The most expensive mistake in AI adoption
I don't come from research. I come from twenty years of operations — from places where a number stands at the end of the month and someone is accountable for it. In those years I kept living the same moment: capable people working in a dozen tools at once, exhausted by evening and still feeling they hadn't moved the actual mountain. And for the past few years a second moment has joined it: those same people talking about "AI" as if it were already everywhere — while the day-to-day still means copying spreadsheets, sorting mail by hand, and hunting down the pieces of a quote.
For a long time I couldn't imagine that the technology hadn't solved this yet. Today I know better: the technology is rarely the problem. The problem is the sequence.
Most AI initiatives don't fail on the model. They fail on selection and sequencing — on wanting too much at once, starting with the wrong process, and writing a plan that ignores budget, time, and the people meant to carry it. Understand that, and you don't need the courage for a grand gesture. You need discipline in the small things.
The wrong first question
The question almost every company starts with is: "Which AI should we buy?"
It's the wrong question. It pushes the decision outward — to the vendor, the tool, the feature comparison. And it produces the pattern I see everywhere: a dozen point solutions bought side by side, each plausible on its own, none connected to the others. A chatbot here, a document assistant there, an automation in marketing. Three months later nobody knows what's actually being used, and what one tool learns never reaches the next.
The right first question is a different one: "Which process do we tackle first — and how do we build the foundation so the next process can sit on top of it?"
That's not a semantic difference. It's the difference between a toolbox and a system. Between features and clarity. AI is not something you install. It is something you build correctly into the foundation — and then let grow, step by step.
Start where it's tangible
You don't build a foundation in the abstract. You build it on a real, tangible process — together with the people who run it every day.
There's a hard reason for that. A process the case worker has known for years brings three things no strategy deck can: the real edge cases, the honest numbers (how long does this actually take?), and the willingness to use the result. Introduce AI over people's heads and you build a monument. Introduce it with them and you build a tool.
That's why we never start with the most ambitious initiative, but with two or three tangible processes:
- clearly bounded, so success becomes visible,
- frequent enough that impact compounds,
- with data that already exists, rather than data that must first be collected.
The early, small, real win is not a "nice to have." It is the currency you use to buy the trust the bigger steps will need later.
The honest assessment
Before anything is prioritised, you need an honest fix on where you stand. Not for its own sake, but because a roadmap is only as good as the picture of reality it sits on.
A solid assessment captures four things:
- Maturity. Where does the organisation really stand — across strategy, data, infrastructure, processes, people, and governance? A company at level 2 of 4 gets a different plan than a leader. Sugar-coating helps no one.
- Data. What data exists, in what quality, and are you even allowed to use it? Most good use cases stand or fall on that answer.
- System landscape. What's in use — email, calendar, CRM, finance, industry software? Every interface is both an opportunity and a dependency.
- The constraints. Budget, time, and risk appetite. This isn't fine print; it shapes the plan from the start.
This assessment doesn't have to take weeks. A few key facts, the documents that already exist, and one good conversation are enough for a first, load-bearing read. What matters isn't completeness — it's honesty.
The roadmap: sequence is a decision
The assessment becomes a roadmap — and that comes not from the gut but from a traceable evaluation. Every conceivable use case is weighed against the same criteria:
- Value — what does it deliver, measured against the stated goal (cost, quality, growth, relief, compliance)?
- Feasibility — is it realistic with the data and systems on hand?
- Strategic fit — does it pay into what this company is actually about?
- Risk and compliance — how sensitive is it legally and in regulatory terms?
- Readiness — can the team carry it today?
From these comes a position for each case: quick win (high value, easy), strategic (high value, but demanding), fill-in (nice, but secondary), or later (neither urgent nor easy). And from that, the grouping into waves:
- Wave 1 — Quick Wins (0–3 months): two to three initiatives, no more. Fast, visible value, low risk, the foundation gets laid.
- Wave 2 — Expansion (3–12 months): what builds on the first wins.
- Wave 3 — Transformation (12+ months): the big levers, for which the foundation has to hold first.
The most important sentence sits in Wave 1: no more than two to three. The temptation to want everything at once is exactly the mechanism AI programmes suffocate on. Discipline in the first wave isn't sacrifice — it's the precondition for there being a second one.
Every initiative also gets clear stage-gates instead of a vague "it'll be fine":
- Pilot — useful in at least 80 % of cases? If not, you improve it; you don't roll it out.
- Validation — does the value hold in real operation, with real people?
- Scale — only now, with proof in hand.
