Build1 distinct publisher3 min readPublished
OpenAI's case study runs Codex through onboarding at Basis, per-account sales subagents at Clay and integration PRs at Exa. All three keep a person at the last inspectable step, and every number comes from the vendor and its customers.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
All three workflows end in something a person can reject cheaply. Exa's ends in a tested pull request [3]. Clay's ends in a morning list whose items keep links back to the material that produced them [11]. Basis's ends in a laptop whose integrations either work or do not [5]. That terminal artifact is the design, and it is doing more work than the model choice.
On a Basis first day, this is what happens. The hire receives Codex plus a Basis-specific onboarding skill holding instructions, company resources and setup steps [4]. Codex then introduces internal concepts and operates the employee's computer to configure integrations in the background [5]. OpenAI puts the before and after at two hours and 30 minutes [6], which is 90 minutes per hire and a 75% reduction [19]. For that to transfer, your two hours have to be mostly deterministic setup rather than judgement, and your integrations have to be drivable by an agent sitting on the machine. The skill also has to exist already. The 30 minutes does not include the afternoon someone spent writing it.
The revision path is the better piece of engineering. HR can add recurring questions and exceptions to the skill before the next hire arrives [7]. That converts onboarding from a habit held by whoever is free into a file with an edit history, so week-two confusion arrives as a defect report. Basis was founded in 2023 on the view that accounting is structured and economically important while much of the work stays manual [16]; this is the same thesis pointed inward.
Clay's input is not sequenced, which makes its shape harder. Sales context accumulates across CRM records, calls, presentations, Slack, email and informal conversation, and Clay says a seller can lose an hour each night reconstructing what changed and deciding which account deserves attention [12]. So each account gets a persistent workspace and a dedicated subagent, and a coordinating agent turns overnight updates into a morning list for the seller to review [8]. The items are concrete, such as an unanswered customer question or a missing member of the buying committee [9]. Now check the arithmetic on the benefit: Clay reports saving one go-to-market engineer about an hour of nightly inbox triage [10], which is the whole of the baseline Clay itself stated [20]. One engineer. The same party supplied both ends of the subtraction.
Exa's version puts a gate at each end. A person selects which developer-integration opportunity is worth pursuing, and Codex carries only the selected ones as far as a tested pull request [3]. The autonomy sits in the middle, between two human decisions.
None of these three are median IT shops. Clay said in December 2025 that it had reached $100 million in annual recurring revenue after growing from $1 million over two years [15], a hundredfold climb [21], and Basis said on February 24, 2026 that it had raised a $100 million Series B at a $1.15 billion valuation led by Accel [14]. runtimewire.com, reading the same case study, describes it as an enterprise adoption pitch built around customers whose founders already have strong reasons to make agents work, with performance figures self-reported and unvalidated [13]. In my context I would copy the skill-file-plus-gate shape and leave the numbers where I found them. What I would want before copying the numbers is a rework rate on the agent's output and a run cost per hire.
Ranked by verification strength, evidence, and original report placement.
OpenAI says the process reduced first-day onboarding at Basis from two hours to 30 minutes.
OpenAI's September 1 case study features three founder-led examples: Basis, go-to-market software maker Clay, and AI search infrastructure company Exa Labs.
Each of the three companies took 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.
Basis uses Codex for onboarding, Clay assigns persistent subagents to sales accounts, and Exa uses Codex to identify developer-integration opportunities and carry selected ones as far as a tested pull request.
New Basis 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.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · September 1, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
product
ChatGPT Work's real ask is your Slack, and somebody has to say yes on everyone's behalf1 distinct publisher
build
Developer habit, priced at $965B: what Anthropic's run actually proves1 distinct publisher
build
OpenAI puts latency on the price list: 750 tokens/sec, gated by workload fit3 distinct publishers
security
OpenAI's Computer History writes a plaintext log of the workday. Decide before staff opt in.1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One vendor document, restated once
Every quantity that matters here — the two hours, the 30 minutes, the hour of nightly triage — originates in OpenAI's September 1 case study or the customers featured in it, and runtimewire.com is the only outlet relaying it. The mechanics are described precisely enough that they could be checked; nothing in this reporting shows anyone has checked them. Exa's workflow arrives with no numbers at all.
Real production use, tiny units
Three named companies describe agents running in live internal processes, not slideware — but the unit of measurement is one new hire's first day, one go-to-market engineer's inbox and one team's integration queue. All three are AI-native startups narrating their own workflows, which is the least demanding population an agent product can be tested against.
Overstated on scale, discounted in the telling
runtimewire.com narrows its own gap by naming the case study an enterprise adoption pitch and marking the numbers as self-reported. What stays inflated is reach: per-person time savings at three companies that sell agent software get carried into a general claim about how agent adoption will work. It also helps the story that Clay's saving happens to equal exactly the baseline Clay supplied.
Everyone in frame sells agents
The document's author sells Codex. The three customers sell agent products — Basis in accounting, Clay in go-to-market, Exa in AI search — and one of them announced a $100 million round at a $1.15 billion valuation six months before this was published. Founders with that exposure are not neutral witnesses to whether agents work, and runtimewire.com says as much, which is more than the source document does.
Descriptions trustworthy, numbers unaudited
We are fairly sure the workflows exist as described — the detail is too specific and too mundane to be invented, and the review gates are the sort of thing a vendor would rather not dwell on. Confidence stops well short of the metrics: one publisher, one vendor document behind it, and no outside party has touched a single figure.