Your AI Has Nowhere to Go
AI is in. The organisational refactor hasn't followed.
A note: this one’s a longer deep dive than my usual posts. Not a permanent shift, just something I wanted to try. Let me know what you think.
This year I’ve made a point of seeking out conversations across a wide range of industries and organisations. Founders at early-stage startups, operators at AI-native companies, executives at blue-chip firms, people working across finance, healthcare, government and not-for-profit, and even a leading frontier AI lab to boot. Different scales, different levels of technical sophistication.
Not everywhere, but in a lot of the more conventional organisations there’s a moment in those chats when something shifts. We start on personal productivity, what AI is doing for individuals, the wins, the time saved. Then I ask about the broader business and the picture doesn’t quite match.
A former client at a mid-size financial services firm put it most plainly. Early adopter type, strong performer, he walked me through everything he’d quietly built: real time saved, genuine leverage, etc. Then I asked about the wider company. Most colleagues were using AI for basic queries at best, a glorified Google, and the ones engaging more seriously had nowhere to take it. No shared systems to plug into, no organisational pull from above, no way to multiply what they were doing individually.
Across more conversations than not, the pattern was the same: the AI was working well for the individuals using it, but had nowhere to go beyond their desk.
The productivity paradox
The individual productivity gains are real, and at this point broadly acknowledged. Most surveys over the past year have landed in roughly the same place: people using AI tools are completing tasks faster, producing better output, reporting genuine time saved.
A 2025 evaluation of Microsoft Copilot across 56 Australian Government agencies makes the point well. Nearly 6,000 licences, 69% of participants reporting faster task completion, 61% reporting better quality work. Hard to argue with. But the same evaluation found that 61% of managers could not confidently identify when Copilot had been used at all. The productivity was real it was just invisible, contained within individuals, leaving almost no trace in the shared work of the organisation. That’s not a surprise when you look at how people were actually using it: summarising documents, re-drafting content. Personal efficiency tasks. Not workflow change. The trial tells us something honest about what AI can do for a willing individual. It tells us rather less about what it does to how an organisation works.
Zooming out, the picture shifts again. McKinsey’s 2025 State of AI survey found that whilst 88% of organisations are now using AI in some form, only a third have begun scaling it at an enterprise level. Jan Hatzius, Chief Economist at Goldman Sachs, reported that despite hundreds of billions invested in AI infrastructure in 2025, the economy-wide productivity impact was ‘basically zero.’
This is the productivity paradox. Whilst AI is clearly making individual tasks faster, cheaper and easier – the same outcomes are yet to be fully realised at the company, economic or societal level.
We have been here before
A useful lens for where we currently sit is Carlota Perez's theory of technological revolutions. She argues that transformative technologies move through two broad acts: an Installation Period, where speculative investment floods in ahead of real adoption, and a Deployment Period, where the gains finally materialise. Within the Installation Period, the Frenzy phase is where capital peaks and hype runs hottest. Most analysts applying her framework to AI place us squarely there right now.
Ben Thompson wrote about this directly in his piece ‘The Benefits of Bubbles’ last year. His argument is that the current AI spending surge, whilst likely to end in a correction of some kind, is also laying necessary groundwork, much like the dotcom bubble's frenzied fibre optic build-out quietly became the backbone of today's internet. The bubble isn't the problem. The question is what it leaves behind.
As someone who can't resist a history analogy: we have been here before. Electrification took roughly 30 years to meaningfully show up in productivity statistics. Office computers, another 15. The Gutenberg press, closer to 70 before it reshaped anything beyond a printing shop (see my earlier article re Ada Palmer). In each case the lag wasn't the technology, it was the time it took organisations to fundamentally redesign around it. The factory floor had to be rebuilt. Business processes had to be rewired. The technology was the ‘easy part’.
Which brings us to the tension that makes AI’s current moment rather different from those earlier examples. A company starting from scratch today carries none of that redesign burden. No legacy systems, no inherited workflows, no org chart built around assumptions that predate the technology. They don't have to transform, they just have to build. They’re AI Native. For everyone else, that’s the hill.
What good looks like
A personal favourite case study is Ramp, a high-growth fintech with a team of 1,000+, who have achieved what genuine organisation-wide AI adoption actually looks like in practice. With 99.5% of the team actively on AI tools, 84% using coding agents weekly – including 12% non-engineers now accounting for human initiated PRs on their production codebase.
You can read a more in-depth case study here by their Head of Product, Geoff Charles, but the key takeaways run downstream from the fact that the culture and leadership made AI a natural part of the work, not just an empty mandate. Some key takeaways:
Made it easier and more inclusive, particularly for non-technical people: They built Glass, their own Claude-powered internal agent fully wired into Ramp’s systems, so anyone could have a capable AI assistant without needing engineering skills. They ran a 700-person hackathon across sales, CX, legal, marketing, and finance. They had designated AI experts whose sole focus was getting colleagues to their own “aha moment”
Removed friction – met people where they were: Connected Claude and Notion AI with existing workplace tools set an immediate low bar that everyone could participate immediately. They also built a structured four-level AI proficiency ladder (L0-L3) to bring people along from wherever they were starting. They pre-connected 30+ tools (Salesforce, Figma, Snowflake, Gong) so when someone opened Glass, everything was ready to go. No token limits, no tiered access, no IT approved queues.
