AI customer service
Jev: faster decisions for Maria and TeleConvo
Someone calls, explains the problem and expects the conversation to move forward. If the assistant takes too long to respond, repeats questions or forgets part of the request, a natural voice only helps so much.
What is Jev, and why are we testing it?
This is what we are working on with Maria, Auditiv’s assistant in Portugal. We are integrating GPT-Live 1 for voice conversations and testing Jev to help interpret what happens during those conversations.
We want to reduce the wait between responses, make decisions more consistent and lower processing costs. In TeleConvo, we also see an opportunity to apply this technology to following up on sales conversations.
Jev is a model from TypeSafe AI, founded by former OpenAI researcher Diogo Almeida. It was built to answer questions with defined outcomes: choosing a category, identifying an intent or assessing a situation. It returns structured answers that software can use directly.
One call can contain more than one request.
To understand how this works, imagine someone saying to Maria:
“I’ve already asked for support, I’m still waiting, and I also need batteries.”
There are two issues here. An earlier support request needs following up, and there is a request for consumables. If the system keeps only one of them, part of the conversation gets lost.
Jev can help identify both. Our application keeps track of outstanding issues, checks the available information and guides Maria towards the next necessary question.
The same applies to apparently simple sentences such as:
“No, I was talking about the other hearing aid.”
The person is correcting information. Maria needs to recognise that correction and revise what it understood, while keeping the rest of the conversation.
In Maria’s new plan, we are organising these workflows: new support requests, follow-ups, questions about visits, consumables and other reasons for calling. Jev can help recognise the appropriate workflow, including when one call involves several needs.
How GPT-Live 1 and Jev share the work.
The division of tasks we are testing is straightforward.
- GPT-Live 1
- Handles the voice interaction.
- Jev
- Helps classify the request and events such as a correction, a pause or a refusal.
- The application
- Stores the context and controls which actions can be carried out.
Some details make this separation particularly useful. “I’m at home until Friday” gives an availability limit. It does not confirm a visit for Friday. Likewise, “I understand” may simply mean that the explanation was understood.
We want Maria to account for these distinctions. Recognising an intent is one step; carrying out an action requires the corresponding information, authorisation and confirmation.
Jev also provides signals of uncertainty. We can use them alongside the application’s rules to identify situations that need clarification. Even a high-confidence answer needs validation in our own context.
What we want to bring to TeleConvo.
In TeleConvo, our focus is on supporting the people making sales calls. A conversation might end with a request to call back, an unanswered question or a refusal. That information needs to be organised so the follow-up makes sense.
We want to evaluate Jev for this classification: helping the team understand what is outstanding and which contacts need attention. Post-call analysis could also flag unanswered questions or situations where a conversation repeated itself without making progress.
Speed and cost: measuring the complete call.
Cost makes this approach interesting. TypeSafe advertises a price of US$0.042 per million input tokens and response times between 70 and 500 milliseconds in its own measurements. These figures apply to the model; the cost and duration of a complete call also include voice, system lookups and other processing.
That is why we are measuring the integration as a whole. A quick decision only helps if it is correct and moves the conversation forward. A cheap call that forces the customer to call again achieves little.
With Maria, we are testing this combination with particular attention to corrections, requests involving several issues and situations where information is missing. In TeleConvo, we want to apply the same approach to organising and following up on sales conversations.
The outcome we want is concrete: less waiting for the caller and better context for the person who continues the work.
Could your customers spend less time waiting?
Bring an example of a conversation that went nowhere or made a customer repeat themselves. We can look at the process together and work out where AI could help.
Talk to one of our doctors