Ramp's internal AI agent platform Inspect passes a million sessions

What "internal AI" means at Ramp

Sobel frames internal AI as the natural next step for analytics engineers, not a separate discipline. He argues that analytics engineers have an edge over software engineers in this area. In his view, they are more comfortable with the "black box" behavior of a model, and more willing to write the prose that agents need, which he half-jokingly claims is harder than Python.

The focus is on the company itself as the user of AI. The tools Sobel described are built for Ramp's own staff, from engineers to sales teams, rather than for customers.

The tools behind the numbers

Sobel walked through several internal systems:

• Ramp Research: the company's data agent, launched about a year ago. It became widely shared enough that the dbt Summit keynote referenced it.
• Inspect: Ramp's remote agent platform. It uses a ticket-style interface in which each ticket is a sandboxed coding agent session. It recently passed a million sessions.
• Glass: it began as an internal prototype that Sobel nicknamed after Anthropic's assistant, before teammates renamed it and it went viral. It is now pre-installed on new hires' laptops and pre-connected to internal systems, and it is especially popular with non-engineers such as the sales team.
• Rampplify: an internal app-hosting platform built for shipping internal tools quickly and tearing them down just as quickly.
• WaveDash: the newest piece, a dashboarding tool in which every dashboard is a SQL file in a repo that renders to Markdown. Agents can therefore read and edit dashboards the same way engineers do. Ramp is actively moving off Looker onto it.

Sobel also shared a chart of data-agent query volume. It has grown from roughly 10,000 to 15,000 questions a month to a level that now dwarfs the human-driven queries it is compared against.

Avoiding lock-in and measuring value

Sobel cautioned that Ramp deliberately avoids locking itself into a single AI lab's models. He pointed to swings such as Cognition's reported $45 billion valuation as a reason to keep options open. Cognition makes Devin.

On the question of return on investment, he was candid. Ramp does not measure the ROI of these AI investments rigorously today, he said, calling it "black boxes and vibes."

In the Q&A, Sobel advised small data teams that lack the resources for custom platforms to lean on a design system, a chart library and a "vibe-coded" internal repo rather than building from scratch. On hiring, he said he looks for analytics engineers with some software engineering instincts. He credited Ramp's security engineering team for making the internal AI push possible, and vented a little about AI-generated writing bloating internal documentation.

What it means for data teams

The talk positions analytics engineers as central to company-wide AI adoption, not only to data pipelines. Ramp's examples show agents reading and editing dashboards as files, answering data questions at a growing scale, and running coding sessions in sandboxes. Sobel's own caveats are just as notable: the company has no rigorous ROI measurement yet, and he suggests smaller teams can start with off-the-shelf components instead of custom platforms.

All keynotes and breakout sessions from dbt Summit 2026 are available on demand, according to dbt Labs.