A buyer sends a WhatsApp voice note. How do you turn it into a reliable B2B order?

The client problem was not voice. It was order-entry time.

Some B2B buyers know the catalogue almost by heart. They do not want to browse hundreds of product pages every time stock runs low. They send a message to the account team: twelve boxes of one reference, six of another, delivered to the usual site. WhatsApp is already part of the commercial relationship because it is quick, familiar and works from a warehouse floor or a vehicle.

The friction appears on the other side. A team member listens to the message several times, searches the catalogue, checks pack sizes, finds the customer account, copies quantities and rebuilds the order line by line. On the project behind this article, a large order could take roughly twenty-five to thirty minutes to enter. The client did not need a novelty chatbot. It needed to remove that repetitive translation job without losing control of the order.

Design the journey as dictate, understand, resolve and confirm

A useful voice-ordering flow is a small operating process, not one magical model call.

Let AI interpret language, but let Medusa own commerce truth

A language model is useful when a person says ‘give me ten of the large blue ones we bought last month’. It can identify intent and propose search terms. It should not decide that a product exists, choose an unavailable variant or manufacture a price. Those facts belong to the commerce system.

We therefore separated interpretation from validation. The AI layer produces a bounded draft. Medusa resolves that draft against products, variants, customer identity, price lists and availability. If two products are plausible, the interface asks. If a quantity conflicts with a pack size, the difference is shown. This keeps the conversational surface flexible while the transaction remains deterministic.

That separation also makes the feature maintainable. The catalogue can change without retraining a private vocabulary into the model. Customer-specific prices remain governed by the same rules as every other order. Voice is simply another way to express purchase intent before the normal commerce contract takes over.

See the original voice-ordering workflow

The product is the ambiguity screen

Voice demos often stop as soon as words become JSON. Real order entry starts there. A buyer can use an old product name, omit a unit, change their mind halfway through a sentence or request a product that has several sizes. Background noise and accents add another source of uncertainty. Hiding those cases behind a confidence score would make the flow look smooth and the resulting orders unreliable.

The confirmation surface should say what the system heard and what the catalogue resolved. Each line needs a clear product, variant, quantity and unit. Alternatives should be offered only when they are genuinely compatible. An unresolved line should block confirmation or move to a human review queue; it should never be silently dropped.

This is where the time saving becomes credible. The operator no longer rebuilds every line, but they can still resolve the few lines that need judgement. The system automates the predictable majority and concentrates attention on exceptions.

A product match is not enough: the order needs customer context

In B2B, the same product request can produce a different transaction for two buyers. Contract pricing, allowed delivery sites, minimum quantities, payment terms and available assortments may belong to the customer account. A voice note that identifies products but loses the buyer’s commercial context has automated only the easiest half of the job.

The WhatsApp identity therefore needs a controlled relationship to a known contact and company. That relationship can be confirmed during onboarding or resolved by an operator when a new number appears. The order proposal should show the account and delivery site it will use, especially when one buyer works across several locations. A conversational channel should not become a shortcut around account permissions.

Previous purchases can improve the experience without becoming hidden authority. They help rank what the buyer may mean by ‘the usual filters’ or ‘the same gloves as last time’, but the current catalogue record still supplies the selectable variant, price and availability. History helps interpret language; it does not rewrite today’s commercial truth.

Roll out around exceptions, not around the perfect demo

A sensible rollout begins with a bounded group of repeat buyers and a catalogue whose references are well understood. Measure how many proposed lines resolve automatically, how often users edit quantities or variants, which phrases create ambiguity and how much operator time remains. These signals reveal whether the bottleneck is transcription, catalogue quality, customer data or the confirmation interface.

Keep the original message, transcription and final confirmed structure connected for support and improvement, with retention appropriate to the business and the consent given. When a correction occurs, record the difference without treating every human edit as training permission. Operational evidence is valuable; customer conversations are not a free dataset.

The success metric is not the percentage of messages processed by AI. It is the share of valid order intent that reaches a confirmed transaction faster, with fewer manual searches and no loss of buyer confidence. A system that escalates ten difficult lines clearly can be more useful than one that claims full automation and creates one expensive wrong order.

Why Medusa was the right foundation

The client needed a new ordering surface without creating a second catalogue or a parallel order system. Medusa already supplied the product, customer, pricing and order foundations. Its extensibility let us add WhatsApp ingestion and AI-assisted interpretation around those foundations instead of forcing the entire workflow into a rigid storefront checkout.

That is the architecture decision behind the feature. We did not choose a composable platform because flexibility sounds attractive in a presentation. We chose it because this client’s buying behaviour did not begin on a product page. The platform had to meet buyers in their existing channel and still produce a governable ecommerce transaction.

The outcome: voice became a practical ordering shortcut

Around thirty large orders were placed through the voice flow during the experience captured in the original field note. A journey that had taken roughly twenty-five to thirty minutes could be completed in about five minutes: dictate, review and confirm. The important result was not that speech could be transcribed. It was that an informal message could enter the same controlled commerce system as every other order.

For merchants considering conversational commerce, the useful question is therefore not ‘should we add AI?’. Ask where customers already express purchase intent, how much manual work sits between that intent and an order, and which decisions must remain attached to catalogue truth. If that gap is large, a bounded voice workflow can create real operational leverage.