Ada and Dialpad Merge to Build the "Agentic OS" for Customer Work

Ada and Dialpad Merge to Build the "Agentic OS" for Customer Work

Toronto's Ada and San Francisco's Dialpad announced an all-share merger valuing Ada at roughly $600M, with the combined company explicitly positioning itself as the "agentic operating system for customer work" [4][5]. This is a serious combination of assets: Dialpad brings analysis of over 14 billion minutes of real conversations, Ada brings 500+ deployed AI agents running in 130+ languages [5][6]. Craig Walker takes the CEO seat, Mike Murchison becomes Chief AI Officer — a structure that tells you who's steering product versus who's steering the AI roadmap.

The deal is expected to close around November 12, and it's a clean signal of where CX consolidation is heading: voice, text, and agentic action in one stack, not three separate vendors you have to stitch together. The "unifying humans and AI agents" language isn't just marketing — it's an acknowledgment that most enterprises still need a human escalation path, and the companies that win will be the ones that make that handoff invisible.

For anyone running customer support or sales ops, this is the clearest evidence yet that point-solution chatbots are consolidating into platforms. Vendor lock-in risk just went up; so did the pressure to pick an orchestration layer rather than a patchwork of agents.

Anthropic Makes It Possible to Prove Which AI Model Actually Wrote Something

Anthropic open-sourced FlashProve, a zkSNARK-based system that lets anyone cryptographically verify a response came from a specific committed AI model — without the model owner revealing any weights [7][8]. They tested it on Llama 4 Maverick (a 400B MoE model, 4-bit) running on a single H100 at ~110 tokens/sec, generating a roughly 19MB proof per 1,024-token request, at a cost of 70-95x the model's normal runtime [8][9]. That overhead is steep, but the direction matters more than the current cost.

This is quietly one of the more important AI safety releases of the year. Provenance has been a trust problem since day one — how do you know an output actually came from the model a company claims, under the safety settings they claim? FlashProve (Apache 2.0, with both a fast C++ verifier and a formally verified Lean version) gives the industry a real tool instead of a promise [8].

Expect this to matter most in regulated industries and anywhere AI-generated content needs an audit trail — insurance, legal, healthcare, government procurement. If you're selling AI into any of those sectors, "trust me" is about to become a much weaker sales pitch than "here's the proof."

SoundHound Pushes Agentic Voice Into Motorcycles

SoundHound AI expanded its partnership with Indian EV maker Ultraviolette to bring agentic voice AI across its global electric motorcycle lineup, showcased at EICMA 2026 in Milan [10]. The feature set is genuinely useful for a vehicle category that can't rely on touchscreens: hands-free traction and ABS adjustments, pre-ride diagnostics, ride stats, and proactive alerts piped through helmet integration [10]. It builds on the OASYS agentic platform SoundHound demoed at CES earlier this year.

Voice is the obvious interface for any context where your hands and eyes are busy, and two-wheelers are a category car makers have mostly ignored. This is a smart, narrow wedge for SoundHound to prove agentic voice works in safety-critical, low-attention environments — and it's a reminder that "agentic AI" isn't just a software category anymore. It's showing up embedded in hardware you strap to your body at 120 km/h.

Mistral Throws 133 Agents at a Real CVE and Patches It in a Time-Lapse

Mistral AI demonstrated something genuinely eye-catching: Mistral Large 4 ("Le Chonk," open-weight, ~1T parameter MoE) coordinating 133 specialized agents to map a Django codebase, hunt for a vulnerability, simulate and confirm an attack, and ship a patch — all on a real CVE disclosed just a day earlier [11][12]. The whole run was captured as a 90x speed time-lapse, and it's the kind of demo that makes the abstraction of "agent orchestration" suddenly concrete [11].

This isn't a toy benchmark. Using a live, just-disclosed vulnerability gives the demo teeth — it's showing the gap between disclosure and exploit shrinking on both sides. Mistral released Large 4's open weights in late October specifically tuned for coding and cybersecurity work, and this is the use case they clearly built it for [12][13].

For European AI sovereignty watchers, this is also a statement: a Paris-based lab matching the agentic-orchestration spectacle coming out of the big US labs, on open weights, aimed squarely at security teams who can't send their codebase to a black-box API.