Nobody in the room had an AI agent running in production

This May, we sat down with 60 senior leaders across 50 Australian businesses at our executive roundtables and discovered not one of them had an AI agent running in production. Not one. A single company had built something, but it was not yet in production and instead presided with the CTO rather than serving the business.

This is the theme we're seeing when it comes to the real adoption of AI in medium sized businesses.

If you spend any time on LinkedIn, you'd be forgiven for thinking agents are already running half the economy. The version of 2026 we heard in those rooms, from the people actually accountable for delivery, looked nothing like that.

What people actually told us

We ran the sessions across Sydney, Melbourne, and a national virtual room for the leaders who couldn't make it in person, and at each one we asked everyone to place themselves on a simple model we use at Adaca: Crawl, Walk, Run. Think of it as a quick AI readiness assessment that the room runs on itself. No marking, no audit of their business, just a show of hands for where they felt they sat on the curve of AI maturity.

Crawl: The Foundations Stage
This means a baseline understanding of LLMs across the business plus a commercial agreement with a provider, so staff aren't quietly pasting company data into a consumer chatbot. A striking 45 of the 50 companies placed themselves here.

Walk: The Governance Stage
This means real AI governance, with proper risk management and some way of seeing the shadow AI that's already in the building. Only 4 of 50 felt they had reached this level of governance.

Run: The Building Stage
This means agents actually deployed into the business. Exactly 1 company of 50, and as I said, that one was still in testing and owned by a single person.

That distribution matters, because nine in ten of the businesses in our rooms, by their own reflection, are still standing at the very beginning of the journey.

The gap nobody is naming

The gap that genuinely surprised me is that the most technical people in these companies, the CTOs and CIOs, are absolutely flying. They are using tools like Claude Code to build the applications they had only ever dreamed about, the systems that used to demand a full quarter and a budget fight, and they're shipping them in days. Their enthusiasm in the room was real and completely understandable, because for someone who has spent an entire career being told there aren't enough resources, this moment genuinely feels like a gift.

But if you looked one desk over, almost nothing had moved, and the same pattern surfaced in company after company:

The engineering function is effectively at "run" for its own work, automating its own corner at pace.

Accounts, operations, and the rest of the team are sitting on a single Copilot licence, and sometimes on nothing at all.

Many of those same people use ChatGPT and Claude confidently in their personal lives, yet have no sanctioned way to touch them at work.

So the adoption curve inside these organisations isn't moving forward as one. It's pulling apart, and the distance between the early adopters and everyone else, all under the same roof, is widening rather than closing.

That is the real risk almost nobody is naming out loud. When your most technical people automate their own work while the operational core of the company stays exactly where it was, you don't end up with a transformed business. You end up with a brilliant engineering team, a company that hasn't actually changed, and a steadily growing pile of shadow AI as frustrated staff reach for personal tools to fill the gap themselves.

Why everyone skips "crawl"

Companies skip crawl because they try to leap straight to building agents, and the rooms made it very clear why that keeps failing. The companies stuck at crawl were not short on ambition. They were short on the unglamorous foundations that nobody posts about:

A clear acceptable-use policy that people can actually follow.

A commercial agreement with a model provider.

Basic AI literacy reaching beyond the technical team.

Any reliable way of seeing where AI is already being used.

AI governance was in a similar state. A handful of companies had a dedicated AI policy, while many more were quietly leaning on their existing data policies and hoping they would stretch far enough to cover the gap. They rarely do.

This is the real state of AI in the Australian mid-market in 2026, and it comes down to three things sitting awkwardly together:

Genuinely high enthusiasm across the board.

Real capability, but concentrated in pockets rather than spread across the business.

Almost no organisational readiness underneath any of it.

The agents are coming, of that I have no doubt. But the companies that win won't be the ones that deployed first; they'll be the ones that did the patient work of laying the foundations, so the rest of the business could follow the technologists instead of being left behind by them.

If you read all of this and recognise your own company, racing ahead in engineering while stalling almost everywhere else, I'd gently point out that you are not behind. You are sitting exactly where 45 of our 50 businesses placed themselves. The work now isn't to chase an agent into production. It's to run an AI readiness assessment of your own, close the gap that's quietly opening inside your own walls, and bring your AI maturity up across the whole business rather than in one corner of it.

So let me put the question back to you: in your business, how far ahead of everyone else have your technologists really pulled, and who is doing anything about it?

A quick note on the numbers. These figures are self-reported. We asked people where they sat on Crawl, Walk, Run, and we deliberately did not audit their businesses to verify it. That choice was intentional, because this is a snapshot of how Australian leaders genuinely see their own AI maturity, and that perception is often the more revealing number anyway.