For a long time, automated support systems were designed as a rigid sequence:
question → answer → next question
This approach works as long as the person follows exactly the expected path. Real conversations, however, are less orderly: users provide information in advance, correct themselves, ask the agent to wait or do not remember a piece of information.
When the system is constrained by a script, the result is often an unnatural conversation, made up of repeated questions and continuous attempts to bring the user back to the predetermined path.
At Estro, we take a different approach: we design task-oriented voice agents, built around the goal to be achieved rather than the exact sequence of exchanges.
From the script to the goal
A task-oriented agent starts with a specific task: collecting information, understanding a problem, opening a support request or guiding the user through the completion of a procedure.
At each turn, the system evaluates:
- the final goal;
- the information already available;
- the data still missing;
- the most useful next action.
If a person provides several pieces of information in the same sentence, the agent should capture them without asking for them again.
If they say: "Wait, I made a mistake", the system should understand which information needs to be corrected. If they do not remember a piece of information, it should help them retrieve it and then resume from the previous point.
Deciding the next action
The agent should not limit itself to asking the next question. It may need to ask for clarification, verify a piece of information, accept a correction, wait, resume an interrupted activity or involve a human operator.
This logic is often referred to as a dialogue policy.
The language model understands what the person is saying and generates a natural response. The orchestrator, on the other hand, maintains control over the state of the conversation and the result to be achieved.
Understanding meaning, not just words
In a voice conversation, transcription is only the first step.
A person may stop, start again, alternate between words and spelling or correct a piece of information several times. The system must therefore understand the overall meaning of the input, rather than merely converting audio into text.
For this reason, the choice of model is important. Very small models can work in simple scenarios, but may have greater difficulty with incomplete sentences, corrections, hesitation and references to previous turns.
As an order of magnitude, for constrained processes it may be reasonable to consider models in the 7 - 14 B parameter range. Parameter count alone, however, is not enough: training, context management and the overall architecture also matter.
The choice must be tested on realistic conversations, not only on ideal examples.
Use natural voices
The quality of a voice agent also depends on how it communicates.
A monotonous voice, unnatural pauses or overly long responses make even a correct conversation feel artificial.
Responses should be brief, clear, suitable for spoken language and easy to interrupt when necessary.
Define broad guardrails
It is important to establish which goal the agent can achieve, which information it can request, which operations it can perform, when it must ask for confirmation and when it must involve a human operator.
Guardrails therefore define the boundaries of the agent’s behaviour without establishing every single sentence in advance. This makes the conversation more reliable and reduces the risk of repetition or inconsistencies.
Keep business logic outside the model
The language model can understand the user and formulate a response. Operational rules, however, must remain under the control of the application.
The validation of a code, the normalisation of an address, the verification of mandatory fields and the saving of data should be entrusted to deterministic components.
Measure the outcome
A voice agent should not be evaluated only on how natural it appears. A conversation can seem fluid and still produce an incorrect result.
The most important metrics concern the outcome:
- Was the process completed?
- Is the collected data correct?
- How many clarifications were necessary?
- When was human intervention required?
The real difference
The difference between a traditional voice bot and a task-oriented agent does not simply consist of adding a language model.
Traditional systems try to make the person follow the path defined by the software. A task-oriented agent keeps the goal under control, but adapts the path to the way the person speaks, responds and corrects themselves.
This is the approach we follow at Estro: combining language understanding, explicit state, controllable rules and sufficiently broad guardrails.
Are you planning to bring an agentic AI voice system into production? Contact us.