Attached to the problem, not the tool: A culture developed where people weren't attached to their tools; they were attached to their problems. Tools shipped in January 2026 were already intentionally obsolete by March.
Peer-led, not permission-based: The playbook spread organically rather than through mandate. An internal leaderboard tracked usage and apps shipped across every team. When you can see a colleague on another team shipping tools that save them hours, you don't need a top-down push to start building.
“That loop - build, share, inspire, build more - does more than any mandate or memo. The biggest surprise wasn't who built the most. It was how many people had been waiting for permission to build at all” — Geoff Charles, HoP at Ramp
The Honest Audit
If the person on your team getting the most out of AI left tomorrow, what would remain? Would the processes they built, the tools they configured, the shortcuts they figured out still be running – or would it walk out the door with them? That's the difference between individual productivity and organisational capability. And it's the gap that's hardest to see from the inside.
Ann Miura-Ko developed a framework to help, drawn from visits to various scaled companies (including Ramp) as part of her work as an investor and researcher. She frames it like the SAE autonomous vehicle level framework, each rung clearly distinct in its capabilities, so there's no ambiguity about where you sit. Each level asks the same four questions: What can AI see? What can AI do? Who can extend the system? How has the org changed?
L0 AI as theatre: Knowledge lives in people’s head, undocumented meetings, and SaaS tools AI can’t read. Org chart unchanged.
L1 Personal productivity: AI sees only what individuals feed it. Helps draft, summarise, brainstorm, code. No action on systems of record. Org hasn’t changed – maybe a ‘Head of AI’ hire or rep.
L2 Team workflow: Inter-team shared context. Functional AI for sales, support triage, code review.
L3 Organisational infrastructure: Agents act across systems, updating CRMs, opening PRs, routing tickets, etc. Non-engineers author skills, not just consume. Evolved org chart.
L4 Compounding operating system: Agents with policy-driven decision authority within scoped domains. Non-engineers ship internal tools. Hierarchy collapses toward ‘channel managers’.
L5: Virtually self-driving: Core operating loops sense, diagnose, initiate work, execute within delegated authority, update shared memory, improving future behaviour. Humans govern strategy, taste, risk, and exceptions.
“An AI-pilled company is not simply an AI-assisted version of an old company. They are organisations rebuilt around a new operating model.” — Ann Miura-Ko
The bottleneck isn’t the tools. It’s a failure to recognise that this is a culture and strategy problem masquerading as a technology one. That reframe has to start at the top, and it requires more than enthusiasm for the space. Leaders who lack a clear-eyed grasp of what AI can and can’t actually do can easily get caught up in making decisions that misdirect towards hype-cycle thinking and vendor influence. Aaron Levie, Box CEO, recently commented that “CEOs are uniquely prone to AI psychosis because they’re sufficiently distant from the last mile of work that still has to happen to generate most value with AI.” The most visible symptom of late: headcount cuts made on the assumption that AI had already closed the gap, only to discover it hadn’t. In some cases, quietly rehiring.
Closing The Gap
The issue is that many organisations are still framing AI primarily as a subtraction exercise. Dan Shipper, CEO of Every, makes the case in ‘After Automation’ that automation doesn’t eliminate the need for human judgement, but that in many ways increases it. His prediction is that the ‘AI jobpocalypse’ is not coming, but that the work shifts, it doesn’t disappear. Much like previous platform shifts the initial instinct is often replacement, but the enduring value tends to come from organisations that redesign how people work around the new capability.
Ramp worked because it had strong conditions for inside-out change: high agency, a tech-first culture and a genuine permission to build. Many incumbent organisations have the potential to follow a similar playbook. However many don’t, and may be best suited to outside-in change.
The recent rise of AI enablement and integration partners is itself a signal, confirming that the deployment gap is structural and that a software licence alone doesn't close it. Consulting firms are moving into this space – both existing incumbents but also native startups as both advisory and toolkit providers. Legacy enterprise systems integrators are repositioning. Private equity is getting in early, backing AI-native operating models as a value creation thesis in their own right. And major model providers themselves (Anthropic, OpenAI) are now more recently moving into help companies integrate directly into their models and harnesses.
This raises its own tensions. Building around a single provider can create dependencies that are easy to underestimate, particularly as models, pricing and capabilities continue to shift rapidly. The Australian Government Copilot report openly flagged this concern. Not as an argument against working directly with model providers, but as a reminder to approach the space with interoperability and flexibility in mind.
The Work Ahead
We’re at what should be a genuine turning point for AI in the workplace. But the gap will only widen between organisations that navigate that transition well and those that don’t. It won’t close because the tools get better, but because of the harder organisational work that many haven’t seriously started yet.
The pattern in those conversations kept pointing back to the same thing: the harder work sits around the technology, not inside it.
If you’re working on this from inside an organisation, advising on it from the outside, or simply watching it unfold, I’d love to hear what you’re seeing.