Voice workflows for commerce and service
We build voice workflows for shop and service channels. They use approved catalog and order data, verify identity before account-specific answers and pass complete context to a person when the request leaves the safe path.
“Where is my order, and can I still change the delivery address?”
The shipment status can be read aloud. The address change is passed to an agent for confirmation.
No account data before verification · no silent write actionChoose the channel deliberately
Not every chatbot needs a microphone. We first identify the channel, environment and task where speech removes friction without hiding important information.
Voice should complement a visual or human route. It should not become the only way to complete an essential task.
A strong fit
Prefer another interface
Voice system architecture
Good voice UX depends on more than transcription. Each stage needs a measurable role and an explicit control before business data or actions become available.
Accept audio from the selected web, app or phone channel and establish consent.
Channel context · recording policyTranscribe language, numbers and product terms while preserving uncertainty.
Correction path · domain vocabularyVerify the user before retrieving customer-specific or protected information.
Least privilege · session scopeSearch approved knowledge or call typed business APIs for the bounded task.
Tool allowlist · validated inputsRead a concise answer, show a visual result or hand the context to a person.
Confirmation · full handover traceConversation design
Each dialogue makes data access, confirmation and fallback visible. There is no autoplay: the user chooses the scenario they want to inspect.
I need red safety shoes in size 43 for less than one hundred euros.
Should they have an S3 safety rating, or is S1 sufficient?
S3, and they need to be available this week.
I found three available S3 models. I am showing them with price and delivery date.
Public catalog
The assistant collects only the missing constraints and returns a short verbal summary alongside products the customer can compare on screen.
Where is my most recent order?
Please confirm the sign-in request on your device before I access the order.
Confirmed.
The order shipped today. I can send the tracking link to your signed-in account.
Authenticated account
Order information becomes available only after the existing account flow has verified the person and scoped the session.
The return window ended yesterday, but the product arrived damaged.
This needs a service decision. May I pass the order, damage reason and this conversation to the returns team?
Yes.
Done. The agent receives the context, so you do not need to repeat the case.
Exception handling
When policy, confidence or customer preference requires a person, the assistant passes the transcript, verified context and unresolved question.
Evaluation before rollout
A transcript can be correct while the workflow still fails. We evaluate understanding, task outcome, response time and the quality of the human handover together.
Each row shows the decision the measure supports.
Privacy and control
Audio, transcripts and spoken account details need an explicit policy. The interface should explain what is happening without forcing the customer through legal copy mid-conversation.
Controls are designed with the customer experience, not added after the pilot.
Provider, hosting and retention choices are documented for the selected channel and jurisdiction rather than promised in the abstract.
Delivery model
A focused vertical slice exposes audio, integration and handover risks before the team commits to a broad multilingual service.
Define where people speak, which requests are in scope and what remains visual or human.
Use support questions, search terms, product vocabulary and difficult audio conditions.
Add the minimum identity, catalog or order interfaces required for the complete task.
Test task outcome, latency, privacy and handover before a staged rollout.
Practical questions
Channel and risk determine the technical answer. These decisions should not be left to the speech provider defaults.
It can be either, but they are different products. We select one initial channel because telephony, browser audio, identity, latency and handover requirements differ.
Yes, through scoped application interfaces after the customer has been verified. The model does not receive broad database access, and failed verification must disclose no account information.
We test representative recordings from the intended channel, preserve transcription uncertainty and provide correction or handover paths. A quiet demo is not sufficient evidence for release.
No. Audio, transcripts and evaluation samples can have different retention rules. The selected policy depends on operational need, consent, provider behaviour and applicable law.
Only through approved, narrowly scoped tools. Sensitive or irreversible changes require deterministic validation, explicit confirmation and, where appropriate, human approval.
