I first understood what technology could really do back in the mid-1990s, on a very specific and very real problem the NHS had: generating critical patient documentation. The process took days, and those days were affecting patient care. Using what was then leading-edge technology, I designed and built a system that turned those days into minutes.
That was the moment the point of software became clear to me. We build it to turn data into information, information into knowledge, and knowledge into the insight that actually changes an outcome. Almost everything I've done since has been some version of climbing that same ladder.
Data science, machine learning and then AI accelerated the climb enormously. Eight years ago I got ahead of myself and started building a spatial-computing platform meant to read real-time interaction signals, turn them into intent, and assemble experiences around the user on the fly. Right idea, wrong time, wrong place. The technology just refused to keep pace with my appetite for it.
That began to change with commercial AI transformers. About two years ago I decided the stack was finally ready for another run at the problem, this time aimed squarely at giving small and medium businesses the kind of superpowers only the largest enterprises had ever been able to deploy.
When you found a company on the democratisation of AI, one of the first questions is how far you're willing to commit. My guiding principle was simple: trust the technology. I couldn't in good conscience bet the farm on a technology I didn't trust enough to use myself. So I went all in. The product's engine is AI, I used AI to help build it, and I built tools with AI that in turn use AI to help manage the building of it. Around it is 35 years of software design rigour and architectural experience of building systems at scale. But the technology being leveraged is AI.
So I'm watching all of this from a particular seat: three decades chasing the same idea, and a company built to put that power into everyone's hands. That's why the present moment unsettles me.
I pay around $800 a month for AI, willingly, because my work depends on it, and I'm about as well-placed as anyone to make sense of this landscape, yet I still find it hard to untangle. Which model leads for a given job changes by the month. Half the advice online is already stale. The thing you carefully chose in March can be pulled out from under you in June. Things that seemed impossible last quarter are a well structured prompt away.
And the complexity is only half of it. The conversation around AI has become mostly noise: hype, fear and vendor spin, confident takes contradicting each other from one week to the next. It's not only that people don't trust the technology yet. The information that's meant to help them is riddled with bias and misinformation. Rather than preparing people for the transformation ahead, far too much of it misdirects them, and too often buries the truth under whatever suits the loudest voice in the room.
If it's hard for me, how does an individual, a small business, a medium-sized organisation, or even a large-scale enterprise navigate these waters?
The promise of AI was never just "better technology." It was democratisation: expertise that used to be reserved for large enterprises, finally within reach of everyone. I still believe that future is possible. I'm just no longer convinced it turns up on its own.
The wrong metric
Whenever the cost of AI comes up, the same reply arrives within seconds: "but token prices are falling." And they are. The trouble is the token is the wrong thing to measure. What matters isn't the price of a token, it's the price of getting to an outcome, and the amount of intelligence you have to burn to reach a useful one keeps going up. Reasoning models think in longer and longer chains, agentic systems fire off several models to do one job, and retrieval pipelines keep stuffing more context into the window. So the individual units get cheaper while the number you need keeps climbing. The cost of intelligence is falling. The cost of actually solving a problem often isn't.
Two markets, not one
AI isn't a single market, which is easy to forget. At the commodity end, intelligence has become almost absurdly cheap. DeepSeek, Qwen, Llama, Gemma and Mistral handle the overwhelming majority of everyday work for a fraction of frontier prices. Frontier reasoning is heading the other way, with each new generation doing more and costing more to do it. The cheap end keeps getting cheaper and the expensive end keeps pulling away, and cheap AI isn't the same thing as accessible intelligence.
The difference lands unevenly. Big organisations have options. They negotiate pricing, put people on optimisation, route the dull work to cheap models and keep the expensive reasoning for the problems that actually need it. Most small organisations have exactly one lever, and it's their credit card. Which raises an honest question: if the people with the most to gain from all this are also the ones with the fewest resources, how do we keep their access survivable?
Underneath all the pricing there's a deeper point we rarely say plainly. What we're rationing here isn't really compute, it's the machinery that turns raw data into knowledge, the thing that lets a person or an organisation understand their own situation and do something about it. That used to be scarce, and reserved for whoever could afford the analysts. If AI ends up sitting in a tier only the well-resourced can reach, we won't just have an uneven market. We'll have rebuilt the oldest gap there is, and handed the biggest lever to the people who needed it least.
Price isn't the only lever
Price rations slowly. There are quicker ways for capability to slip out of your reach.
Last week made that plain. Anthropic released its most capable model yet, and within about three days it was gone. Not for some of us, for all of us. A government export-control order citing national security had the company suspend access worldwide overnight. Live sessions just started erroring and fell back to older, weaker models. It's worth getting the story straight, because the easy version has it backwards. The model wasn't turned into a weapon. Its maker had declined to let it be used for autonomous weapons and mass surveillance, and everyone else's access vanished as collateral. Forget the specific company and the specific administration. The point is that frontier intelligence can now be switched off over your head by decisions you had no part in. Which is an uncomfortable argument for owning your capability rather than renting it.
Architecture matters more than models
For a long time the whole conversation was about the model. Bigger, smarter, more powerful. I've become convinced the architecture around it matters just as much, partly because most people shouldn't have to navigate any of this on their own.
Most organisations don't need frontier reasoning for every task. They need the right level of intelligence pointed at the right problem at the right cost, without turning themselves into a research unit to work it out. The future probably isn't one model doing everything. It's more likely a mix: open models handling the predictable load, local infrastructure giving you sovereignty and a bill you can actually predict, frontier models kept back for the genuinely hard problems, and a governance layer that makes the routing visible instead of hidden. Get that balance right and you don't just spend less. You get access to intelligence you can actually rely on.
The playing field doesn't level itself
I'm still optimistic, for what it's worth. The progress of the last few years has been remarkable and the road ahead looks better still. But I think we made one quiet mistake. We treated democratisation as an inevitable by-product of the technology, when it never was. The technology only creates the possibility. People have to create the outcome.
If we actually want AI to level the playing field, we have to build for it on purpose: open models, local-first where it makes sense, and orchestration and governance that optimise for the outcome rather than the prestige of whatever model happens to be fashionable this week. Accessibility isn't something the pricing page hands you. It's something you engineer.