Enterprise AI Is Broken: Gokul Rajaram on Trust, the New Kingmakers and Burn Discipline
Key points
Key takeaways from the The Information TITV interview with Gokul Rajaram, Founding Partner of Marathon Management Partners (September 2026):
Trust is the blocker, not capability. Rajaram uses consumer agents such as Muse and Instinct as a personal assistant for travel, restaurant bookings and errands — but deliberately keeps his email out of reach. He cites a recent Instinct incident where one user's messages leaked into another user's chat with the agent as proof of the stakes: agents cannot yet be trusted with sensitive content, and trust is the single biggest barrier to enterprise adoption.
Models assume trust instead of earning it. The fix is guardrails and progressive autonomy: start with full human review of output and relax it only as reliability is demonstrated. He points to Abridge in medical scribing, which began with doctors approving around 90% of transcriptions and now needs review of only a fraction — an "earn trust" pattern agentic AI should follow.
Transparency and control separate the trustworthy from the rest. Rajaram gives Claude and ChatGPT access to his email only because their connectors are explicit about what they can do, keep drafts in draft mode and separate read access from send rights. Agents must clearly state the controls they offer and be transparent about what happens to data once access is granted.
Enterprises should rent models, not hand over data. In a multi-model world expected to hold roughly ten open-weight and closed-weight models, betting everything on a single lab makes little sense. His prescription for buyers: build routing capability and switch models per task while keeping data with the enterprise — echoing scepticism about trusting the labs with proprietary data.
Fortune 500 sales are no longer the mark of product quality. In his "new kingmakers" argument, selling to companies such as CVS, Walmart or Pfizer can deliver revenue with little genuine signal — those buyers are not yet sophisticated enough to rigorously evaluate a product, so wins can be pilots or one-off relationships. The highest-quality first customers for an AI startup are the AI labs and AI-native companies such as Notion, Granola, Decagon and Sierra, which force products through a competitive gauntlet before adopting them.
Diversify beyond the AI-native echo chamber. For seed and Series A companies, the first two customers should be sophisticated AI-native buyers that push the roadmap; as companies scale they must move outside the Silicon Valley AI ecosystem to non-tech customers while watching customer-concentration risk.
Burn multiple and net revenue retention are the metrics that matter. With ARR multiples "divorced from capital-market reality", Rajaram focuses on how much new revenue a dollar of sales and marketing spend buys, whether customers expand over time, and net revenue retention as the number one measure of business quality.
The growth-at-all-costs era is over. He argues the Rule of 40 should apply to Series C and D AI companies, and estimates that roughly 90% of fast-growing companies burn so heavily they can only raise in a good market. The winners keep growing efficiently and can raise even when the market turns.
Read more: The Information · Watch on YouTube
