SaaS Says: Smart Enough Already

Smart Enough Already

Ali Ghodsi thinks the frontier could stop moving tomorrow and enterprise software would be fine. "the models are smart enough. Um, but they just don't have the context that exists inside of any organization."

That is a product argument, not a safety one. A model has not sat in your meetings, does not know who actually decides things, and has none of what an employee accumulates by staying. "so it's the difference between a person who is a new employee in the company and just started today and a person that has worked there 5 years." His fix is a precomputed graph of the organization, read fast, rather than an agent crawling one system at a time. "For that, we actually don't need smarter models."

Every vendor holding a system of record said a version of this. What changed is that they now argue it on the bill rather than on quality.

Tamar Yehoshua ran the test. Atlassian's chief product and AI officer says coding agents routed through the company's graph of projects, people and decisions rather than a raw tool interface came back better and cheaper: "the quality increased by 44% and very importantly, the token usage went down by 48%." An internal A/B test with no published method. Still the first time anyone has put a token price on institutional memory. "An agent needs the context of an organization."

Keith Peiris walked away from a product with millions of users after deciding it could not hold enough context, and rebuilt on that premise. The Lightfield co-founder on the category he landed in: "the hardest part of doing this work was actually making sense of all of the data across all of these disparate systems." What customers say they will build themselves has moved with it. "We don't hear. I'm going to build my own system of record, we hear I'm going to build my own company brain."

Ann Davis sells the same asset with the software peeled off. Crunchbase's chief revenue officer treats the interface as the part that depreciates. "Doesn't mean it only has to be accessed through your software. You can have a way to license the data within that software application and have another revenue stream." The pressure behind it: "Anyone in the software business that's not seeing a hit to their revenue from AI, I would be surprised."

None of it settles who holds the thing. Per McKinsey research cited on the show by the analyst Peter Buchanan, 32% of organizations have already declined to buy a product because coding agents made building it credible. Hailey McDonald, who runs revenue marketing at Sprout Social, draws the line where most buyers do. "sprout social absolutely is not going to like vibe code a website" and then hand it to a junior to maintain.

Last week the argument was who pays for the tokens. This week it is who owns the thing that decides how many get burned.

Sources: Interviews from a16z (Sep 18, Ali Ghodsi of Databricks; and Sep 16, Keith Peiris of Lightfield), CXOTalk (Sep 16, Tamar Yehoshua of Atlassian), Revenue Builders (Sep 17, Ann Davis of Crunchbase), Metrics that Measure Up (Sep 14, Peter Buchanan of New Plan), and Exit Five (Sep 17, Hailey McDonald of Sprout Social).