Key Takeaways

  • Ramp CTO Rahul Sengottuvelu launched Inspect, an internal collaborative coding agent that runs directly in Slack and embeds enterprise codebase context.
  • Inspect was engineered to run an actual version of Ramp within the agent environment, letting engineers test changes against live product context.
  • Model leadership switches every few weeks, making single-provider lock-in an operational risk for engineering teams.
  • Sengottuvelu built routing that tracks spend by department, session, token volume, model tier, and reasoning effort to prevent runaway token bills.

Every Coding Agent Eventually Lands on the CFO Desk

Engineering leaders treat AI coding agents as productivity tooling. Finance teams see them as unbudgeted cloud infrastructure. Sengottuvelu puts the dynamic bluntly: “AI is a CFO problem. It will become more of a CFO problem. Every CFO is probably sweating right now. They're seeing their bills go up a lot.”

When developers experiment with reasoning models, costs scale unpredictably. A single engineer testing complex multi-step refactors can burn thousands of tokens in minutes. Because most tooling ties billing to broad API keys rather than individual developer workflows, finance teams only see the aggregate spike weeks later.

To solve this, Ramp designed Inspect to link directly with Ramp Router. The system tracks consumption at the individual session level. Finance teams and engineering leads can inspect which specific department generated the bill, which model handled the request, the exact service tier selected, and the reasoning effort applied to each prompt. Sengottuvelu notes: “There's a lot of complexity that we're trying to expose to finance people.”

The Model Shuffle and Multiplayer Workflows

Committing your engineering workflow to one AI provider is a losing bet. Model benchmarks change almost weekly, and picking a permanent winner is impossible. Sengottuvelu explains the dilemma: “Right now I'm not sure who's winning. And the leading model provider changes very very quickly every few weeks. And so if you had to pick I don't know what the answer is. And so we wanted to build something works across all the tools and you can always switch.”

Inspect addresses this by routing tasks across frontier providers based on task requirements and cost constraints. But switching models is only half the problem. If an agent operates in isolation on a single developer's laptop, team coordination fails and duplicate work multiplies.

Ramp built Inspect around multiplayer interaction inside existing communication channels. “We made it extremely collaborative, multiplayer first, works in Slack,” Sengottuvelu says. By embedding the agent directly where engineers talk, entire teams can view agent threads, review code diffs, and audit context collaboratively. Even more challenging was context depth. “It actually can run a version of ramp inside inspect which is it's an enormous engineering challenge because we just wanted to get have it work with all the stuff we have.” The agent tests logic against real application architecture rather than guessing against isolated files.

What to Do With This

Audit your company's AI spend by department this week. If your frontier model API keys sit in a single pooled account with no session-level tagging for reasoning effort or model tier, add metadata tags to your internal proxy tomorrow. When an engineer runs a reasoning-heavy task, finance should know which team triggered the run before the monthly invoice arrives.