Voice AI dispatcher: from a call to a verified service request
Hear clarification. Consider interruption. Confirm the action based on the system response.
Modelled engineering case. Developed by ITTECHE to analyze architecture and solutions. Company, educational data and calculations; the confirmed result of client implementation is not declared.
Prepare a full context for the request and transfer it to CRM or to a person. The system must distinguish between the recognized phrase, the confirmed field and the actual action taken, while maintaining a clear transmission path to the dispatcher.
Go to demo
From the business task
to a verifiable solution.
The model service department takes calls for servicing industrial equipment. The dispatcher needs the site, the exact unit identifier and a description of the situation. During the conversation, the client corrects himself, names similar designations and asks to speak with a specialist. Loss of these clarifications makes the request record unusable.
- educational conversation script
- 3
- stages of dialogue analysis
- 5
- confirmation of request acceptance
- CRM
- complex treatment route
- Man
- 01
Define the role of the assistant
The greeting informs you that the automated service assistant is answering. The order of recording and data processing is agreed upon with the process owner prior to the pilot.
- 02
Wait for the clarification to be completed
Intermediate decoding is needed for the response of the dialogue. The significant field is fixed after the completed replica and client confirmation.
- 03
Prepare an enquiry
The site, equipment and description are separated from assumptions. If the designation is ambiguous, the system clarifies it or transfers the conversation to a person.
- 04
Confirm action
The acceptance message appears after the CRM responds. In case of refusal, the context for the dispatcher is preserved, without the promise that the request has already been created.
What happens after the client's response?
Text demonstration using pre-prepared cues. Choose a situation and go through five stages. Microphone, phone call, speech recognition and CRM recording are not used here.
Stage 0 out of 5
- 01
Greetings
Will open in the next step
- 02
Client Replica
Will open in the next step
- 03
Clarification and interruption
Will open in the next step
- 04
System action
Will open in the next step
- 05
Result
Will open in the next step
How the demo works
Scenarios, utterances and mock responses are defined in the code. The dialogue branch and mock CRM availability determine the outcome. An interruption changes the third stage without creating an extra action. Changing the conditions resets the walkthrough. A successful human transfer is not simulated: the demo shows a handover request with context.
From measurement -
to an explicable action.
- 01
Telephony
SIP session management and separate media stream with connection disconnection control.
- 02
Flow of speech
Speech detection, ASR and response synthesis; canceling the sounding replica when interrupted.
- 03
Dialogue state
Confirmed fields, clarification rules, allowed actions and transfer to a person.
- 04
CRM/dispatcher
Verifiable confirmation, duplicate protection and conversation context for the employee.
Infrastructure, storage and integration periods are clarified during the survey. The event's identifier, time, source data, model version, and decision are saved.
The difficulty lies in the conditions
errors and verification.
The fix is more important than the first version
The phrase “pump seventeen... no, nineteen” can change the intermediate transcript several times.
- Solution
- Store the clarification state and confirm significant details. Changing the text version should not trigger the creation of a new request.
- How to check
- A set of records with self-correction, pauses and similar titles. Check the resulting fields, and not just the quality of recognition of individual words.
Interrupt response
The client begins to speak while the assistant is still saying the previous line. Continuing a long answer prevents you from correcting the data.
- Solution
- Cancel the current synthesis and playback, listen to a new cue and update the state. A command in CRM remains a separate, confirmable action.
- How to check
- Interruption at greeting, clarification and before confirmation. Lost words, response stopping time and lack of repeated action are checked.
Passing along with context
Transferring a call in itself does not guarantee that the dispatcher answered and received information about the client.
- Solution
- Check the status of the transfer and confirmation of the workplace. If an employee is unavailable, explicitly communicate the available next step without promising a successful connection.
- How to check
- Busy dispatcher, CRM failure, call drop and human request. The log differentiates between “requested”, “accepted” and “unacknowledged”.
What needs to be confirmed
on real data.
Request quality
The percentage of correctly completed confirmed fields, routing errors, and cases of fictitious equipment are assessed using labeled dialogs.
Call quality
The delay to the first response speech, the stopping of the response when interrupted, the success of clarification and the proportion of repeated questions are measured. Average call time is not a substitute for evaluating results.
Actual transmission path
The test loop checks for busy line, CRM unavailability, and employee receiving context. A limited stream of matched service calls is then launched.
Path from survey to transmission
- 01
Scenarios and boundaries
Permitted requests, reasons for transfer to a person, order of data and acceptance criteria.
- 02
Voice stand
Test telephony, marked recordings, noise and interruptions without production actions.
- 03
CRM and dispatcher
Assertions, idempotency, request cancellation, and context transfer checking.
- 04
Limited pilot
Observation of errors, analysis of conversations and regulations for disabling automatic reception.
This is a sequence of work, not a promise of a deadline or budget. The assessment comes after checking the source data and integrations.
Where a solution is required
a different approach.
- The assistant does not give repair instructions and does not determine the safety of the equipment based on the voice description.
- Noise, accents and rare markings require separate checking; universal accuracy is not promised.
- A text demo shows the logic of the process. Calling, speech recognition, CRM and human translation are not supported here.
Let's start with the data
and acceptance criteria.
For the first discussion the following are useful:
- Impersonal examples of conversations and equipment reference.
- Enquiry scenarios and rules for handing over to a dispatcher.
- Test telephone circuit and CRM with a description of confirmations.
- Coordinated procedure for working with recording, access and storage periods.