only after it has already been used on real decisions.
You need: Lesson 5.1's reasoning for why an AI-use policy exists and Lesson 5.2's picture of the risks an ungoverned tool creates. This lesson turns both into a concrete pre-purchase checklist.
By the end of this lesson you can:
Not covered here: writing the AI-use policy itself (Lesson 5.1) and assigning accountability after something has already gone wrong (Lesson 5.4).
| Term | Meaning |
|---|---|
| Data ownership check | Confirming who owns and controls the information an AI tool is fed, and what the vendor may do with it. |
| Failure plan | The agreed answer to what happens, and who is told, if an AI tool's recommendation or output turns out to be wrong. |
| Vendor claim | A statement a supplier makes about what its AI tool can do, which needs checking rather than automatic acceptance. |
| Pilot period | A limited trial of a new AI tool on a small, reversible decision before it is used on a larger or harder-to-reverse one. |
| Reversibility | Whether a decision made with an AI tool's help can be undone or corrected without lasting cost if the tool turns out to be wrong. |
These three checks apply whether the AI tool drafts text, recommends a stock quantity, or reads a technical document. Data ownership matters because a vendor's own terms may allow it to keep or reuse information the business considers its own. A failure plan matters because every AI tool will eventually be wrong about something, and a business that has not decided what happens next will discover it in the middle of a real loss. A human review step matters because it is the point where a wrong recommendation is caught before it changes real stock, a real order or a real commitment.
A tool that fails any of these three checks is not automatically unusable, but it needs a stronger review step, a shorter pilot period, or both, before it is trusted with a decision that is hard to reverse.
| What can differ | The shapes you will meet | Why it matters to you |
|---|---|---|
| Export order under Make to Order | An AI tool proposed for drafting buyer replies is fed real, confirmed order information. | The data ownership check matters most here, since confirmed buyer information is exactly what must not leave the business under the vendor's own terms. |
| Domestic replenishment under Make to Stock | An AI tool proposed for stock reorder recommendations is fed the business's own sales history and acts on a decision (how much to reorder) that is only partly reversible once stock is committed. | Reversibility matters most here: a wrong reorder recommendation ties up real cash that is slow to recover, so a pilot period on a small product line is worth insisting on. |
| Hybrid order handled through a mill | An AI tool proposed for material specification drafting is shared across more than one buyer's hybrid order. | The failure plan must specify how far a wrong specification is checked back once found, since it may already be shared across more than one order. |
| Buying office or agent role | An AI tool proposed to summarise messages from several buyers at once is fed information from all of them together. | The data ownership check must confirm the tool does not let one buyer's information influence what is shown about another. |
| Different product families | Socks, knit-to-garment, terry and home-textiles businesses proposing an AI tool for forecasting or specification work all face the same three checks; only the specific data at risk changes. | The three checks apply the same way regardless of product family; only the detail behind each check changes. |
Cedar & Co. runs 12 shops on its own Make to Stock forecast, buying from Delta Garments on 60-day open account. Omar is offered an AI tool that promises to recommend a monthly reorder quantity for each product line from Cedar & Co.'s own sales history. Before agreeing, he checks the three things this lesson teaches.
On data ownership, Omar reads the vendor's terms and finds that Cedar & Co.'s own sales data remains its property and is not shared with other customers of the vendor, which passes the first check. On the failure plan, he asks the vendor directly what happens if a recommendation is wrong, and is told only that the tool "improves over time," which is not an answer to what happens on the order that was already wrong; he records this as a gap Cedar & Co. must cover itself. On human review, he decides that no reorder recommendation will be placed as an actual order until a member of his own planning team compares it against the current shop-level sales trend, at least for the tool's first three months of use.
Because the failure plan gap is real, Omar starts with a pilot on the shop's smallest product line rather than its highest-volume one, so a wrong recommendation in the pilot period costs little to correct.
The worked numberSuppose the AI tool's first pilot recommendation for the trial product line overstates demand and Cedar & Co. reorders more stock than it can sell, tying up cash the business would rather have used elsewhere. As a reference for the scale of exposure a full-line rollout would represent, Order A's own fixed order value shows what a single export order already carries:
Running the pilot on the smallest line first is exactly what limits the cost of this kind of mistake: the wrong recommendation is caught and corrected on one small line, at low cost, before the same AI tool is trusted with a reorder decision across Cedar & Co.'s full range, where a similar error would tie up far more cash.