And every initiative gets KPIs with a baseline and a target. Not "email handling gets better," but: today 8–12 minutes per request, target 3–5 minutes, measured weekly. Without a baseline there is no proven value — only a good feeling. And good feelings don't carry a roll-out.
Reality belongs in the plan
A roadmap that only "takes budget, time, risk, and law into account" at the end is a wish list. For us, those four shape the plan from minute one.
Budget and time aren't limits you plan against, but dials you turn. What happens to the roadmap with six months instead of three? With a tighter budget? Initiatives shift between waves, cost and timeline change — and you see the consequence before you commit. A plan should be a negotiable model, not a document carved in stone.
Cost we estimate deliberately vendor-independent — as a rough market range for build and run, not as a quote. Someone who knows an honest order of magnitude up front makes better decisions than someone who only senses, after the third sales call, what they've gotten into.
Compliance is built in, not bolted on. Every use case gets an EU AI Act risk classification; wherever personal data is involved, GDPR applies; for especially sensitive initiatives there are additional checks, such as a fundamental rights impact assessment. And since February 2025 the EU AI Act already requires demonstrable AI literacy in the organisation — training is not optional, it's an obligation. Those who need data sovereignty can run fully local: no data outflow, on-premise, cloud-free on request.
And finally: the people. In our experience success comes roughly 70 % not from the technology but from the people — from a visible sponsor, from champions in the business units, from clear communication and, where it belongs, from involving employee representation early. A technically perfect system nobody uses is a failed project.
The foundation that grows with you
Now comes the part that decides between toolbox and system — and that simply is missing from most AI initiatives.
When the first process runs, you reach a fork. The easy path: stand the next tool up beside it. The right path: dock the next process onto the same foundation.
For that to work, the foundation has to be designed for it from the start. Concretely, for us that means:
- A shared memory. What the system learns in the first process — about the company, its vocabulary, its quirks — is immediately available to the second. Knowledge is shared, not rebuilt inside every tool.
- A shared knowledge graph. Documents, cases, and relationships don't live in a dozen separate silos but in a structure every new process can reach.
- Shared governance. One place for permissions, audit trails, data sovereignty — not twelve settings dialogs that eventually drift apart.
- Modular and switchable. Every capability can be turned on or left out. You don't buy the maximum; you buy what fits your maturity — and extend as it grows.
That's how individual use cases gradually become something coherent. The second process ships faster than the first, because the foundation already stands. The third benefits from what the first two learned. That's the compounding interest point solutions never reach.
Why "operating system" isn't a label
We call the goal an operating system for the business. That's not a marketing word; it's a technical description.
An operating system on a computer does three things: it manages resources, it holds state, and it lets individual programs work together without each reinventing the wheel. That's exactly what an AI operating system does for a company — except the "programs" here are the business processes: the inbox, quotes, knowledge retrieval, planning, finance.
Instead of every process having its own little AI tool with its own memory and its own rules, they share a backbone. The shared memory is the state. Governance is the resource management. The knowledge graph is the shared address space through which processes reach one another. An answer in customer service may know what was discussed in sales — controlled, traceable, within the same system.
For us this architecture sits on ZenBrain technology and on publicly verifiable research; anyone who wants to go deeper will find it under ZenBrain technology. But the decisive thing isn't the depth of the algorithms. It's the stance: a company is not an assortment of tools. It is a coherent whole — and its AI system should be one too.
Try the approach for yourself
This approach — assessment, evaluation, waves, gates, KPIs, reality priced in — sounds like a lot. So we poured it into a tool you can run through yourself in a few minutes: ZenEntry.
You throw in a few key facts — size, sector, goal, constraints — or add existing documents and an interview transcript. From that, ZenEntry derives a complete draft: the right scope of functionality, the building blocks and interfaces needed, a prioritised roadmap with waves, stage-gates and KPIs — and a vendor-independent cost estimate across the first year. If you like, you turn the dials directly: more budget, more time, more or less risk — and watch the plan and the cost shift.
The framing matters to me: this is meant as neutral orientation, not a quote. It runs right in the browser, no sign-up. It doesn't replace the conversation about the edge cases every business has — but it replaces the blank page. And it makes visible what the method does at its core: turning a diffuse "we should do something with AI" into a concrete, sequenced, honestly costed first step.
If you want to see it for yourself: How to adopt AI the right way.
What remains
It took me a long time to understand why so many good intentions run into the sand. Not because the technology is missing. Because the sequence is missing — and the patience to build the foundation correctly before building on top of it.
So my ambition is unspectacular: no magic, just a disciplined path. Start with the people, at a tangible process. