Proaction's OpenAI case study landed quietly, but it's one of the most concrete examples I've seen of how Codex changes the shape of early-stage sales work—not just the speed.
The headline numbers are strong: 50–60% more deals moving from first contact into solution development, 75+ engineering hours saved monthly, and a non-technical founder building four to six fully customized product demos per month. But the interesting part isn't the efficiency gain. It's that Codex made a category of work economically viable that wasn't before.
The demo economics problem
Proaction builds fleet management software. Every prospect—whether they run delivery vans, construction equipment, or corporate car pools—operates differently. Showing them a generic demo is like showing a construction company screenshots of Salesforce and asking them to imagine their workflow.
Custom demos were essential to closing deals, but custom demos required engineering time. Colin Knudsen, Proaction's non-technical co-founder and COO, estimates each engineer-built demo would take about 10 hours. At four to six demos per month, that's 40–60 hours of engineering capacity the team simply didn't have.
So Proaction did what most early-stage companies do: founders ran sales calls with slide decks and hand-waving. "I used to have to loop engineers in if I wanted a demo," Knudsen says. "Now I do it myself in Codex."
What changed: 30 minutes instead of 10 hours
After a sales call, Knudsen feeds Codex the Granola call recording, email threads, and any spreadsheets the prospect shared. Codex builds an HTML demo environment that mirrors Proaction's product with the prospect's actual vehicles and workflows.
Each demo takes him 30 to 45 minutes to build. When he screenshares with the prospect, they see their own trucks and equipment, organized the way they actually work. "You work together to generate the end solution without getting engineering involved at all," Knudsen notes.
This isn't a Figma prototype or a Loom walkthrough. It's an interactive environment where prospects can click around, point to what needs adjusting, and co-create the solution. That concreteness matters: Proaction estimates 50–60% more deals now advance from initial contact into real solution development rather than falling into nurture limbo.
The downstream benefits are better than the time savings
The 40–60 hours of monthly engineering time saved is real, but it's almost secondary to the workflow change.
When a prospect converts to a customer, Knudsen hands engineers the customized demo as a visual spec. That reduces ambiguity and back-and-forth about requirements. Engineers build from a concrete reference, not from meeting notes and memory.
Proaction also built a customer solution center in Codex where prospects log in and explore workflows tailored to their business. Non-engineering team members can turn customer conversations into clearer specs before engineers ever get involved. By the time engineering starts building, the picture is sharper.
This is the interesting part: Codex didn't just make demos faster. It changed who could build them, when they could be built, and what they could be used for downstream. The value isn't in replacing 10 hours with 30 minutes. It's in making a category of work—granular, one-off customization for every serious prospect—economically feasible at seed stage.
Codex as operating system for a generalist founder
Knudsen's daily work spans sales, customer support, and product management. He uses Codex with plugins for Granola, Gmail, Slack, Linear, GitHub, and HubSpot to pull context and act on it without tab-switching.
He pulls call transcripts and email history to prep follow-ups, creates Linear issues, updates HubSpot deals, and set up a scheduled automation that reviews recent calls and drafts sales updates for the team.
"Everything that I do is centered around working in Codex. I don't leave it much," he says. He estimates 15–20 distinct tasks per day, and believes Codex saves him 25–33 hours monthly.
The quote that stuck with me: "I can't really imagine being a startup founder without Codex."
That's not AI hype. That's a founder describing infrastructure.
Voice agents with GPT-Live-1: the "Managed Execution Layer"
Proaction also uses OpenAI models in-product. ChatGPT-5.6 Sol helps identify vehicle damage from customer-submitted photos. But the more ambitious piece is what they call the Managed Execution Layer: voice agents built on GPT-Live-1 and GPT-6 Astra that handle operational work, not just tracking and dashboards.
Customers can ask specialized agents to handle tolls or service scheduling, or set up workflows that trigger the right agent automatically. The agents make voice calls, review documents and images, analyze text, and respond in chat.
One agent, Marty, coordinates vehicle maintenance. Marty talks with a driver about a problem, calls repair shops, arranges service, and helps get estimates approved and paid. Proaction's team steps in when the work needs human review.
"We're building the ability for Proaction to execute work for our customers, beyond helping them manage and track it," Knudsen explains. "The advances in OpenAI voice are a big reason we can do it."
Danny O'Halloran, Proaction's Head of Product, notes that GPT-6 Astra's computer-use capabilities run more efficiently than GPT-5.6 Sol: "Astra's computer-use runs are more succinct. With GPT-5.6 Sol, I had a much longer run to execute the same work."
What this case study actually shows
The Proaction story is useful because it's specific and unglamorous. This isn't a consumer app with viral loops or a research lab pushing benchmarks. It's a B2B SaaS company using Codex and OpenAI models to do ordinary startup work—sales demos, customer support, product specs, operational coordination—in ways that weren't economically viable before.
The pattern I'm watching for: tools that change the economics of previously unscalable work. Not "make X faster," but "make X possible."
Custom demos for every serious prospect? Wasn't feasible at seed stage. Voice agents that call repair shops and negotiate service appointments? Wasn't feasible without a large ops team.
Codex and GPT-Live-1 made both categories of work viable. That's a different kind of leverage than raw speed improvements, and it compounds in ways that are hard to predict from the outside.
The question for other founders isn't "Can AI make my team 20% faster?" It's "What work is currently too expensive or too specialized to do at all?" That's where the unlock is.