# Basis, Clay and Exa turn onboarding, sales and integrations into agent workflows > Source: > Published: 2026-09-02 01:18:09+00:00 # Basis, Clay and Exa turn onboarding, sales and integrations into agent workflows **Basis, Clay and Exa Labs show three ways founders are turning bounded internal processes into agent workflows, from onboarding and account triage to tested developer integrations.** By [RuntimeWire Staff](/author/runtimewire-staff) ยท Published Primary source: [OpenAI](https://openai.com/index/ai-native-company-workflows/) ## Why it matters Basis, Clay and Exa are turning founder knowledge into repeatable software. Their results suggest agent adoption will depend on bounded workflows, measurable outcomes and review gates instead of broad mandates to "use AI." [Matthew Harpe](https://www.getbasis.ai/blogs/basis-raises-100m-series-b-led-by-accel-and-google-ventures?ref=runtimewire) and [Mitchell Troyanovsky](https://www.getbasis.ai/blogs/building-a-company-for-the-agi-era?ref=runtimewire) founded Basis to turn the repetitive, structured work of accounting into jobs for AI agents. They are applying the same premise inside Basis, where a Codex workflow now handles much of a new employee's first-day setup. Basis is one of three founder-led examples in [OpenAI's September 1 case study](https://openai.com/index/ai-native-company-workflows/?ref=runtimewire), alongside go-to-market software maker Clay and [AI search infrastructure company Exa Labs](https://exa.ai/?ref=runtimewire). Each has taken an internal process that previously depended on employees gathering context, following a sequence and handing work across functions, then encoded much of that process into an agent. The workflows cover three different stages of building a technology business. Basis uses Codex for onboarding. Clay assigns persistent subagents to sales accounts. Exa uses Codex to identify developer-integration opportunities and carry selected ones as far as a tested pull request. That selection is useful for OpenAI. The article is an enterprise adoption pitch for Codex and OpenAI's broader agent products, built around customers whose founders already have strong reasons to make agents work. The performance figures remain self-reported by OpenAI and the featured businesses, without independent validation. ### Basis turns a founder thesis inward Harpe and Troyanovsky started [Basis](https://www.getbasis.ai/?ref=runtimewire) in 2023 around a specific view of agent adoption: accounting is structured and economically important, yet much of the work remains manual. Harpe previously worked on machine-learning strategy at Boston Consulting Group, while Troyanovsky held product roles at Goldman Sachs and fintech startup Puzzle. That background shaped Basis around longer-running accounting work rather than a general-purpose assistant. Basis [said on February 24, 2026, that it raised a $100 million Series B at a $1.15 billion valuation](https://www.getbasis.ai/blogs/basis-raises-100m-series-b-led-by-accel-and-google-ventures?ref=runtimewire), led by Accel with participation from GV, Khosla Ventures, Lloyd Blankfein and angel investors. Inside Basis, the same preference for defined workflows now governs onboarding. New employees receive Codex and a Basis-specific onboarding skill containing instructions, company resources and setup steps. Codex introduces internal concepts and operates the employee's computer to configure integrations in the background. [OpenAI says](https://openai.com/index/ai-native-company-workflows/?ref=runtimewire) the process reduced first-day onboarding from two hours to 30 minutes. The more durable benefit may be the mechanism for revising it: HR can add recurring questions and exceptions to the skill before the next hire arrives. Knowledge that once lived with an available employee becomes an operating asset that Basis can inspect and edit. The workflow also introduces new hires to Basis through the tool Basis expects them to use. [Troyanovsky argued in a 2025 essay](https://www.getbasis.ai/blogs/building-a-company-for-the-agi-era?ref=runtimewire) that powerful AI will arrive faster than organizations can absorb it. Encoding onboarding is one attempt to close that diffusion gap inside Basis rather than waiting for employees to invent their own habits. ### Clay gives every account a memory Clay's use case tackles a less orderly process. Sales context accumulates across CRM records, calls, presentations, Slack, email and informal conversations. A seller can lose an hour each night reconstructing what changed and deciding which account deserves attention, according to Clay. Founded in 2017 by Kareem Amin and Nicolae Rusan, with [Varun Anand joining as a co-founder in 2021](https://capitalg.com/portfolio/clay/?ref=runtimewire), [Clay](https://www.clay.com/about?ref=runtimewire) describes itself as a self-learning revenue engine for go-to-market teams. Amin previously worked at Microsoft and The Wall Street Journal