Proaction, a company that builds software for businesses managing vehicle fleets, says it has folded OpenAI's Codex into its sales motion and now produces four to six customized, interactive demos a month. Each demo takes 30 to 45 minutes to build, according to co-founder and COO Colin Knudsen, who does the work himself rather than routing requests to engineers.
OpenAI published the account as a customer case study on its own site. The numbers in it are Proaction's own estimates, relayed through OpenAI's write-up, not an independently audited benchmark.
The setup is straightforward. After a sales call, Knudsen feeds Codex three inputs: the Granola call recording, the prospect's email threads, and any spreadsheets that prospect has shared. From that material, Codex assembles an HTML demo environment shaped around Proaction's product and the customer's own fleet, so prospects see their own cars, trucks, or construction equipment organized around how they work.
"As a non-technical person, I used to have to loop engineers in if I wanted a demo," Knudsen said. "Now I do it myself in Codex."
What the case study claims
The headline figures are about time and conversion. Knudsen estimates the custom demos avoid 40 to 60 hours of engineering work a month, based on a rough 10 hours per demo if engineers built a comparable one. He also estimates that the proportion of deals advancing from first contact to solution development, bypassing the nurture stage, has increased by 50% to 60% since the demos were introduced.
Proaction says the demo work carries past the sale. When a prospect becomes a customer, Knudsen hands engineers the customized demo as a visual reference, which the company says reduces back-and-forth about what to build. The team also built a customer solution center where prospects can sign in, browse workflows configured for their operations, and look through sales collateral.
Knudsen's own workload is part of the pitch. He says his day spans sales, customer support, and product management, and that Codex plugins for tools including Granola, Gmail, Slack, Linear, GitHub, and HubSpot let him pull call transcripts and email history, create Linear issues, and update HubSpot opportunities without switching tabs. He estimates 15 to 20 distinct tasks a day and says Codex saves him 25 to 33 hours a month.
Agents move into fleet operations
The case study also covers Proaction's product work. The company says ChatGPT-5.6 Sol can help assess damage when a customer sends in photos alongside a report of a vehicle problem. With GPT-Live-1, Proaction says it is building agents that take on a larger share of the routine operational work involved in running a fleet — what the company calls its Managed Execution Layer.
Customers can ask specialized agents to handle tasks such as tolls or service, or set up workflows that dispatch the right agent automatically. Drawing on OpenAI models, including GPT-Live-1 and GPT-6 Astra, these agents can make phone calls, work through documents and images, process written text, and hold chat conversations.
One agent, Marty, is designed to coordinate vehicle maintenance: it discusses the issue with the driver, phones repair shops, schedules the service, and helps move the estimate through approval and payment. Proaction's team steps in when work needs human review or intervention.
Danny O'Halloran, Proaction's Head of Product, said GPT-6 Astra also lets the company build these experiences faster. "Astra's computer-use runs are more succinct," he said. "With GPT-5.6 Sol, I had a much longer run to execute the same work."
Why it matters
Case studies like this one are how model vendors argue that coding assistants pay off outside engineering. Proaction's claim is narrower and more interesting than a generic productivity pitch: the assistant is being used by a non-technical executive to produce customer-facing artifacts that used to require engineering time. That is a different buyer than the developer already paying for a coding subscription.
It also puts a number on a familiar bottleneck. For small software companies, pre-sales demos are exactly the kind of work that is too custom to schedule and too valuable to skip. If a sales lead can build one in under an hour, the constraint moves from engineering capacity to how many conversations the team can have.
Our take
The interesting part is not the 60% figure — it is that a COO, not an engineer, is shipping the demo. That changes who inside a company is a candidate for a coding assistant, and it is the kind of claim OpenAI needs to make to sell seats beyond developer teams. It is also the kind of claim that stays unverified: the stage-advancement lift depends on deal definitions, pipeline mix, and sample size, and none of that is disclosed here.
