Last year I kept hearing the same two sentences in the same meeting. Everyone is using AI now. Nobody has changed how the work is organised. Both were said with confidence, and neither was measured, so I built a workforce intelligence platform to measure it.
There are two corporate phrases I hate more than anything: "be more strategic" and "use more AI." These phrases reveal a fundamental misunderstanding of what both strategy and AI are.
For the past couple of years everyone has been using AI, and companies aren’t yet seeing the results, because the work around it hasn’t changed.
Verellax reads the hiring record: 24,944 live job specifications from 217 employers across ten markets, as at 14 September 2026; every AI-central role is scored on four tests.
- 01Does the person work with the model?
- 02Can they overrule it?
- 03Do they own its output?
- 04And which decisions stay with them?
What the data shows is a distance, and the distance has a direction. Adoption arrives at the top of an organisation and stops before it reaches role design.
Three in four roles that mention AI say nothing about how decisions are divided between the person and the system: 75.6 per cent of 12,547 live specifications where AI is core, supporting, or peripheral to the work. This is not a forecast. It is the hiring record, written by the companies and hiring managers themselves, with a date on every line.
I am not the first to say that roles lag tools. What I can do that the last person couldn’t is measure it.
AI has fundamentally changed knowledge work, and tasks have been redistributed from humans to machines.
AI isn't transforming work by wiping out whole jobs. It's changing work at the task level. Cognitive work that used to span different roles is being automated, accelerated, or reshaped, so most jobs won't vanish outright, but the composition of each role is changing.
The record shows it from the bottom up. Execution work, the layer a model absorbs first, is 65.1 per cent of an entry-level role and 21.9 per cent of a C-suite one, and the layer that goes next is the one middle management was built to run: the handoffs, the sequencing, the approvals (all live specifications, as at 14 September 2026). Jack Dorsey said the same from the top. He cut Block from over ten thousand people to under six thousand in February, and in March wrote with Sequoia's Roelof Botha that hierarchy has only ever been a way of routing information and a system can now do the routing.
The coordination layer is collapsing before the execution layer. Lead remains the inflection point.
And the users are arriving. Across all employers on a posting-date basis, shares within each quarter's own postings, the share of roles that mention AI hired to use a tool went from 51.3% in Q4 2025 to 60.3% in Q3 2026 (to 14 September 2026, partial). Companies are hiring people to use tools nobody is accountable for, at an accelerating rate. The job that changed is the user's job, and nobody rewrote it.
Users gained 9 points in three quarters. Governors stayed inside a 2-point band.
Companies need to redesign how they operate. The role didn't move with the task.
The seat that answers for the model never gets hired. On the same basis, roles hired to govern the model's output ran at 2.5, then 4.3, then 3.5, then 3.7 per cent across the four quarters. Users gained 9 points in three quarters. Governors stayed inside a 2-point band.
Companies write AI policy far more often than they write about who decides. Across canonical specifications, active and closed, 1,354 roles name a governance framework and 36 name decision authority, 38 to one, as at 1 September 2026. The designed minority exists in every cohort, so the deficit is a choice rather than a property of the technology. Financial services designs best, and the companies building the models have barely designed for them either.
That is a design failure rather than a technology failure, and the organisations that own it are the ones writing the specifications, not the market they hire in. Adoption is not the question any more. Design is, and almost nobody is hiring for it.
People need to learn new skills.
Roles are changing shape, careers are losing their old ladders, and businesses are realising that many of their speed and efficiency problems are system design problems, not people problems.
Graduates used to learn judgement inside execution work. Employers are removing that apprenticeship without designing its replacement: of 608 live Australian entry-level roles, 8 are AI-central. On the same posting-date basis, the share of roles that mention AI hired to build the model fell from 27.8% to 17.7% in three quarters. The skill to build is leaving the hiring record while the skill to use arrives, and no tool rollout teaches judgement.
Australia reads this the hard way. Foreign-headquartered employers post in Australia and the United States in the same quarter, and they post a different job in each: builders are 15.0% of their United States roles that mention AI and 4.8% of their Australian ones, across 187 Australian roles as at 4 September 2026. The accountability for the model sits in another time zone and the Australian office gets the seat that uses the tool.
The same employers hire builders at home and users here. Employer rows publish at cohort level when the branch view lands.
The work changed faster than the organisation, and faster than the people inside it were taught. That is the premise. The thesis beneath it is the part I can measure:
AI is redistributing tasks faster than organisations are redesigning roles, decisions, and learning systems.
Where this came from
I've spent my career inside big agencies and global companies, Disney and M&S among them, and I see technology changing how work gets done every day. I had been building with AI a lot in my spare time, and I started to notice that it was fundamentally changing the way I worked. What improved my adoption of AI was redesigning how I worked, down to the task level.
I don't have a technical background. I spent countless hours after the kids went to bed learning, through AI, about engineering, systems, workflows, databases, agents, agentic workflows, all the buzzwords. I first built CareerOS, a career-context platform to help people apply for jobs and learn new skills, and that's when I switched to building intelligence first, not tools. Like any strategy work I've done before, I was thinking about a problem and reverse-engineering the data points I would need to understand it.
This led me to something I'm now calling workforce intelligence: a new way to measure, redesign and build the businesses, operations and workflows of the future. I wrote the argument down in January, under the name Future Skills Intelligent Work, before the instrument existed. At the end of July, I quit my full-time job to build it properly. I couldn't shake the desire to get as close as possible to redesigning how work is done.
What I think it means
People expected AI to save them time. It redistributed the work. People are saving time with AI and don't know where to put it, because nobody has redesigned how the time should be allocated.
AI provides the foundation for the process; it's not the output. The output remains the same. The work is to redesign the process and workflows, down to the task level, and to write down the decisions before the workflow rather than after.
Always start with the work, not the tools. I learned that by building the tool first.
This technology is not going away and is becoming a fundamental commodity. Businesses need to completely redesign how they operate, and continue to do so, and people need to learn, and relearn, new skills. My belief is that the economy will collapse because these won't happen quickly enough. It's less another wave of tools and more a structural reset: closer to the Industrial Revolution than a software update.
The path to agentification is the journey. Agents aren't the destination. People are overselling now because everyone is racing to the finish, and many are overlooking the work still to be done. There is so much work to be done.
So Verellax measures where a company sits, redesigns the roles, decision rights, and workflows to close the gap, and builds the systems and governance to keep it closed. My goal is to soften the collapse by supporting businesses in redesigning how they work and helping people learn the skills they need for the work that's coming, including standing in front of universities with the Skills Index data.
The smallest proof I have
The platform that produced every figure on this page is run by agents. Before any of them could act, I documented the frameworks and the thought process, built the context files they decide from, and codified the decisions: what each one may promote, what it must discard and why, and where I hold the gate. I wrote the decision rights for my own agents before I measured whether any company had written them for its people. That's why the tools were worth anything.
So read your own specifications back. Not the values page, the specifications. Whichever role first names who decides when the model is wrong is where the redesign has already started.
Designing work, not reacting to it.