There’s a growing unease in boardrooms. The AI budget was approved, the tools were bought, the team was trained. Everyone agrees it’s impressive. And yet, when you look at the numbers that actually matter — revenue per employee, cycle time, cost to serve — very little has moved.
The conclusion most people are reaching is that AI was oversold.
I think the conclusion is wrong. What’s been oversold is the idea that you can buy AI without changing anything else.
The retrofit trap
Here’s what it looks like in practice. Your agency now writes campaign copy in four minutes instead of four hours. Genuinely remarkable. And your campaign still takes six weeks to go live.
Why? Because the six weeks were never about the copywriting. They were about the brief being written on Monday, reviewed on Thursday, revised the following week, sent for approval, batched into a monthly review, and scheduled around a production calendar that assumes everything takes time.
That entire structure — the handoffs, the review gates, the weekly cadence, the approval chain — exists because human beings were the bottleneck. It was built to manage scarcity. Remove the scarcity at one step, and the scaffolding built around it doesn’t fall away. It just absorbs the gain.
This is the retrofit trap: we’ve made steps faster inside processes that were designed around slow steps. The task speeds up. The system doesn’t. And it’s the system that shows up in the P&L.
Almost nobody designed these processes. We inherited them. They accumulated. Which is why this isn’t an indictment of anyone’s judgement — it’s just that AI has made the seams visible for the first time.
The same mistake, in four different rooms
Marketing. Every step is faster — research, copy, design variants, reporting. The campaign still takes six weeks, because approval is monthly and production is a queue.
Hiring. Five hundred applications used to take a recruiter a week to shortlist. AI does it in an hour. But the company still runs five interview rounds over three weeks, because that process was designed for a world where shortlists were unreliable. Hiring still takes a month. You fixed the one-week part and left the three-week part exactly as it was.
Reporting. Someone used to spend three days a month pulling numbers into a deck, because leadership couldn’t see the data themselves. AI now produces that deck in ten minutes. It feels like a clear win. But the deck only ever existed as a workaround for not being able to ask a question and get an answer. The AI-native version of that process doesn’t have a faster monthly report. It has no monthly report.
Customer support. AI drafts responses in seconds. But the L1/L2/L3 tiers remain, and the SLAs are still written around queue time — both of which exist because human attention was expensive and had to be rationed. The AI-native version isn’t a faster ticket. It’s a product that answers the question at the point of confusion, so most tickets are never raised.
Collections is the same story: instant AI-drafted reminders, still sent on a monthly cycle behind an ageing report that only exists because nobody could watch every invoice at once.
Notice the pattern. Retrofit makes work faster. AI-native makes work unnecessary.
The real opposition isn't inefficiency. It's inertia.
Inefficiency has no defenders. Nobody stands up in a meeting to argue for the six-week campaign.
Inertia has plenty of defenders — and they are reasonable people with reasonable fears.
What if the new system fails? The old process is slow, but it is survivable and predictable. Nobody has ever been fired for a six-week campaign. People do get fired for a broken one. Slowness is a shared, invisible cost. Failure has a name attached to it.
What if this is a bubble? Everyone in a leadership role has lived through at least one transformation programme that quietly died. Rebuilding your core operations around a technology that might deflate feels reckless. So people wait — and waiting looks a lot like prudence.
And the question nobody says out loud: if the process shrinks, what happens to the person who managed it? Redesigning a process usually means fewer handoffs, fewer roles, fewer meetings. You are asking people to redesign themselves out of relevance and then wondering why adoption is slow.
Once you see this, the popularity of tool-buying makes complete sense. Buying a tool is a low-risk gesture. It looks like progress, it threatens nobody, and it changes nothing. That, more than any limitation of the technology, is why AI investment isn’t showing up in outcomes.
I’ll admit we learned this the hard way ourselves. We built a large internal AI framework for our content team — skills, SOPs, prompt libraries, hundreds of documents. It was good work. The team barely touched it, because using it meant each person voluntarily redesigning their own way of working, on top of their actual job. It only started producing value once we stopped handing out capability and rebuilt the process itself, so the new way was simply how work happened. The lesson was uncomfortable and useful: toolkits don’t get adopted. Systems do.
About the bubble
Let’s be honest about the froth, because it’s real. Valuations are stretched. Companies are raising on narrative. A great many products have bolted on a chatbot and repriced themselves as AI companies. Some meaningful share of what’s being funded today will not exist in three years.
But I want to separate two things that keep getting collapsed into one: the market and the capability.
I’m not a reflexive believer in new technology. I never thought blockchain would fundamentally change anything, and a decade after the peak of that hype, here we are. It has had real impact in a few narrow places. It has not come close to what was promised. And the tell, in hindsight, was that you always had to be told why it mattered. It needed a whitepaper, an evangelist, a future tense.
AI needed none of that. Nobody had to be convinced. People used it once and were amazed — and I don’t think there’s a single honest person who can say otherwise. Beyond the feeling, there’s a harder signal: it is already doing real work, at scale, every day, paid for out of operating budgets rather than innovation budgets. Blockchain never crossed that line.
So yes, the market may correct. Bubbles pop; that’s a statement about prices, not about whether a capability disappears. Railways, telecom, the internet — the valuations collapsed, the infrastructure stayed, and the real productivity arrived afterwards, once the noise cleared and people finally rebuilt their operations around what had been laid down.
Which makes this an unusually asymmetric bet. Convert to AI-native operations and the market crashes — you still have a faster, leaner, cheaper business. Wait for certainty and it doesn’t crash — you’ve handed your competitors a two-year head start that can’t be bought back later at any price.
We are not going back to the pre-AI world. That much seems settled.
The question worth asking
Stop asking where can AI help us do this faster? That question can only ever produce a retrofit.
Ask instead: if we were starting this business today, with these tools, would this process exist at all?
It’s a harder question, and it produces uncomfortable answers. Most processes fail it. That’s the point. The real measure of how AI-native a company is isn’t how many tools it has bought — it’s how many processes it has deleted.
None of this is a technology problem. The technology is available to everyone, at roughly the same price, today. It’s a leadership problem, because redesigning a process means someone has to accept risk they aren’t currently paid to take. That’s the actual bottleneck, and no vendor can solve it for you.
If you want a place to start, don’t start everywhere. Pick one process — one that everybody privately agrees is bloated — and redesign it from zero, as if the old version never existed. Don’t touch anything else. Get it working, let people see the difference, and let the appetite build from there.
The companies that do this over the next two years won’t be the ones with the best AI. They’ll be the ones that were willing to give up processes that still, technically, work.
