The industry has found its new mantra: AI-first. It shows up in strategies, in sales pitches, on the homepage of any website that wants to look modern. And I, who do this daily, am convinced it is the wrong order.
In my previous article I described the three levels of AI mastery and finished with a question I deliberately left open: why do so few professionals reach the highest level, the one where AI gets integrated into real processes? I have thought about it a lot since then, and the answer is not technical.
The obstacle is not the tool. It has never been so easy to get hold of one.
The real bottleneck in AI adoption is not mastering the tool. It is never having learned to think in processes.
The mirage of the tool
Stop and think about it for a moment. If the adoption problem were access to the technology, adoption would already be universal. Anyone can open ChatGPT, Claude or Gemini in thirty seconds. Anyone can create an account on Make, n8n or Zapier this afternoon. The technical barrier to entry, which was real a decade ago, is practically zero today.
And yet most companies still get no real value out of it. Not because they do not have the tool, but because they do not know where to put it.
That is the underlying misunderstanding of "AI-first": it confuses having access to a capability with knowing how to apply it. They are two completely different things. Having the most powerful engine in the world is useless if you have no chassis, no steering and nowhere to go.
A faster process that is broken is still broken
The "AI-first" reflex is to look for somewhere to fit the AI in. And it almost always gets fitted onto a process nobody has stopped to understand.
The result is what I call AI theatre: a chatbot stuck on top of a broken flow to look innovative. It looks modern. It changes nothing underneath. Sometimes it even makes things worse, because now the error happens faster and at greater scale.
There is an idea attributed to Bill Gates that sums it up better than anyone: automation applied to an efficient operation multiplies its efficiency; applied to an inefficient operation, it multiplies the inefficiency. AI is the most powerful amplifier we have ever had in our hands. And an amplifier does not discriminate: it amplifies whatever you put in front of it.
That is why optimising cannot start with the tool. It has to start with understanding the process. And you cannot optimise what you cannot see.
The four questions that come before AI
When I come across a problem, a task or a process, I always do the same exercise. And AI does not appear anywhere in it. Not yet.
Before thinking about automating anything, I break the process into four elements:
- Data input. Where does the information come from? In what form does it arrive: an email, a form, a spreadsheet, a sensor, a phone call? Is it reliable? What triggers the process? If the input is chaos, no amount of intelligence downstream will fix it.
- Connection nodes. Where does that information have to travel? Which systems or people pass it along? This is almost always where the real friction lives: at the points where a figure gets copied by hand, rewritten, reformatted or lost between two tools that do not talk to each other.
- Processing. What transformation does the data undergo? What decisions get taken? What is a fixed rule and what needs judgement? This is the only point where, sometimes, AI earns its place: when something has to be reasoned, interpreted or generated that a simple rule does not cover.
- Expected output. What is the deliverable? In what format, for whom and to what standard? How do I know it is right? In practice, this is the best point to start thinking from: if you define the output precisely, the rest of the process orders itself backwards almost on its own.
The flow is always the same: input → connection nodes → processing → output.
And here is the important part: none of the four questions mentions AI. You can — and should — answer them before deciding whether AI has anything to contribute. Most of the time, the best optimisation does not even involve adding intelligence: it involves removing a redundant step, connecting two tools you already have or fixing the quality of the input. AI is one possible answer to the processing question, not the starting point.
An everyday example: that report someone compiles by hand every month by pulling data from three different places. "AI-first" would say: "let the AI write it". But when you break it down you discover that eighty per cent of the problem is in the input and the nodes — scattered data someone copies and pastes — and that gets solved by connecting the sources, with no AI involved. Only the last stretch, writing the report’s narrative, justifies using it. You have optimised the whole process; the AI does a small part, but it does it where it genuinely contributes. That is exactly how we approach any process automation project: the map first, the tool afterwards.
The discipline can be trained (and I train it daily)
The most interesting thing about all this is that it is not a talent you are born with. It is a discipline. And, like every discipline, it can be trained.
I do this exercise constantly, even with processes I know I will never implement. I watch how a café takes its orders, how my bank’s support handles me, how emails pile up in my inbox, and I break it down mentally: what is the input, where are the nodes, what processing is there, what would the ideal output be? The vast majority of those exercises lead nowhere. It does not matter. Each one is a repetition that sharpens the instinct.
Because that is the real objective: that when a real opportunity turns up, you do not have to strain to analyse it. You see it already broken down. The professional who has done this exercise a thousand times looks at a process and, at a glance, knows where it bleeds and what would fix it. That fluency is not bought with a tool and not downloaded with a subscription. It is built repetition by repetition.
As I once said: the difference is not talent, it is decision.
From tactic to company philosophy
If this stays an individual skill, in "the AI person" on the team, its impact is limited. The real leap happens when it stops being a technique and becomes a company philosophy.
An "AI-first" company buys tools, orders people to adopt them and measures "AI usage". It ends up, almost always, with AI theatre spread everywhere.
An optimizer-first company does something different: it teaches all of its people — not only the technical ones — to look at their own work through those four questions. The question in the air across the organisation stops being "where do we put AI?" and becomes "where is the friction and what is the simplest way to remove it?". When that question is the culture, AI gets deployed with a scalpel, exactly where it earns its place, instead of being sprinkled over everything to look modern.
The paradox is that this second company ends up adopting AI far more deeply than the first. Because it rests it on processes it genuinely understands.
The conclusion, against the current
The industry repeats: be AI-first.
I would put it the other way round:
First optimizer, second AI.
And let us be clear: putting AI second does not mean it matters less. It means it comes second in the sequence, not in importance. AI’s power is real and enormous. But power without direction is just noise. The optimisation mindset is the steering wheel; AI is the engine. An engine with no steering wheel takes you nowhere: it just crashes you faster.
The companies that win the next decade will not be the ones that adopted AI earliest. They will be the ones that learned to think in processes first, because they are the only ones who will know where to point it.
If you are interested in exploring how to install this way of thinking — optimise first, automate afterwards — inside your organisation, in our AI consulting for companies we always start there. Let us talk about your operation and we will tell you frankly where the friction is and what makes sense to optimise before touching AI.