Voice AI From POC to Production - The Key Lessons - Chris Bailey, Maersk

Prefer audio?

Who should listen: Voice and contact centre engineers, M365 service owners and UC architects evaluating voice AI vendors or moving a proof of concept towards production.

Guest: Chris Bailey — Lead Infrastructure Engineer (Voice and Contact Centre), Maersk

Chris Bailey leads voice and contact centre infrastructure at shipping giant Maersk, where he has spent the past year taking voice AI from proof of concept through to production across 50 countries. He talks through vendor assessment, telephony integration models, hidden costs and who should actually own voice AI in an enterprise.

Many thanks to Ribbon, sponsor of this episode.

Key insights

• The demand for voice AI came from the business — customer services, finance and IT services approached the voice team, not the other way round. Bailey is blunt that it fails if IT pushes it: you need the people who answer the calls to supply the SOPs and define what a good customer journey looks like. ▶ 2:38
• Budget for more than the AI licence. PSTN costs rise once calls touch the network, and the resourcing effort was far larger than anticipated — unlike a traditional IVR, a voice agent is never set-and-forget and needs continuous monitoring and fine-tuning. ▶ 6:12
• Specialist AI vendors tend to have excellent voice quality and naturalness but fall down on telephony integration — escalation to a human agent, contact centre reporting and call recording. Incumbents are stronger on integration; the job is balancing the two. ▶ 8:15
• PSTN transfer as a hand-off model is what many newer vendors propose and it does not scale — Maersk tried it and hit problems. API media streaming, which keeps the call anchored in the contact centre while media is streamed to the third-party agent, is the gold standard; Maersk is currently on a basic SIP transfer (not even a REFER) and is pushing vendors towards API integration. ▶ 9:17
• Check where the AI vendor's infrastructure sits. A US-hosted agent serving EU contact centres means calls crossing the Atlantic — latency plus GDPR exposure if recordings and transcripts end up stored outside the EU. The business will not spot this; it is IT's job to flag it. ▶ 10:48
• Cisco's 'bring your own virtual agent' gives a supported API route to third-party agents while retaining reporting visibility in Cisco; with more primitive integrations, most vendors just expose reporting APIs and you build your own dashboard to bridge the gap. ▶ 12:48
• Treat the agent like a new hire: train it, grant it system and data access, define what its job is and is not, and monitor its performance. At scale, use a second AI agent to review transcripts rather than listening back to hundreds of recordings — but check first whether the vendor even exposes transcripts via API. ▶ 13:49
• Outbound finance collections proved the biggest win — a daily Excel upload to a campaign dialler triggers a multilingual agent that chases outstanding invoices and writes promise-to-pay data back to a finance platform. Benefits were too small to claim at one or two countries; only after scaling to around 50 did the numbers stack up. Another outbound use case was killed off because it showed no benefit at small scale. ▶ 16:22
• Traditional voice teams are not automatically the right owner — prompt engineering and API integration were new skills alongside PowerShell and telephony. Decide early who owns voice AI, because ServiceNow, Salesforce and contact centre vendors all ship their own and it will otherwise appear everywhere unmanaged. ▶ 23:32

Insights summarised by AI from the episode transcript, reviewed by the Empowering.Cloud team.

Listen: Apple Podcasts · Spotify · Other platforms