A demonstration is built to show the tool working well, not to show what happens when its recommendation is wrong on the business's own real data. Adopting business-wide on the strength of a demonstration skips exactly the three checks that would have caught a data-ownership gap, an unclear failure plan, or a missing review step, and it does so at full scale from the first day.
The better rule is simple: run the three checks and a small pilot before any AI tool touches a decision the business would find hard or costly to reverse.
Before you let a new AI tool make or shape a real decision, ask it three plain questions: who keeps my information once I give it to you, what happens if you are wrong, and who checks your answer before it changes anything real. Then try the tool first on your smallest, easiest-to-fix decision, not your biggest one.
If the tool cannot answer the first two questions clearly, that is not a reason to refuse it outright; it is a reason to keep a person checking its work for longer before you trust it on its own.
The vendor offering the stock-reorder AI tool asks Omar to sign a full-service contract immediately, covering all 12 shops and the business's complete sales history from the start, in exchange for a lower first-year price than a smaller pilot agreement would cost.
Should Omar sign the discounted full-service contract as offered? Decide before opening the answer.
No. Omar asks for a smaller pilot agreement on one product line first, even at a higher unit cost, before considering the full-service contract.
1. No pilot period means Cedar & Co. would adopt an unproven tool across all 12 shops from day one, exactly the exposure Lesson 5.3's method is meant to avoid. 2. The data clause allows aggregated use of Cedar & Co.'s own sales data, which is a real term to negotiate or accept knowingly, not to sign without reading. 3. A lower first-year price does not offset the cost of a wrong recommendation reaching every shop at once, since that cost could easily exceed the price difference between the two contract options.
The cost of signing the discounted full-service contract without a pilot is not the price difference itself; it is the exposure of finding out how the tool performs on real decisions only after it is already running across the whole business.
Order A's fixed total order value remains a useful reference point for the scale of exposure any single AI-assisted decision in this trade can carry:
The three pre-adoption checks in this lesson cost a few hours of reading a contract, asking direct questions, and setting up a small pilot. What they protect against is discovering, only after the fact, that a vendor's data terms were unfavourable, that no one at the vendor could say what happens when the tool is wrong, or that a wrong recommendation was already applied business-wide before anyone reviewed it. Any one of these, discovered late, costs far more than the checks themselves ever would.
The cheapest point to catch a bad AI tool is before the contract is signed, not after the first wrong recommendation is acted on.
Use the T18-C05 AI governance and risk checklist workbook. In the AI tool evaluation sheet, list each AI tool under consideration, the answer to the data-ownership check, the answer to the failure-plan check, and whether a human-review step and a pilot period are planned before wider use. Start with Cedar & Co.'s own stock-reorder tool, then repeat the exercise for a tool your own business is considering.
[Download the AI governance and risk checklist workbook](../../templates/T18-C05-ai-governance-and-risk-checklist.xlsx)
1. Why does Omar start Cedar & Co.'s AI reorder tool on its smallest product line instead of its full range, given the vendor could not describe a clear failure plan?
2. A hybrid-order mill adopts an AI drafting tool business-wide immediately, based on a vendor demonstration, and later finds the vendor's contract allows the vendor to reuse the mill's specification drafts for its own product examples. Which pre-adoption check would have caught this before signing?
| EN | AR | FR | TR | Meaning |
|---|---|---|---|---|
| Data ownership check | فحص ملكية البيانات | Vérification de la propriété des données | Veri sahipliği kontrolü | Confirming who owns and controls the information an AI tool is fed, and what the vendor may do with it. |
| Failure plan | خطة التعامل مع الخطأ | Plan en cas de défaillance | Hata planı | The agreed answer to what happens, and who is told, if an AI tool's recommendation or output turns out to be wrong. |
| Vendor claim | ادعاء المورّد | Affirmation du fournisseur | Tedarikçi iddiası | A statement a supplier makes about what its AI tool can do, which needs checking rather than automatic acceptance. |
| Pilot period | فترة تجريبية | Période pilote | Pilot dönemi | A limited trial of a new AI tool on a small, reversible decision before it is used on a larger or harder-to-reverse one. |
| Reversibility | إمكانية التراجع | Réversibilité | Geri döndürülebilirlik | Whether a decision made with an AI tool's help can be undone or corrected without lasting cost if the tool turns out to be wrong. |
Next you will learn how to name one accountable person for a business decision, even when an AI tool helped produce it. Back to the course map