Lore

AI agents in product management

Inside Ramp, the $32B Company Where AI Agents Run Everything | Geoff Charles

Ramp's CPO Geoff (called "Jeff" in the transcript) argues that AI agents like Claude Code have driven the cost of building software toward zero, so Ramp now lets every function—not just engineering—write prompts/specs that AI agents execute into real, shippable product, and this forces the PM role itself to split into hands-on builders and business/GM-strategy owners or become obsolete.

Peter Yang · 2026-03-15 · English

Key ideas

  1. AI has driven the cost of code and analysis toward zero, which requires PMs to write specs/prompts for AI agents to read and execute rather than for human engineers, collapsing the old 'spec then build' pipeline into a 'prompt to product, back to prompt, back to product' loop.

  2. Ramp deliberately gives unrestricted, company-wide access to AI tools across every function rather than mandating one tool, betting that visible 'aha moments' and open sharing drive faster adoption than top-down tooling mandates.

  3. Because AI agents can now do end-to-end product work—research, spec, build, release, and even enablement writing—with real codebase and design-system context, output from PMs, designers, and salespeople is genuine production code and documentation, not throwaway prototypes.

  4. Since building is no longer the bottleneck, planning has compressed from roughly 3 years to about 3 months, and leadership's job shifts from directing individual execution to diagnosing systemic breakdowns in a prompt, skill, or design system so fixes generalize instead of being one-off.

  5. The traditional PM value proposition (best spec-writer, best at stakeholder management, best at understanding customers/data) is argued to be obsolete now that 'code is free,' and the function is bifurcating into hands-on builders who ship directly in the product and GM/strategy roles who own positioning, competition, distribution, and monetization.

  6. Ramp operationalizes this shift org-wide through a formal AI-proficiency ladder (L0-L3), an AI-fluency hiring bar, and cultural mechanisms (office hours, showcases, usage tracking) meant to push the whole company up the ladder, on the premise that self-starters can be elevated but low performers likely won't remain.

  7. Because the end-state of AI-native software is an agent doing the user's job rather than a UI for the user to do it themselves, builders must personally use AI co-workers to understand that their own products should become co-workers too, and requiring a login/UI at all may become a design failure.

  8. Ramp frames its generous, seemingly inefficient investment in AI tool access as a deliberate bet to exploit a limited competitive window before AI can 'one-shot' a full platform like Ramp's, rather than as a cost-cutting exercise.

  9. Claude Code — Anthropic's coding agent, given company-wide access at Ramp rather than being restricted to engineers, described as 'a product for builders.'. Apply: Any employee, not just an engineer, uses it directly to build or modify real product features instead of routing the work through an engineer.

  10. Inspect — Ramp's internal AI agent tool that plans, understands the existing codebase, and implements full-stack (front-end plus back-end) product changes using Ramp's existing design component library. Apply: A PM writes a short spec describing a dashboard or feature and Inspect produces a real, mergeable pull request that reuses existing components and endpoints.

  11. Voice of the Customer agent — An agent that replaces the former human 'voice of customer' role by mining Gong calls, Salesforce notes, support tickets, in-app chats, surveys, Snowflake data, and analytics to answer PM research questions. Apply: Ask it a research question such as recent procurement-product feedback from support tickets, then follow up conversationally for quotes, session recordings, or a drafted outreach email.

  12. Ramp Research — An earlier internal chatbot for querying Ramp's database in natural language, launched about six months before the interview and already considered outdated. Apply: Teams such as sales, support, or marketing asked it direct data questions (e.g., campaign open rates, customers by region) and got answers in minutes instead of filing a request with a data analyst.

  13. Data analyst skill (Snowflake CLI + Claude + Skills) — A Claude skill encoding full knowledge of Ramp's database schema and analytics best practices, used with the Snowflake CLI to replace Ramp Research. Apply: State a broad goal, such as building a performance report on a product with growth ideas, and Claude generates a grounded, actionable HTML report from real data.

  14. Product spec skill — A conversational Claude skill built by a Ramp PM that acts like a peer reviewer, asking clarifying questions about goals, trade-offs, and overlaps with other initiatives using Notion context. Apply: A PM drafts a rough spec and iterates with the skill's follow-up questions until it converges on a refined spec instead of relying on a human manager review.

  15. Product shaping skill — A Claude skill oriented toward defining the problem, researching competitors and customer evidence, and synthesizing findings into forcing decisions before a spec is written. Apply: Feed it an early problem statement so it pushes for simplicity, surfaces trade-offs, and returns a synthesis with decision-forcing questions.

  16. Goal-oriented prompting — Geoff's stated principle that users should tell AI their end goal rather than a narrow question, since the AI can generate the right sub-questions itself. Apply: Instead of asking one data question, state the outcome you want and let the AI decide what to investigate to get there.

  17. Automated, complexity-aware PR review — Ramp's process for auto-approving a double-digit percentage of pull requests and routing others for review based on an assessment of the change's complexity and scope. Apply: An AI-generated PR is automatically reviewed and either merged directly for low-complexity, quality-of-life changes or escalated to the appropriate human reviewer.

  18. Staged release process — A four-stage rollout consisting of dogfooding, alpha with a research-group of customers, a beta opt-in tier of roughly 10% of the customer base, and a fully automated review gate before GA for large or material features. Apply: Ship a change internally first, then to a small customer research group, then to the beta cohort, before it clears the automated gate for full release.

  19. "Ramp releases" bot — An automated bot that compiles release reports (product preview, Snowflake impact data, Slack work summaries) and drafts help-center articles, internal enablement docs, and announcement posts. Apply: After a feature ships, trigger the bot to generate documentation and communications artifacts automatically instead of a PM writing them by hand.

