A vendor with a few hundred million users picked our agency as an example worth highlighting. I had to sit with that for a while before it felt real.
OpenAI for Business showcased RAB2B this week, and I'm not going to play it cool. There's a specific feeling that comes with years of quiet, hard work suddenly being held up in public. It's not pride, actually. It's much closer to relief. The kind you feel when something you believed in private turns out to be real enough for someone else to see from the outside.
That started 3 years ago, when the path forward was genuinely unknown and every fork mattered. Three of them, in particular. The decision we chose is marked with an asterisk.
None of those had a clean answer. So we did the only things you can do when there's nothing concrete to stand on: bring as much data to the decision as possible, research it as thoroughly as possible, make the best choice you can, and be willing to be wrong, with potentially huge impact.
Here's what the OpenAI post captures: 100% of our people on ChatGPT Business licenses. 100% adoption. Proficiency measured, not assumed. The everyday work got faster: creative production by 56%, landing page deployment by 68%, competitive research by 44%, quality assurance by 61% and those are just a few examples. 41% overall efficiency gain across the agency, and the equivalent capacity of 30 new hires with zero added to payroll.1
What it can't capture is that none of those numbers were the plan. They were the residue of a hundred smaller decisions, most of which felt too small to matter when we made them.
An easy bet was to lean on the AI already baked into established SaaS tools, letting the vendors bolt the intelligence on for us. But I've yet to see an AI add-on I'd call genuinely effective. Most are features in search of a problem: impressive in the release notes, hollow the moment real work touches them. Choosing the right platform and getting it into every person's hands was real work in itself, and we treated it that way.
Instead, we did it the hard way. That meant a rule we kept even when it cost us: we never started with the technology. We started with a problem someone in the business actually had: a real bottleneck in daily work, with a number attached. Then we asked the question almost nobody asks in the middle of a gold rush. "Does this even need AI?" Sometimes the honest answer was no. A better process, a spreadsheet, or a human making one faster call. We let those answers stand, and we kept a person's hands on the wheel of everything that survived.
It was the less certain path and definitely the unsexy one. There was nothing flashy to point to. Just ChatGPT coverage rolled out across the whole company, weekly trainings built not on "prompt engineering" but on the work itself, and a steady stream of CustomGPTs we stood up to test one opportunity after another.
Most teams run it the other way. Buy the impressive capability first, then go hunting for a problem big enough to justify it. It's the more comforting way to work. It feels like progress. It produces demos that get a round of applause in a meeting and are never opened again.
Some of you know I have a beloved freediving past. My twenties were spent chasing depth on a single breath, which led me to world records. It taught me to distrust that instinct, as many other sports do. There was always a temptation to believe the gear was the answer. A better fin, a hypoxic training machine, a more streamlined suit. But the right way to win was the disciplined training of my body to do more with less.
Growing our discipline with AI was our breath-hold. Every build had to begin with a genuine problem, stay in human hands, and prove itself in daily use as a bare CustomGPT before it earned a single line of engineering. No graduation to anything heavier until people reached for it when no one was making them. Most of what we tried, we let die in that bare form, before it ever cost us a development cycle. That restraint was the part that couldn't be bought, and it was the whole point.
But here's what I didn't see coming. The discipline didn't just filter out weak ideas. It built the foundation for everything that followed. Once a workflow survived that test, we knew exactly what it needed to become. Some stayed as CustomGPTs because that shape was right. Others grew into agentic workflows, into skills, into infrastructure that now runs across every team. The thing has grown into something far larger and more capable than the CustomGPTs people see from the outside, and it grew fast because we weren't guessing anymore. Every new build inherits the context the last one earned. We're not at the start of this. We're at the point where it compounds.
None of it happens without the people who made room for it. I owned the vision, but a vision is just a nicely worded intention until someone clears the road in front of it. Boris Rabinovici gave us the resources and patience to do this the slow, correct way instead of the fast, impressive way. Alexandra Roux carried it through the parts I couldn't, getting teams to engage and removing every roadblock before it became an excuse. One person holds the why. Another makes the daily yes possible. And it took the whole company choosing this path: every person deciding to lean into the unknown rather than wait it out, to ask what they could do with this rather than why it couldn't be done. That collective yes is the part no strategy can fake. I've stopped believing execution looks like anything else.
Being seen by OpenAI is the kind of thing you'd think would feel like an arrival. It doesn't. It feels like a checkpoint on a road we're still very much walking. I know what it feels like to come back to the surface after another deep breath-hold dive, and this isn't it. It's the moment when you realize you can go deeper than you thought, and the only question left is "what do we do next?"
1: The figures are before-and-after deltas on the same work at the same scope, measured from time tracked in Jira, not estimates. One RFP workflow's 75% reduction (the example OpenAI showcased), for example, came from ~25 pre-pilot RFPs over a year against ~13 post-pilot, with tasks structured to capture effort by RFP size and scope. We held brand compliance and every other AI-enabled workflow to the same standard.