Salesforce buys Cresta for $2.7B to put AI agents on the phone
Salesforce has signed a definitive agreement to buy Cresta, the contact-center AI company founded out of Stanford's AI Lab, and The Information puts the price at $2.7 billion. Salesforce's own announcement confirms the deal and the customer list but not the number, and no filing on Salesforce's EDGAR record mentions Cresta at all — so the price is well-reported rather than disclosed. The company expects to close in the first quarter of its fiscal 2028, roughly February to April 2027, subject to regulatory approval.
What Salesforce is actually buying is the conversation, not the switchboard. Cresta sells an autonomous voice and chat agent, real-time assistance for human reps, conversation QA, and a workflow layer it calls Deputy Agent, all running on its own orchestration stack. The customers it brings are the interesting part: Salesforce named Marriott, Comcast and Alaska Airlines, and Cresta's own materials add Verizon, United, Cox, Hilton, CVS, Intuit and CarMax — logos that already run Salesforce Service Cloud or a rival desk and now come with a voice agent attached.
The price needs context before it means anything. Cresta was last valued at $1.6 billion post-money in March 2022, when Tiger Global led an $80 million Series C; it raised $125 million more in November 2024, led by WiL and QIA with Accenture, Qualcomm and Workday Ventures in the round, and said at the time it had brought in more than $270 million total. At $2.7 billion, Salesforce is paying about 1.7 times the 2022 mark for a company that has spent four years selling into the hardest, slowest enterprise buying cycle in software. That is not a distressed price, but it is not a 2021 price either — and it says something about what the market thinks a standalone agent-assist vendor is worth once the platform owners bundle the same feature.
There is a strategic tell in the terms Salesforce did disclose. Cresta will keep supporting third-party contact-center and telephony platforms after the close, which means Salesforce is deliberately not forcing a migration to its own service stack. Reading that plainly: it bought the voice and reasoning layer that sits on top of whichever desk a customer already owns, rather than buying a desk. Kishan Chetan, who runs Agentforce and Service Cloud, framed the rationale around voice being where service is hardest to get right — an admission that the hard part of this market was never the ticketing system.
The obvious counter-argument is that Salesforce just paid up for a feature its competitors can ship for free. Genesys and Five9 own the telephony and routing layer, and both were early Cresta investors — a detail that cuts two ways. It meant Cresta had distribution into the incumbents' installed base, and it means the incumbents had a front-row seat on the product for years. Microsoft bundles Dynamics and Copilot; Zendesk sells its own resolution-priced agent; Intercom's Fin, which Salesforce effectively bought in September when it closed the Intercom deal with 30,000-plus customers attached, already does the autonomous-resolution job. Agent assist on top of someone else's switch is a thinner moat than a switch.
Where the deal gets interesting is pricing, not features. Customer service is the one AI category where buyers have already moved to outcome pricing: per resolution, per verified conversation, per contained ticket. That makes a voice agent's value measurable to the cent, and it makes the vendor's margin a function of how many resolutions it actually earns. Cresta's 100-million-plus conversations give Salesforce a dataset of resolved calls with outcomes attached — the raw material for pricing Agentforce seats against work done rather than licenses sold. That is the asset, and it is the reason a 1.7x markup is defensible.
Two things to watch. First, whether the $2.7 billion survives contact with a filing: if Salesforce's next quarterly report discloses a purchase price materially below the reported figure, the reporting was wrong, and if it discloses nothing, the deal closed small enough not to require it. Second, whether Genesys and Five9 keep reselling Cresta's agent inside their own platforms once their investor relationship becomes a competitor relationship — the partnership language in the announcement is the part most likely to quietly evaporate.
A viral research post claims a model edited another model's weights by hand and sextupled its math score. The intriguing part holds up better than the number does. The authors — Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab at MIT's CSAIL — gave GPT-6 Astra a Python kernel with direct read and write access to a small model's parameters, plus practice problems and a budget of forward passes, and let it hunt for the bug. In the published writeup, Astra traced Qwen3.5-0.8B's habit of looping in its reasoning to a direction written by two mid-stack MLP layers and subtracted it from 64 output columns per layer — 131,072 weights, 0.016% of the model.
The headline results are large: MATH-500 from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, IFEval from 21.0% to 62.3%, on 500 problems with four samples each. The mechanism check is the honest part of the paper — answers hitting the token cap fell from 93.0% to 33.1%, which is what loop removal looks like. A control model that doesn't loop gained nothing.
Then the caveats land. The dose-response is one lucky run: 10 passes got 12.3%, 25 got 14.1%, and only the 50-pass run reached 48.1%, with no seeds and no error bars. Commenters on the Hacker News thread found the untouched Qwen3.5-0.8B scores 77.6% on MATH-500 when given the context length Qwen's own model card recommends, above the "fixed" model — and the evaluation here capped output at 32,768 tokens against a card that suggests 81,920, so the baseline was measured in a regime where almost everything truncated. Independent reviewers made the same point from the other direction: this is the abliteration and steering-vector playbook, which a competent researcher can run with a script, and training is still the better way to change what a model knows.
The useful takeaway isn't sixfold math. It's that a frontier model can now be pointed at raw parameters, given a measurable failure, and left to find the functional circuit behind it — a debugging loop over weights rather than data. That is worth watching. The benchmark graph is not.
If you run contact-center AI, would you rather buy the agent or the phone system it sits on? Write to us via our contact page.
