After the Prototype - agents, applied AI, and models put to work

A voice assistant that explains a telecom bill looks finished in a demo. The moment it has to identify the caller, read billing data and take an action through a CRM API, it becomes a different system: every conversational turn has a latency budget, the CRM’s API contract drifts, calls fail halfway through, and some actions should never be taken without a human. This talk walks through the integration layer of an agentic voice assistant built as a proof of concept for telecom customer-operations use cases – explaining a bill, preventing a suspension and warning a customer before bill shock. It has not taken live customer traffic; the talk is about what building it exposed, and what would have to change before it could.

Most agent demos stop at the model. In customer operations the hard part is the boundary between the agent and the systems of record: what it may read, what it may change, how it recovers when a dependency fails, and how a human takes over mid-call with the context already attached. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027 because of cost, unclear value or weak risk controls – and integration and control design is where many of those problems start. (Source: Gartner, June 2025)

Key takeaways

1. Treat every CRM operation the agent can call as a typed tool contract – validated inputs, explicit error semantics and idempotent writes – instead of letting the model compose raw API calls.
2. Give each conversational turn a latency budget: decide what is fetched synchronously, what is pre-fetched when the call starts, and what is deferred until after the call.
3. Put guardrails on actions, not just words: separate read and write tools, require confirmation above set thresholds, and design the hand-off to a human with the conversation context attached.
4. Test with scripted call scenarios and replayable transcripts before any customer hears the bot.
5. The gap between a working POC and production for voice agents: what the POC proved, what it could not prove, and the checklist we would apply before real traffic.

Outline (30 minutes, no live demo)

• 0-4 min: the demo that worked, and the first CRM integration that did not.
• 4-12 min: the integration layer – tool contracts, authentication, read vs write paths, error handling.
• 12-19 min: latency and turn-taking on a voice channel; pre-fetch and deferral patterns.
• 19-25 min: guardrails, human hand-off, and testing with scripted calls and transcripts.
• 25-30 min: POC-to-production checklist and questions.

Who is this session for?

Engineers and architects building agents that touch CRMs, ticketing or billing systems, and product or operations leads deciding what an agent should be allowed to do on a live call.

About me

Ashok Bugude is the founder of Srishti GenAI in Bengaluru, where he builds AI agents, RAG systems and LLM applications for enterprises, including customer-service chatbot and voicebot integrations. He has more than 13 years of experience in enterprise software engineering and integration, including engineering-management roles at Sutherland Labs and Codvo.ai. https://srishtigenai.com/

Slides and video

Draft slides (architecture diagrams and anonymised scripted-call transcripts) will be shared for review if the proposal is shortlisted. The work was a proof of concept for a customer-operations team; no client names, data or screens will be shown.