Jasper Polak captured something this week that most AI Ops leaders know but rarely say out loud. At a 75-person consulting firm, he sees an analyst running Claude for market research, a few associates drafting deck sections in Gemini, and a partner who built a private GPT for proposal writing and didn't tell anyone. All real work. None of it changes the firm's throughput, win rate, or margin.
His post was a riff on McKinsey's "Agentic Organization" series, which argues most companies are stuck in pilot mode because the work itself hasn't changed. The tools are fine. The prompting is getting better. The models keep doubling their effective task horizon every four months. And still, none of it shows up on a P&L.
We've been watching this exact pattern for two years. It has a name inside Gadoci, and it's the whole reason the firm exists.
Pilot purgatory is a workflow problem, not a tool problem
When we walk into a 75-person services firm, we usually find the same three things Jasper named. A few power users running Claude or ChatGPT for ad hoc tasks. A couple of teams experimenting with Copilot. Somewhere, a partner or a director has built a private GPT they are quietly proud of and have never shared. These are real pilots. They are also completely disconnected from the firm's operating model.
The mistake most leadership teams make is reading this as a tooling problem. Buy more licenses. Run another training. Try the new model. None of that fixes the underlying issue, which is that the firm's actual workflows, the way a pitch gets built, the way a project gets delivered, the way a proposal gets staffed, have not been touched.
McKinsey's framing lands for us because it's the one we use with clients. Going from pilot to production is not about scaling AI tool adoption. It's about rebuilding one specific workflow agent-first, measuring it against the old one, and rolling it out. That is a different kind of project, and it's the one most firms skip. We wrote about McKinsey's earlier framing here, and the newer piece extends the same argument in a more practical direction.
The step most firms skip
Picking the workflow is the hard part. There are a lot of places you could redesign, and most of them aren't worth it. We spend the first phase of any engagement figuring out which workflow is both high-leverage and tractable. That's the Discovery Discipline, and it's the step where we keep clients from chasing the wrong thing.
The second hard part is rebuilding without breaking. Most consulting firms can't afford to stop and rebuild their proposal process in a clean room. The existing process has to keep running while the new one gets shaped, tested, and validated against the old one's outputs. That's why we send people on-site for the Embed engagement. One of our AI Solutions Engineers sits inside the department for a few days, watches how the work actually happens, and redesigns it while the team keeps delivering. By the end of the engagement, the department has a new version of the workflow that has been shaped in the real environment, not in a deck.
The third hard part is rollout. A redesigned workflow that only works for the engineer who built it isn't a workflow. It's another pilot. Rolling it out means documenting the new process, training the team, tracking what breaks, and iterating. This is where most firms fall apart, because it looks like project management rather than AI work. It isn't glamorous, and it's the only thing that moves the number.
Why the private GPT matters more than you think
The partner with the private GPT deserves its own moment. On the surface, it's a success story. Someone senior, without any central support, figured out how to make AI do useful work for them. The issue is that it's invisible. Nobody knows it exists. Nobody has reviewed what data goes into it. Nobody can reuse it. If the partner leaves, it leaves with them. If the partner uses it to do something risky, nobody notices.
Shadow AI is now the norm at most mid-sized firms we work with. It is not a governance problem to solve by banning tools, which never works. It is a signal that your people are ahead of your operating model. The job is to catch up to them, not to shut them down. That is what an AI Readiness Audit is actually for: finding where your firm is already using AI informally, and turning those pockets into something the whole organization can benefit from.
What this looks like if you do it
Most firms don't skip the redesign step because they don't understand it. They skip it because it's harder than adding more tools. Redesigning the proposal workflow is harder than buying Copilot. Running the Embed is harder than running a 101 training. Rolling out a production workflow is harder than celebrating a successful pilot demo.
The firms that are pulling away right now are the ones that did the harder thing once, saw it work, and are now doing it again. Not ten workflows at once. One workflow, measured, rolled out, then the next. McKinsey is right that this is the paradigm shift. Jasper is right that most firms won't take it. The question for the firms we work with is which side of that line they want to end up on.