What does Didrah’s AI agent need from a distribution invoice?

Item, quantity, net amount, date, customer identity, and region or beat. For a campaign, a mechanic line. Without these a claim should not be built.

What should you not ask a distribution agent?

Anything outside sales, promotion, campaign, customer and channel, and managerial synthesis. Consumer intent and ad copy, if they do not join to invoices, are out of scope.

What should the agent do when the data is incomplete?

Stop and name the data gap. Filling the hole with large-language-model text is a hallucination, not supply-chain analysis.

What invoice data an AI agent can use in distribution, and what you should not ask it

Didrah’s AI agent works on invoices, customers, and campaigns. If a field is not on the invoice, the claim must not be built from meeting talk. The fields that usually suffice are item, quantity, net amount, date, customer identity, region or beat, and a mechanic line when the row is a campaign. What you should not ask: consumer intent, national share with no join, and “which brand should we build next year.” Those are not the distribution operating layer. FMCG distribution needs a bounded agent, not a model that fills every question with prose.

Which invoice fields will hold a claim

Without customer identity you cannot build concentration or coverage. Without a date you cannot build a baseline. Without net amount you cannot build incremental margin. Without item you cannot build cannibalization. Data normalization means those keys match in the sales system and the campaign system. If they do not, the agent should write a data gap. Hallucination is filling that gap with a sentence that sounds like analysis.

Which questions a distribution agent should leave unanswered

A question outside the families of sales, promotion, campaign, customer and channel, and managerial synthesis should stay empty. “Brand awareness after the spot,” if it does not join to invoices, is out of scope. “The best Ramadan slogan” is out of scope. A large language model can write the text. Didrah’s AI agent must not seat that text in place of evidence. The scope is closed on purpose so a guess does not fill the blank.

When data quality allows a claim to be published

When keys repeat cleanly, orphan rows are few, and provenance walks back to the invoice. Medium confidence means a reversible decision. Low confidence means ask again or fix the data. If a policy check blocks a figure, the answer does not get softened. That is the same split as in reading an agent answer: empty is better than falsely presentable.

What to ask from the invoice, and what not to ask

Pin the question to item, customer, date, and net amount.
If the field is missing, write a data gap. Do not leave a hallucination in the blank.
Do not ask this agent for national share or consumer intent.
Treat the closed scope as a feature, not a product gap.

What to check on the week’s invoices before you ask the agent

Scan five fields quickly: item, customer, date, net amount, region. If one is empty, fix that first. Then ask the question. Those two minutes of data quality buy back an hour of hallucinated meeting. Agentic AI on a broken invoice is faster at being wrong, not smarter. If the team asks out-of-scope questions every week and resents the empty answer, read the scope to them again. The product was not built for that question. Guessing is a betrayal of the supply chain.