and co-founded Frame, which was acquired by Sailthru. Clay [said in December 2025 that it had reached $100 million in annual recurring revenue](https://www.clay.com/blog/100m-arr?ref=runtimewire) after growing from $1 million over two years. Clay uses a persistent workspace and dedicated subagent for each account, while a coordinating agent turns overnight updates into a morning list of actions for a seller to review. Those actions can include resolving an unanswered customer question, identifying a missing member of the buying committee or finding a reason to restart a stalled conversation. [Clay says the process saves one go-to-market engineer](https://openai.com/index/ai-native-company-workflows/?ref=runtimewire) about an hour of nightly inbox triage. The recommendations retain links to their supporting evidence so a seller can inspect the underlying material before acting. In [Clay's product roadmap](https://www.clay.com/blog/build-on-clay-clays-product-roadmap-with-ceo-kareem-amin?ref=runtimewire), Amin described the company as a self-improving revenue engine that remembers which go-to-market work succeeded and recommends subsequent actions. Clay's account agents test that thesis against the messy source material and long feedback cycles of enterprise sales. ### Exa uses agents to expand its own distribution Will Bryk and Jeff Wang started Exa in 2021 after meeting at Harvard. Bryk had been an early engineer at Cresta, while Wang worked on data and web infrastructure at Plaid. Exa initially set out to build a higher-quality search engine before pivoting toward an AI-focused search API in early 2023. That pivot placed integrations at the center of Exa's distribution. Every coding agent, research tool or AI application that adds Exa can generate more search volume, but finding and supporting those opportunities requires developers to monitor repositories, research technical context and coordinate with outside maintainers. Exa calls this goal "Exa everywhere." Under the workflow described by OpenAI, Codex monitors sources including Slack and Notion for promising integrations, gathers context, creates pull requests, runs tests and prepares weekly updates. Codex can also draft an announcement for review when an integration is ready. People at Exa still decide which projects warrant attention and which external commitments Exa should make. Tests and review gates sit between generated code and publication. Codex is handling the repeatable path from discovery to a proposed artifact, while Exa retains decisions that involve technical risk or an outside relationship. The workflow arrives after Exa [raised a $250 million Series C on May 20, 2026, at a stated $2.2 billion valuation](https://exa.ai/blog/announcing-series-c?ref=runtimewire). Andreessen Horowitz led the round. [Exa said at the time](https://exa.ai/blog/announcing-series-c?ref=runtimewire) that its API served more than 400,000 developers across over 5,000 organizations. Automating integration discovery is a direct attempt to stretch Exa's developer-relations and engineering capacity as that footprint grows. ### The founders are productizing management work The common thread across the three examples is process ownership. Basis, Clay and Exa did not begin by giving an agent an open-ended instruction to improve operations. Each selected a repeated job, identified the context and tools required, defined where the work stops and kept a person accountable for the outcome. The workflows also sit close to each founder's commercial thesis. Basis packages expert accounting processes and uses the same structure for onboarding. Clay sells programmable revenue workflows and applies persistent account context to its own sellers. Exa supplies search infrastructure to agents and deploys an agent to put that infrastructure in more places. OpenAI's [Enterprise Signals data](https://openai.com/signals/enterprise-data/?ref=runtimewire) says customers in the top 10% of AI usage generated 8.3x as many output tokens per active user as customers near the middle of its distribution in June, compared with a 2.6x gap in January. OpenAI acknowledges that token volume is an imperfect proxy for value. A longer output can consume more compute without producing a better business result. The case studies offer more meaningful measures: minutes removed from onboarding, hours removed from account triage, pull requests tested and exceptions captured for the next run. Those numbers still come from OpenAI and its customers, but they point toward the standard founders will ultimately face. An agent earns broader permissions by completing a bounded process reliably and leaving enough evidence for someone to check its work. Harpe, Troyanovsky, Amin, Rusan, Anand, Bryk and Wang are treating those controls as part of operating design. Their experiments suggest that durable agent workflows will resemble carefully maintained company playbooks that can execute defined tasks under human review.