Delphina Raises $7.5M to Let AI Agents Manage Analytics Context

Delphina raised $7.5 million on November 12, 2026, doubling its total funding to $15 million and launching the Context Factory, a system that assigns AI agents the job of building and maintaining the business context behind analytics answers. Radical Ventures and Costanoa co-led the round. Tenacity Capital participated, alongside angel investors including Fei-Fei Li, Nobel laureate Guidon Imbens' fellow laureate Guido Imbens, and early employees of Anthropic and OpenAI.

The company says it will spend the money scaling the Context Factory and expanding its go-to-market and customer-engineering teams. That is a conventional use of proceeds. What is less conventional is the product premise: accurate AI analytics depends on knowing which tables are trusted, how metrics are defined, and what experienced analysts know but have never written down.

Delphina's claim is that most companies assemble this context manually. It then goes stale as the business changes. That gap, not model quality, is the company's target.

What do the agents actually do?

The Context Factory puts agents in charge of maintaining that context. Delphina says they can assemble an initial baseline from existing company data and documents in hours, identify and repair gaps as employees ask questions, and route consequential changes to domain experts for approval. Experts stay responsible for definitions; agents handle the search and updating.

The system, per Delphina's product overview, reads dbt projects, dashboards, documents and query logs, and keeps reading them as they change. Approved changes land in a versioned knowledge base that the customer owns. Evaluation tests can draw on dashboards employees have already verified.

The workflow the company illustrates is concrete. When a question exposes a missing metric definition, the agent searches dbt, Looker and Notion, identifies the owner of the relevant data table, and requests confirmation in Slack. After approval, it publishes the definition, regenerates the original answer, and saves the fix as a regression test. The point, Delphina says, is that approved logic cannot silently regress.

How does Delphina handle enterprise security?

The security architecture matters as much as the agents. Connections use per-user, read-only credentials, and each customer gets a separate data plane. Python and DuckDB run in an isolated Firecracker microVM with no internet access. Prompts, queries and knowledge references are logged and exportable.

Those controls make agent activity inspectable, in Delphina's telling, while restricting what agents can reach inside company systems and where they can send data. The company offers no third-party audit of these claims in its announcement.

Who is using it?

Delphina reports that business teams now ask 73% of the questions on its platform, and that CEOs are active users at more than half of its customers. At Substack, more than 70% of employees use the product, including CEO Chris Best. Other named customers include LATAM Airlines, Upstart, Udemy and Coursera.

Best described the effect in a quote: "Now our teams get accurate answers in seconds instead of waiting weeks in a queue. It's changed how we make decisions." A CEO testimonial is a strong signal, though adoption percentages reported by the vendor remain unaudited.

The founders carry weighty credentials. CEO and cofounder Jeremy Hermann previously built and led Uber's Michelangelo machine-learning platform and cofounded Tecton, which Databricks acquired. President and cofounder Duncan Gilchrist led global pricing data science at Uber. Gilchrist said operations, sales, marketing and finance teams are now Delphina's heaviest users — a shift away from the data teams one might expect.

What is the real test?

The $7.5 million buys Delphina room to expand the product and the teams selling and supporting it. The central question is whether adoption by business users and CEOs can keep growing without weakening the expert approvals, versioned definitions and regression tests on which its promise of trusted analytics depends.