This is Design At The Most AI-Installed Company: Diego Zaks

Executive summary

Ramp’s AI adoption was presented as an organizational redesign, not simply a rollout of AI licenses. Design lead Diego Zaks described a company-wide effort to help employees build tools and agents, while reserving specialist design time for experiences where quality and customer emotion matter most.

The approach combines strong communication, performance expectations, broad access, hands-on coaching, and tools tailored to Ramp’s own data and workflows.

Business metrics and claims

• 99.5% of Ramp employees reportedly use AI every day.
• 1,500 internal apps were reportedly shipped in six weeks.
• 12% of production code reportedly comes from non-engineers.
• Ramp’s analysis of millions of U.S. businesses found that heavy AI users grew revenue at 27% annually, compared with 3% for companies that did not spend on AI.
• Anthropic reportedly visited Ramp after finding that about 80% of Ramp’s engineers were using Claude Code.
• A three-person research team supports a much broader research effort: design and product teams conducted roughly 150–200 research sessions in one month, with about 1,000 employees enabled to conduct research.
• An account-management task—drafting a meeting follow-up email that had taken 30–40 minutes—was automated in a 40-minute working session.
• In a host-reported survey, designers who heavily used AI rated workflow satisfaction at about 7.5/10, versus about 5/10 for those who did not use it. 37% said they were not using AI. These are survey results, not Ramp operating metrics.
• Zaks cited an AI-forward business in manufacturing or a related trade growing 65% year over year, compared with competitors at 4–8%. The specific industry was unclear in the subtitles.

Frameworks and operating playbooks

Four levels of AI fluency

1. Use AI for answers and everyday assistance.
2. Use LLMs to write code or create prototypes.
3. Build personal tools to support your work.
4. Build agents that create and use tools to perform substantial parts of your job.

Ramp’s goal is to move employees toward the fourth level—not merely provide access to chatbots or coding assistants.

AI adoption playbook

• Communicate repeatedly that AI is a company priority.
• Include AI fluency in performance reviews and discuss employees’ progress.
• Provide real access to tools.
• Pair early adopters with colleagues and help them automate a real, frustrating task together.
• Share successful examples so coworkers can learn from them and adoption spreads through practical results and word of mouth.

Zaks’s takeaway: mandates and access help, but a firsthand “aha” moment makes the value tangible.

Design infrastructure: make everyone a designer

• Build reusable components, standards, and AI-supported workflows so employees can turn ideas into usable, on-brand experiences.
• Let teams build and test low-risk ideas without waiting for a designer to handle every detail.
• Bring designers in when an idea shows promise or when the experience is especially important; focus their effort on making it exceptional.
• Quickly discard weak ideas rather than spending weeks polishing them.

Customer research at scale

• Start with a specific question.
• Use agents to find customers matching the research profile and help structure questions and plans.
• Have the research team coordinate sessions, record conversations, and feed findings into a shared, searchable system.
• Make prior insights available to employees so research can answer later questions as well.
• Follow up with customers to explain how their feedback informed the product.

Tools gained adoption on merit: employees switched when they saw a more effective process working for colleagues.

Fast experimentation, with a higher bar for launch

• In early alpha, test with a handful of friendly customers and learn quickly—even if the work is rough.
• Avoid polishing an unvalidated idea simply to make it look finished.
• Apply a more rigorous quality bar before general availability (GA).

Zaks framed willingness to look imperfect and admit uncertainty as a way to learn faster.

Internal-tool governance through trust and momentum

• Give employees room to build and experiment; let useful tools attract users and collaborators.
• Invest more when a tool demonstrates value and gathers momentum.
• Keep customer-facing products subject to clearer quality and launch processes.

Ramp’s internal approach is deliberately loose; the transcript characterizes it as high-trust and somewhat chaotic.

Product and organizational examples

• Glass: Ramp’s internal desktop AI application, described as similar in concept to Claude Cowork. It gives employees a simple interface to company knowledge, Slack, Google Workspace, and other connected sources, with sign-in and setup handled centrally. The goal is to avoid requiring each employee to learn terminals, configure MCPs, or connect systems individually.
• Ramp Inspect: A coding agent with access to company data in Snowflake, used to identify customers and feature users for research. It improved through repeated use and iteration before gaining traction.
• AI research assistant: Available through Slack and Notion, it helps employees formulate research questions and plans and makes past customer insights more accessible.
• Automating meeting follow-ups: An account-manager supervisor was coached to use Claude directly. Together, they created a workflow that takes a Gong transcript and generates an email summary consistent with the employee’s previous writing.
• Human-feeling agents: Zaks described an AI engineer in Slack that felt more like a colleague than a tool because of small interaction details, such as emoji reactions, revising an initial message, and returning after considering a question. The example illustrates that interaction design matters, not just model capability.
• Codifying design taste: Zaks described building and refining an internal skill for website design through cycles of generated output and feedback. This turned his preferences—such as spacing, typography, and visual grouping—into reusable guidance for producing work that met his standards.
• Delight in internal products: Employees added playful details, including a small terminal game and ornate animated visuals for an internal token-spend briefing. Zaks argued that curiosity and personal investment can produce more memorable work than a purely functional brief.

Leadership and design implications

Ramp set a company goal, about two years before the conversation, to become the world’s most operationally efficient company. Employees were encouraged to question whether recurring tasks and processes were necessary or could be improved. GPT and Claude Code access was extended broadly, including to non-engineering teams, despite concerns about cost.

Zaks described managers as player-coaches: a director might spend roughly 40% of their time as an individual contributor and 60% coaching and supporting others. At Ramp, leadership is tied to craft and the ability to make; managers are expected to demonstrate work, guide people, and help generate new ideas.

He expects AI to increase the amount of work a company can pursue, rather than simply reduce headcount. Greater productivity can let teams explore more products, markets, and revenue opportunities.

As more people become builders, functional strengths still matter. Designers contribute empathy, visual judgment, and sensitivity to how products make people feel; engineers bring technical rigor and scalability; product practitioners bring commercial and selling instincts. The anticipated model is more fluid collaboration around problems.

In Zaks’s view, design increasingly means shaping production work, intelligence, and knowledge—not only producing screens and flows. AI can handle more execution, giving designers room to concentrate on experience, emotional impact, and taste.

His product principle for Ramp is to make the customer’s work disappear where possible: favor a calm, reassuring experience that completes tasks and lets users leave, rather than maximizing notifications and time spent in the product.

Actionable recommendations

• Anchor AI adoption in a meaningful company goal, then make progress visible and discussable.
• Start with a painful, frequent workflow; build alongside the employee who owns it rather than relying only on training sessions.
• Treat agents and internal tools as products: make them easy to access, connected to the right context, and specific to the company’s needs.
• Use shared design and research infrastructure to expand what non-specialists can safely build and investigate.
• Test concepts with customers before investing heavily in polish; set a distinct quality bar for general release.
• Turn successful workflows, research findings, and design standards into reusable skills and shared knowledge.
• Keep human judgment central: automate execution, but preserve ownership of thinking, customer empathy, and product taste.

Presenters and sources

• Diego Zaks, who leads design at Ramp.
• Tommy Geoco, interviewer and host of State of Play.
• Supporting claims and examples cited in the conversation include Ramp’s internal AI and research data, Ramp’s analysis of U.S. businesses, the host’s designer survey, and the account of Anthropic visiting Ramp.