Teams Insider

How Pure IP Is Transforming Voice With AI and the Intelligent SIP Edge

Prefer audio?

Who should listen: Microsoft 365 and UC service owners choosing a voice carrier or SIP provider, contact centre and CCaaS architects planning voice agent deployments, and anyone responsible for telephony fraud, call recording and compliance across multi-platform estates.

Guest: Adnon Dow — Chief Strategy and Product Officer, Pure IP

Adnon Dow, Chief Strategy and Product Officer at Pure IP and formerly of Bandwidth, Cisco and Motorola, sets out how the carrier is moving from transport-only SIP trunking to an "intelligent SIP edge" that enriches calls in the network. The conversation covers the three layers where AI lands in a voice business, best-of-breed partnering versus building, LRN and reputation scoring against AI-driven fraud, and why network fundamentals decide whether voice agents work.

Many thanks to Pure IP, sponsor of this episode.

Key insights

Dow splits AI in a carrier business into three layers: machine-learning anomaly detection to monitor and manage core network infrastructure 24/7, extracting and visualising data held in systems such as CRM, and AI-enabled services (recording, transcription, voice intelligence, routing) sold to customers. Pure IP is working on all three. ▶ 3:35
The distinction to press carriers on is where the intelligence sits: traditional providers sell minutes and sessions and let the far-end platform add value "over the top", whereas enriching at the SIP edge means recording, transcription, sentiment, fraud detection and routing decisions happen at the point of entry, before the call reaches Teams, Webex, Zoom, a CCaaS or a legacy PBX. ▶ 6:38
The architecture is three-layered: global core routed network with intelligent SIP edges, a containerised microservices enterprise service bus as the integration layer, and northbound connections via open APIs, connectors and SDKs to best-of-breed recording, analytics, fraud detection, UCaaS and CCaaS vendors — Pure IP builds some services and augments the rest rather than writing its own recorder or LLM. ▶ 12:50
Voice cannot tolerate the packet loss email can — a dropped packet in an email is invisible, but in a call it breaks the conversation — so observability (MOS scoring, latency, quality of service) is layered across the whole call path rather than measured only at the endpoints. ▶ 11:15
Fraud handling uses least cost routing plus a location routing number (LRN) engine with NNIs and peering agreements with major carriers, real-time dips into third-party reputation databases and STIR/SHAKEN, so Pure IP is the numbering authority and originating calls arrive with fewer hops and a high trust score. Different customers buy different outcomes — some want lowest rate, others (hospitality chains was the example) want to be certain their outbound calls are not blocked as suspected fraud. ▶ 16:25
Getting listed on the FCC's Robocall Mitigation Database in North America means calls are blocked and numbers must be pulled from service; implementing your own LRNs and trusted peering networks is the route out. ▶ 20:31
Voice agents fall over on the basics: any latency, pause, stutter or background noise throws the agent off, and Pure IP saw its own agents hallucinate or lose context when rolling them out — so secure, low-latency network and edge fundamentals are a prerequisite for agentic AI, not an afterthought. ▶ 25:37
An agentic call path adds obligations a plain voice call does not: media anchoring, intent detection, transcription and summarisation, sentiment scoring, secure storage with retention periods set by legal requirements, contextual search and replay, plus 911/112 location and compliance enforcement. ▶ 22:34