  20. Process-level root-causing of feedback — A leadership practice of diagnosing which prompt, skill, or design-system component failed when a quality issue recurs, rather than giving one-off feedback to an individual. Apply: When a design mistake such as a misplaced call-to-action appears, fix the shared design system or skill so it is structurally prevented going forward, instead of repeating verbal feedback to each person.

  21. Automated Figma "create process" — An automated design-generation process in Figma with established UX principles, such as placing calls-to-action above the fold based on years of A/B testing, baked in by default. Apply: When a new design is generated through this process, it automatically complies with the baked-in principle rather than relying on a reviewer to catch violations.

  22. Three-pillar planning framework — Geoff's framework for what a compressed roughly-3-month planning cycle must still accomplish: aligning on strategic trade-offs, securing team commitment/accountability, and giving sales a forward-looking baseline. Apply: Structure a shortened planning cycle around these three objectives instead of skipping strategic alignment for the sake of speed.

  23. Automated sales enablement generation — A process that turns a team's finished Notion backlog or plan directly into on-brand one-pagers and slide content for the sales organization. Apply: Once a roadmap is finalized in Notion, trigger the automated generation instead of assigning a person to build sales collateral by hand.

  24. AI proficiency levels (L0-L3) — Ramp's internal maturity ladder for AI adoption: L0 is occasional ChatGPT use, L1 builds custom GPTs/agents and uses Claude Code for some tasks, L2 has built an app automating part of their job and can commit code or feedback, and L3 builds fundamental systems that influence the rest of the org. Apply: Use the levels to diagnose where an employee currently sits and target them toward the next level, with L3s expected to spread practices org-wide.

  25. Internal AI usage tracking tool — A vibe-coded internal dashboard that tracks token usage per employee across Notion AI, ChatGPT, Claude Code, Claude Coworker, and internal Inspect tools. Apply: Leadership consults it to see who is pushing AI adoption forward and who may need intervention or coaching.

  26. Hands-on "build a product" interview — A dedicated interview stage for Ramp's product team requiring candidates to present a product or prototype they personally built and explain how and why it works. Apply: Require candidates to bring a working build to the interview as proof of AI-tool fluency, rather than only discussing product sense abstractly.

  27. GM mindset — Geoff's proposed reorientation of the PM role away from spec-writing toward business-strategy ownership — competition, positioning, distribution, monetization — since AI already handles execution. Apply: A PM shifts time from writing detailed specs to answering strategic questions, such as how to win long-term against competitors, in the manner of a general manager.

Insights

With 'infinite coders' now available via AI, the real constraint moves to human decision speed and PM throughput rather than engineering capacity: 'if you have infinite coders at your disposal, you are actually the bottleneck.'

Making data-querying easier is claimed to increase the number of people asking questions rather than just speeding up existing questions, making the company more data-centric as a side effect.

The token-spend-versus-salary comparison reframes AI compute cost as a headcount cost rather than a tool budget line: if an agent does 10x a person's work, paying more in tokens than that person's salary is treated as reasonable, not wasteful.

Ramp explicitly tracks and tries to reduce the time users spend inside its own product, which the source frames as the opposite of the engagement-maximizing playbook most PMs are trained on at ad-funded companies like Facebook and Netflix.

Sales enablement material, release reports, and help-center docs are described as generated automatically as a byproduct of the shipping pipeline itself, rather than as separate deliverables a PM or writer owns.

Geoff singles out Ramp's own highest-performing PMs as the biggest adoption risk, because their pre-AI approach has already proven successful for their careers, making them least likely to change methods.

Coding is treated as the leading indicator for other white-collar automation: because it was the first knowledge-work function substantially automated by agents, its trajectory is used as the evidence for what is coming next for PMs, designers, and other roles.

The staged rollout (dogfood → alpha → beta ~10% → GA) is paired with automated Snowflake-based impact analytics inside the release bot, effectively folding what used to be a separate experimentation/reporting function directly into the release pipeline.

«If you're not using cloth code this year, no matter what your role is, you're probably underperforming compared to others on the company.»

— 00:00

«My job is to automate my job. And our all our jobs is to automate our jobs.»

— 00:34

«At the end of the day it's just prompt to product, back to prompt, back to product.»

— 02:56

«This is like this was done in, you know, from 38 to 40 so, about 8 minutes... something that would have taken 8 days for a human to actually do across the entire volume.»

— 05:47

«You should just tell AI what your goal is. And you'll actually be surprised at the questions that the AI can actually ask themselves to get to the goal.»

— 10:33

«We've hit coding escape velocity and it's a brave new world out there.»

— 15:19

«If you have infinite coders at your disposal, you are actually the bottleneck and you actually need to start moving faster.»

— 16:10

«This is — this is the real product. Back-end, front-end.»

— 18:07

«You want to figure out within the process what broke down and fix that process so the next time you never have that feedback again.»

— 24:31

«You no longer have to write the spec anymore. You need to actually like be in the product itself.»

— 30:47

«We do not have committees and we do not have sign-offs. You just need to prove that you've added value, and then you can go go for the raises.»

— 32:19

«if you have agents that are able to do 10 times more work than you, then why would you not pay them twice as much as you?»

— 38:14

«management is probably dead... now is not the time to build that skill set.»

— 39:21

«fundamentally software is dead. It's all going to be like co-workers.»

— 41:13

«a login in your product in the future, um I think it's going to be a failure, right?»

— 42:33

«Most of the things that you saw here are things that we built in the last months.»

— 43:48

This is a first-person, insider case study from a well-funded, recently-raised company with strong incentive to showcase transformation, so its efficiency and adoption claims should be read as self-reported rather than independently verified; Geoff himself concedes most of what's shown was built in the prior few months and expects it to be outdated soon.

44:02

↳ Peter Yang · YouTube

Watch original