AI order entry: check product, unit and price separately

“Please send two cartons of A at the same price as last time.” The intention is understandable. A reliable order still needs more information. Which pack does the customer mean? What was agreed last time? Is that price still applicable?

Our recommendation is to use the language model to interpret the message. Product matching, unit conversion and the applicable price should follow approved business data and rules. Plausible wording cannot supply a missing commercial agreement.

What the message actually tells you

The hypothetical catalogue has two entries: A-6 contains six units per carton; A-12 contains twelve. Both are called “product A” in everyday conversation. The customer previously ordered A-6. That history is a clue, not proof that the same pack is intended now.

Twelve units and twenty-four units are different orders in this example. Automation should keep that uncertainty visible. It can prepare a focused clarification question or route the transaction for review.

Separate the proposal from the order

A useful draft keeps the original message with the proposed match. For each line, make clear which information came from the message, which came from the catalogue and which decision remains open.

Check the price against the business's authoritative price source. A previous invoice may reflect a promotion, another quantity or different terms. The model should not infer a current price entitlement from that amount.

Once the customer, product, unit and applicable price are established, the draft can be approved under the agreed rules. A repeated delivery of the same message also needs handling that avoids an unintended second order.

Does this require an agent?

Anthropic distinguishes predefined workflows from systems where the model selects subsequent steps. A fixed sequence of reading, matching and checking may be sufficient here. Including a model in that sequence does not by itself call for an agent with freedom to choose actions.

More discretion can help when each case requires different investigative steps. The readable data and permitted actions still need boundaries. A product label cannot settle that design choice.

When automatic acceptance can be appropriate

A restricted catalogue and explicit customer codes can change the situation. If “A” unambiguously means A-6 by agreement, the price rule is established and the other checks pass, automatic entry may be part of the approved process. Human approval is not inherently necessary for every line.

Test difficult cases separately: another pack size, expired terms, an unclear sender and a repeated message. Specify the correct result for each. A high overall success rate tells you little if an ambiguous price is being accepted automatically.

The expected improvement in an initial project should therefore be specific: less manual transfer while product identification, quantities and pricing remain reliable. Assess the whole transaction, including clarification and corrections, to find out whether that improvement occurs.

Sources

• Anthropic: Building effective agents: the conceptual workflow/agent distinction. The recommendation and order example are editorial analysis.