Salesforce agrees to buy Cresta for about $2.7B, pricing enterprise voice AI at roughly 27x ARR

The number to extract from this deal is the multiple: about $2.7B against roughly $100M ARR, so around 27 times revenue. That is the public benchmark for enterprise voice AI that has real deployments behind it, and it says two things. Voice is where service AI pays, because one resolved call removes cost and unlocks revenue in the same interaction. And Salesforce would rather buy than build: it already sells Agentforce, and it still paid up for what a voice team accumulates over years, which is telephony plumbing, turn detection tuned against real callers, and guardrails that survive production.

What gets bought here is an engineering discipline more than a model. Voice latency is unkind: a reply that lands comfortably inside a chat window feels broken on a phone call, so the budget splits into streaming ASR, first-token latency from the live-loop LLM, text-to-speech first byte, and the transport itself. This is a tail-latency problem rather than an average one: a caller hangs up on the worst call of the day, not the median one, which is why Dean and Barroso's techniques for latency tail-tolerant services matter more here than any benchmark average. Most of the failures we see are not model quality. They are silence timers that cut a caller off mid-account-number, reasoning models left inside the live response loop, and a latency budget nobody wrote down.

The Gulf angle is latency geography and lock-in. Serving a market where your inference provider has no region adds round-trip latency you cannot optimise away, and that is the position of most UAE and Saudi contact centres whose model endpoints sit in Europe or the US. Add Arabic dialect coverage on ASR, Gulf accents on TTS and data residency rules, and this becomes a deployment problem rather than a procurement one. It goes wrong in three predictable ways: buying on demo medians instead of measuring p95 through your own SIP or WebRTC path, putting a frontier reasoning model in the live loop because it tops a leaderboard, and letting a CRM renewal settle the voice architecture before anyone has written a latency budget. The upside is symmetric: connect conversation outcomes to customer records and you hold retention and revenue signals your competitors do not.

How Azrty would approach it: a stitched ASR, LLM and TTS pipeline with a hard end-to-end latency budget, semantic turn detection over a voice-activity fallback (in OpenAI's Realtime API the difference between the default silence-timer server_vad mode and semantic_vad is exactly whether a classifier decides that a caller who trails off has finished speaking), guardrails as concurrent classifier calls that can interrupt the main response, and hedged LLM calls behind one gateway so routing, budgets and access control sit in a single place. Open-source frameworks such as Pipecat keep those stages explicit, with transport, ASR, turn detection, LLM and TTS as separate components you can measure and swap. In our work that gateway is FastLLM Proxy, which fronts both hosted providers and your own model servers. Regional inference and GPU capacity, where residency or Arabic coverage demands it, runs on Kuvryn AI, and our team in the UAE does the deployment and the support. We keep the live-loop model small and fast, push reasoning models to asynchronous work such as post-call summaries and coaching, and trace every call end to end. The starting configuration shape is roughly this:

voice_pipeline:
  transport: webrtc
  asr:
    chunk_ms: 50
    target_latency_ms: 300
  turn_detection:
    mode: semantic
    vad_fallback_silence_ms: 600
  llm:
    live_loop_first_token_ms: 300
    hedge: true
    reasoning_models: async_only
  guardrails:
    mode: concurrent
  sli:
    p50_reply_ms: 300
    p95_reply_ms: 800