An illustrative scenario: a sales adviser finishes a promising conversation, opens the CRM and writes only ‘call back next week’. The next colleague cannot tell which need was discussed, who makes the decision or what was promised. In support, an equally thin note can force a customer to explain a delayed order again. The first useful AI project may therefore be a better handover rather than a new automated voice.
Start with the handover problem
Conversation summaries are an established product feature: Microsoft documents summarisation for chat and transcribed voice conversations in its contact-center tools. That does not establish what an integration will achieve in your operation. Treat the following workflow as a practical design proposal to test, not as a claim about measured Viva Call results.
Choose one queue and one clearly defined output. For telesales, that might be a structured note containing the need, qualification facts, unresolved questions, agreed next step and responsible person. For customer service, it might capture the issue, checks completed, outcome and any promised follow-up. A short schema is easier to review than an impressive-looking paragraph that hides an incorrect detail.
Separate facts, interpretation and commitments
- Facts: information actually stated by the customer or confirmed in an authorised system.
- Uncertainty: a missing detail or an inference that still needs checking.
- Commitments: actions explicitly agreed, with an owner and a date when provided.
- Next step: what the receiving colleague should do, without inventing a promise.
For example, ‘interested in a demonstration’ is not the same as ‘approved the purchase’. ‘Asked about a refund’ is not the same as ‘refund authorised’. Keep these distinctions visible in the template. The agent should correct the proposed note before it becomes the working record or drives a follow-up message.
Design multilingual checks, not just translation
A Romanian multilingual call center may handle several European markets in the same operation. Agree which language the customer conversation uses and which language the internal record requires. Preserve product names, dates and important wording, and retain a route back to the original conversation where authorised. A translated summary should not quietly change the degree of interest, urgency or approval.
Review examples from each pilot language separately. Include interruptions, accents, brief answers, changes of mind and conversations where no sale or resolution occurs. If a recording or transcript is unavailable, do not pretend the system saw it. Define the permitted data sources and review any sensitive information before using external tools.
Measure a narrow pilot before widening it
NIST’s AI Risk Management Framework provides a general basis for considering trustworthiness throughout AI design and evaluation. For this particular pilot, use a small review sheet: was the customer’s need correct, was the next action accurate, was anything invented, and how much editing did the agent need? Compare documentation time with a baseline, but also check handover quality and repeat work.
- Select the queue, language and note template.
- Agree access permissions and the human approval step.
- Review representative examples, including unsuccessful and ambiguous calls.
- Track omissions and unsupported statements alongside time spent editing.
- Expand only when the team understands the failure patterns and fallback process.
Keep automation proportionate. A useful summary does not mean the tool should automatically send a sales email, approve a refund or close a complaint. Those are separate actions requiring their own rules. At Viva Call, summary assistance, knowledge search and quality insights are options to scope with the client around real processes and agreed access. A good brief describes the workflow to improve before choosing the technology.