A five-person fleet-software startup is showing what happens when an AI coding agent stops being confined to engineering. Proaction says Codex now sits across sales, product discovery and day-to-day operations, allowing a non-technical founder to build customer-specific software demonstrations himself while engineers remain focused on production systems.
The results reported in OpenAI’s September 25 customer story are substantial: Proaction estimates that customized Codex demos avoid 40 to 60 hours of engineering work each month, while co-founder and COO Colin Knudsen estimates Codex saves another 25 to 33 hours of his own time. OpenAI also reports a 60% increase in sales, although the underlying case study describes the more specific metric as an estimated 50% to 60% increase in the share of deals progressing from initial contact into solution development rather than nurture.
The case is therefore less about a generic AI productivity gain than about collapsing boundaries between selling, prototyping and product specification. Proaction is using Codex to turn customer conversations directly into interactive software, then handing the validated result to engineers as a clearer description of what production needs to become.
Fleet management is difficult to sell with a generic demo
Proaction builds software for businesses operating fleets that can include cars, trucks, trailers, refrigerated vehicles and construction machinery. Those operations vary significantly from customer to customer, which makes the fit between software and workflow difficult to communicate through a standard product walkthrough.
Historically, personalized demos created a resource problem. The founders could describe what Proaction might do using conversations and slides, but building an interactive experience around a prospect’s actual operation required engineering time the startup could not spare.
Codex changed who could perform that work. Knudsen, who describes himself as non-technical, now creates four to six customized interactive demos per month without requiring an engineer to build each one. OpenAI says each takes roughly 30 to 45 minutes.
A sales conversation becomes a working software prototype
After a customer call, Knudsen can point Codex at the Granola recording, relevant email threads and spreadsheets the prospect has shared. Codex uses that material to customize an HTML demo environment that mirrors Proaction’s product while reflecting the prospect’s own fleet and operating model.
The effect is more concrete than conventional personalization. When Proaction shares the demo, a prospective customer can see its own types of vehicles and workflows represented in the software, identify what should change and participate in shaping the solution.
OpenAI Academy previously documented an example in which Knudsen could ask a prospect how a roadside-service issue should be reported and then build that workflow live with Codex. A form could request a license plate, ask about the problem, collect a photo, determine whether the vehicle remained operable and send the resulting information into the dashboard.
The sales demo therefore doubles as product discovery. Instead of translating a conversation into notes and then translating those notes again into engineering requirements, Proaction can validate an interactive representation with the customer before production development begins.
OpenAI reports a 50% to 60% improvement at a specific sales stage
The headline says Proaction boosted sales 60%, but the body of OpenAI’s case study provides a more precise interpretation. Knudsen estimates that the percentage of deals moving from initial contact into solution development, rather than being moved into nurture, increased by 50% to 60% after Proaction began using customized demos.
That distinction matters. OpenAI does not publish a controlled experiment demonstrating that total company revenue increased exactly 60% because of Codex. The reported result is Proaction’s estimate of improved progression through a particular part of its sales funnel.
The mechanism is nevertheless easy to understand. A prospect no longer has to imagine how generic software could adapt to its fleet. The customized environment turns that hypothetical fit into something the buyer can inspect and modify.
Four to six demos can avoid 40 to 60 engineering hours every month
The largest directly quantified operational saving comes from work that engineers no longer need to perform. Knudsen estimates that an engineering-built version of a comparable demo would take about ten hours. At four to six demos per month, that represents approximately 40 to 60 engineering hours avoided.
For a five-person startup, the opportunity cost is significant. Every hour spent creating disposable sales prototypes is an hour unavailable for production reliability, customer-facing features and the core platform.
Codex does not eliminate engineering from the process. Proaction deliberately separates the sandboxed demo from secure production work. Engineers still build the production implementation after the workflow has been validated.
What changes is the quality of the specification they receive.
The prototype survives the sale as an engineering reference
Once a prospect becomes a customer, the tailored demo does not simply disappear. Knudsen hands the customized environment to engineering as a visual reference, reducing the questions and back-and-forth normally required to convert sales discovery into a build plan.
Proaction has extended the approach into a customer solution center where prospects can log in, explore workflows built around their businesses and review sales materials. Non-engineering teammates can use those interactions to capture clearer requirements before engineers become involved.
This creates an unusually direct chain from conversation to prototype to specification. The artifact produced during selling becomes useful input to delivery.
Codex is also becoming the founder’s operating environment
Knudsen’s use of Codex extends well beyond demos. Proaction connects customer and company context through Codex plugins for Granola, Gmail, Slack, Linear, GitHub and HubSpot. From one environment, he can retrieve call transcripts and email history, prepare follow-ups, create Linear issues and update HubSpot opportunities.
He has also configured a scheduled automation that reviews recent calls and prepares sales updates for the team. OpenAI says Knudsen estimates he handles 15 to 20 distinct tasks each day and saves 25 to 33 hours per month through Codex.
