Two of the CRM market’s biggest platforms revealed opposite bets on where their AI actually comes from this week. Salesforce trained its own reasoning model from scratch. HubSpot deepened its bet on renting intelligence from OpenAI. The split is not a footnote. It is the industry’s first real test of whether proprietary AI is a durable moat or an expensive detour.
The Builder’s Bet
Salesforce’s answer, announced September 15, is Koa, a CRM reasoning model built by post-training NVIDIA’s Nemotron 3 Super on a synthetic dataset drawn from nearly 27 years of Salesforce deployment history across more than 14 industries. No customer data went into training. The scenarios were generated to mirror real enterprise workflows: deal structures, service case lifecycles, industry-specific approval chains. Salesforce says the result matches or beats leading general-purpose models on CRM tasks with three times fewer errors, though that benchmark is Salesforce’s own and has not been independently replicated, as SalesTech reported when Koa launched.
“The knowledge is put inside the model itself. We trained a reasoning engine that understands deal structure, service case lifecycle, and industry-varying workflows,” said Marc Benioff, Chair and CEO of Salesforce.
Advertisement
“NVIDIA Nemotron open models give Salesforce the foundation to turn decades of enterprise expertise into specialized AI with Koa, creating a CRM model that can reason and securely take action,” said Jensen Huang, Founder and CEO of NVIDIA.
The Renter’s Bet
HubSpot took the opposite route. On September 16 it deepened its partnership with OpenAI rather than build a proprietary model, becoming the first CRM to integrate directly with ChatGPT Ads so businesses can build, manage and measure ad campaigns without leaving HubSpot. It also expanded its existing HubSpot connector for ChatGPT to cover email campaigns, landing pages, closed-won deal analysis and AI search optimization, and rolled out an AI Growth Bundle discounting HubSpot Starter and ChatGPT Business seats together through September 30.
HubSpot’s case for renting rather than building rests on adoption it already has: it says weekly active users of its existing ChatGPT connector grew 250% as sales teams used it to analyze pipeline and track deals from inside conversations they were already having.
“Connecting ChatGPT with the customer knowledge in HubSpot helps teams do more with what they already know,” said Brian Landsman, VP of Global Partnerships at OpenAI.
“With our HubSpot connector for ChatGPT, new ChatGPT ads integration, and new AI Growth Bundle offer, we’re giving every scaling business a real shot,” said Duncan Lennox, Chief Product and Technology Officer at HubSpot.
Why the Split Makes Sense for Both
The two strategies are not really competing on the same axis. Salesforce can justify building because it owns something OpenAI does not: 27 years of enterprise CRM workflow data across a customer base large enough to make a proprietary model economically defensible. That is a moat few competitors can replicate regardless of engineering budget.
HubSpot’s calculation runs the other way. It serves a smaller-business customer whose buying decision turns on speed to value, not model provenance. Partnering with the lab already running ahead on general capability, and wrapping it in CRM context, gets a usable AI feature to market faster than a from-scratch training run would, and its own adoption numbers suggest customers are not waiting to be sold on which approach is more sophisticated. Salesforce itself is not purely a builder either: it also ships job-shaped Agentforce agents built for specific sales roles rather than left as configurable prompts, showing the same platform can build proprietary reasoning while still packaging AI for fast time-to-value.
Salesforce Is Hedging Its Own Bet
Even Salesforce is not purely a builder. The same week it announced Koa, it also expanded Agentforce’s access to Amazon Bedrock’s model marketplace, letting customers choose which third-party model powers a given agent rather than relying on Koa alone. Joshua Stern, Director of GTM Systems at Engine, one of the customers cited in that announcement, framed the appeal in terms sales leaders will recognize: having a broad selection of foundation models available “gives us flexibility to choose the right model for each use case.” That is functionally the same rented-intelligence logic HubSpot is using with OpenAI, running inside the platform that just spent a year building its own model. The two strategies are not mutually exclusive. They are a hedge against betting the entire AI roadmap on one approach turning out to be wrong.
What It Means for the Sales Leader
The practical risk is treating “we built our own model” or “we partner with OpenAI” as a quality signal on its own. Neither claim tells a buyer whether the AI will actually improve close rates or forecast accuracy. Before renewing or shortlisting a CRM AI feature, ask three questions instead of taking the marketing framing at face value: What data trained or grounds the model, and is any of it independently verifiable? What happens to that feature if the underlying model partner changes terms or the in-house model needs retraining? And does the vendor’s own benchmark, like Salesforce’s three-times-fewer-errors claim for Koa, come from an internal test or a third party with no stake in the result.
Buyers evaluating either path should also weight switching cost. A rented-model feature can be swapped if OpenAI’s terms shift, but it also means core AI capability sits partly outside the CRM vendor’s control. A proprietary model locks in differentiation but ties the vendor’s AI roadmap to its own retraining cycle. Neither is automatically the safer bet. The right question is which tradeoff matches how fast your sales org actually needs new AI capability to ship, not which vendor talks about ownership more confidently.
Source: HubSpot
