Anthropic Runs Its Go-To-Market on Salesforce. That Should Tell You Something.
I had a specific assumption about how frontier AI labs run their go-to-market, and it was wrong.
I assumed they had built something homegrown. Something I was not smart enough to understand. Then SaaStr published a breakdown of the hyperscaler GTM stack, and the answer was six tools most of the people reading this have already implemented: Salesforce, LeanData, Gong, Clay, Ironclad, Slack.
The companies with the deepest access to frontier models on earth are running their revenue motion on off-the-shelf software.
We hosted Paul Wilson to figure out why. Paul spent 20-plus years in GTM systems, including Marketo through the Adobe acquisition and Slack through the Salesforce acquisition, and now sits at Coastal advising mid-market and enterprise companies on this exact question. He opened by saying roughly 96% of any conversation on this topic is speculation, because nobody has a playbook yet. That is the honest starting position, and it is the right one.
The question everyone is actually asking
Should AI be the control tower for how your commercial team interfaces with your systems, or should you point it at specific problems?
Paul's answer came in three parts, and the first one is the one that stops most conversations cold. A CRM is not a database. It is a set of business rules written in software. Which products can sit on which opportunity in which geography. Whether this customer has entitlement to open this case. What happens when a form gets filled out and routed. If you move that into a model, you have not eliminated the software. You have agreed to become a software developer, permanently, without the team of one.
Second, models are non-deterministic. That is architecture, not a bug you can prompt away. A rules engine that can be talked into a discount it was never authorized to give is not a rules engine.
Third, cost. We are in a land grab where capability is priced at almost nothing. Build a bespoke system on that assumption and the bill arrives in eighteen months, at which point you need to modify software you did not write and cannot maintain.
Why the hyperscalers compose
Paul's analogy landed for me for obvious reasons. If you make wine, are you also going to manufacture the bottles? The corks? The cardboard? Your product genuinely needs all of it. None of it is your business.
Anthropic's core business is shipping their product. The fastest, most stable path to market was composing a stack from mature products that had already solved routing, CPQ, contract workflow, and conversation intelligence. In two years they may have built something of their own. Today they bought the bottle.
What this means for the mid-market
Three things we are telling clients.
Point AI at friction inside the rules engine, not at the interface. The moment worth chasing is not "reps stop logging into Salesforce." It is lead-to-opportunity, prospecting agents, in-app coaching for sellers building quotes. Paul has seen seven-figure monthly AI bills at companies that cannot yet trace the spend to a single improved business process, because the licenses were handed out without connecting them to anything that runs the business.
Classify your data before you try to govern it. Core data supports your business rules and your relationships. It has to be protected and kept clean. Transactional journey data, the transcripts and page visits and call topics, informs the customer experience. The fidelity bar is different for each. You do not need every timestamp to be exact. You do need one agreed definition of a customer.
Stop waiting for clean data. Josh Vandenbor made the best point of the session, and it reframed my own question. The question is not when to introduce AI. It is what the first problem is. One company in paralysis over data quality correlated exec-turnover records from Workday against their funnel data and found that every new CRO and CMO had quietly redefined the stages. That explained years of inconsistency nobody could account for. A team of a hundred people would have spent weeks finding it.
The part for the operators
Paul's advice for anyone doing this work: keep running into the burning building. Business process is humans, movement, and data. The tools will keep changing and the technical skills are learnable. The foundational understanding of how work actually moves through a company is what makes you useful when the next capability shows up and everyone else is guessing what it does.
There is no agent you hand your Salesforce login to. Not yet.
Check out the full RevOps Masterclass: The GTM Stack Hyper-Scalers like OPenAI & Anthropic Use