Combined with the 40 to 60 hours of engineering effort avoided by the demo workflow, Proaction’s reported monthly savings exceed 65 hours at the low end and reach 93 hours at the high end. The “75+ hours” framing is therefore directionally supported by the two estimates, although the components measure different people and workflows rather than a single audited time study.
This is the broader Codex shift from coding assistant to work agent
Proaction’s workflow fits a wider change OpenAI has been documenting throughout 2026. In June, the company reported that knowledge workers had become roughly 20% of Codex users and were growing more than three times as fast as developers. Their work increasingly included research, analysis, workflow automation, reports and lightweight tools rather than conventional software engineering alone.
OpenAI’s later economic research describes the same movement in terms of task duration. By May 2026, more than 70% of sampled individual Codex users had made at least one request estimated to represent more than an hour of human work, while users increasingly orchestrated multiple agent tasks in parallel.
Proaction makes that abstract trend tangible. A COO can use the same agentic environment to prepare a sales interaction, create working software, update the CRM and hand structured work to engineering.
The customer-facing product is becoming agentic too
Codex changes how Proaction builds and sells software, but OpenAI models are also changing what Proaction’s customers can buy. The company is developing what it calls a Managed Execution Layer: specialized agents that perform operational fleet work instead of merely recording and displaying it.
OpenAI says GPT-Live-1 and GPT-6 Astra are among the models powering these agents. They can make voice calls, review documents and images, analyze text and communicate through chat. Proaction also uses ChatGPT-5.6 Sol to help identify damage in photos submitted with vehicle issue reports.
This moves the fleet-management product from a system of record toward a system that can execute portions of the operational workflow.
“Marty” shows what execution means in practice
One Proaction agent, Marty, is designed around vehicle maintenance. According to OpenAI, Marty can talk with a driver about a problem, call repair shops, coordinate service and help move an estimate through approval and payment. Proaction’s human team remains available when the workflow requires review or intervention.
The example is important because fleet operations contain many tasks that are individually simple but operationally expensive: phone calls, status checks, document handling, scheduling, approvals and coordination among drivers, shops and fleet managers.
A voice-capable agent does not need to replace the fleet-management system. It can become an execution layer on top of the data and processes already inside it.
GPT-6 Astra reduces the length of computer-use work
Proaction is also using GPT-6 Astra as a development accelerator for these experiences. Danny O’Halloran, Proaction’s Head of Product, says Astra’s computer-use runs are more succinct than equivalent runs with GPT-5.6 Sol, requiring a shorter execution path for the same work.
OpenAI does not provide a quantified benchmark for this Proaction comparison, so it should be treated as the company’s reported experience rather than a general performance claim. But it points to an increasingly important agent metric: not merely whether the model can finish a task, but how efficiently it can navigate the sequence of actions required to finish it.
For production agents that operate interfaces repeatedly, shorter and more reliable trajectories can affect latency, cost and the number of opportunities for an execution to go wrong.
AI lets a small startup move work across organizational boundaries
The most consequential aspect of the Proaction story may be organizational rather than technical. The company has not simply made its engineers faster. It has moved selected technical execution into sales and product roles while preserving engineering responsibility for secure production systems.
That changes the handoff structure. Sales can prototype instead of describing. Customers can react to a working interface instead of a slide. Product teams can validate workflow details earlier. Engineers can begin with a concrete reference rather than reconstructing the customer’s intent from notes.
For a five-person company, reducing those translation steps can matter as much as reducing the time required for any single coding task.
The case study is first-party evidence, not an independent audit
The performance figures should be interpreted in context. The source is an OpenAI customer story about a company using OpenAI products, and several key numbers are estimates attributed to Proaction’s co-founder. OpenAI does not publish the raw sales pipeline, time logs or a counterfactual analysis.
That does not make the figures uninformative, but it does limit what can be concluded from them. The case supports the claim that Proaction reports substantial time savings and better deal progression after adopting Codex. It does not establish that every company should expect a 60% sales improvement or more than 75 hours of monthly savings.
Proaction’s real gain is a shorter path from customer intent to execution
Viewed together, Codex, GPT-Live-1 and GPT-6 Astra are serving three different layers of Proaction’s business. Codex helps the team convert customer context into demos, operational actions and clearer requirements. GPT-6 Astra helps build and execute more capable agent experiences. GPT-Live-1 gives those agents a voice channel for work that still happens through phone conversations.
The reported sales and time savings are the visible outcomes. The deeper change is that fewer translations are required between what a customer says, what a salesperson understands, what a prototype demonstrates and what an engineer ultimately builds.
For Proaction, AI is not simply generating more code. It is making the entire path from selling software to operating fleet workflows more executable—and allowing a very small team to carry more of that path without adding another handoff.