At 07:15, a bakery calls its packaging distributor. It needs twelve boxes of the usual bags, a different label size and delivery to the second branch. Halfway through, the buyer changes twelve boxes to twelve packs. A conversational assistant can sound entirely confident throughout that exchange. The distributor still needs to know what to pick, where to send it and who confirmed the final version.
A useful voice ordering product ends with a reviewable order, a clear customer confirmation and a dependable route into the existing order desk. For distributors in the Netherlands, Belgium and Germany, that is a more valuable development brief than simply asking for an AI phone agent. The opportunity is to capture demand while staff are busy, including requests that need a person to finish them.
Speech technology has opened the door
Current speech tooling offers more than a plain transcript. Mistral's audio documentation describes separate batch and realtime transcription capabilities, including terminology support and speaker information in the appropriate modes. These features make richer ordering experiences possible. Their availability does not establish whether a model can understand your catalogue or your customers' accents.
Google's speech adaptation documentation explains how supplying relevant vocabulary can improve recognition of particular words and phrases. That is useful for manufacturer names and product codes. It still leaves a second, commercial task: deciding which sellable item the caller meant. Treat recognition and catalogue matching as two measurable stages.
Start the project with recordings you are entitled to use, representative accents and the vocabulary your sales team actually hears. Synthetic examples can supplement this material, but a quiet demonstration by the developer is a weak acceptance test for a noisy trade counter.
Make the order visible before it becomes binding
The application should maintain a draft with separate fields for product, quantity, unit, destination and requested date. Each new statement changes that draft. A correction must replace the intended value rather than adding an extra line. The caller should be able to hear a concise readback or open a confirmation link with the same draft revision.
Consider three similar statements: twelve individual bags, twelve packs and twelve cartons. Your catalogue needs explicit pack relationships and an agreed way to display them. The assistant may ask a clarification question. It should not select the most profitable interpretation or silently use the unit from a previous order.
Returning customers create another trap. “The usual” may describe an old order, a branch-specific standing request or a product the supplier has discontinued. Present the candidate clearly, including the relevant size and pack quantity. Where several candidates remain plausible, send a concise exception to the order desk with the caller's words and the unresolved choice.
Availability needs a clock
Catalogue knowledge changes slowly; stock can change during a call. Separate product identification from availability and commercial pricing. The app should show when each answer was checked and which system provided it. A sales assistant should not promise tomorrow's delivery because that promise appeared in an old product description.
Agree what the voice app is allowed to reserve. Some businesses will accept confirmed orders through an ERP API. Others need staff approval because prices, credit limits or delivery routes require a decision. Both are viable first releases. Describe the actual status to the customer: request received, awaiting confirmation, accepted or unavailable.
A dropped call also needs an outcome. Keep the draft long enough for the permitted recovery process, associate it with an internal reference and prevent a retry from creating a second confirmed order. Staff should see whether the customer completed the readback and whether an ERP acknowledgement exists. These are different pieces of evidence.
Design a useful handoff for the sales team
An exception screen should answer the questions staff would otherwise ask again. Show the unresolved line, candidate items, confirmed delivery details and the relevant transcript excerpt. Provide a way to correct the draft without overwriting the original request. The person resolving the issue should know whether the caller expects a callback or is waiting on the line.
Use a small set of practical exception categories: ambiguous item, conflicting quantity, unavailable stock, account identification and commercial approval. Avoid a single “AI failed” bucket. Different problems belong to different people, and the categories help you decide what to improve after launch.
Recording and retention choices belong in discovery. Explain the channel to callers, limit who can replay audio and decide when recordings can be removed while retaining the necessary order record. Establish the applicable requirements with your privacy and legal advisers for each operating market. A recording toggle in an admin screen is not a complete operating policy.
Buy this demonstration before expanding the channel
Ask a development partner to run a short set of awkward calls using a test catalogue and a sandbox order system:
- ▸The caller changes a quantity after the assistant has read it back.
- ▸Two branches use the same trading name but have different delivery addresses.
- ▸A familiar product has been replaced by a new pack size.
- ▸Stock disappears between the first availability check and confirmation.
- ▸The ERP accepts the order, but the connection drops before the app receives the response.
- ▸The caller switches language midway through a product code.
Evaluate the final order and staff workload, not just the transcript. Track how many requests reach the correct status, how long exceptions take to resolve and how often staff must contact the customer again. Compare these results with the existing phone process over a defined sample. Set your own commercial threshold before commissioning additional languages or automated fulfilment.
A first release that earns its expansion
Choose one product family, one customer group and one language for the initial order flow. Keep manual ordering available. Include a clear stop control, a documented catalogue update process and ownership of the telephony account. The supplier's quote should separate integration, voice usage, testing and ongoing maintenance so the operating cost is understandable.
For the surrounding system, our guides to API integration scope and customer branch purchasing rules explain two frequent sources of hidden complexity. Voice is an additional input channel; the underlying account and order rules still need an owner.
TuniCyberLabs develops custom software and business integrations. If your sales team already knows which calls consume the most time, discuss a voice ordering pilot. Bring a sample catalogue, the current order path and a few representative exceptions so the first scope solves a real sales bottleneck.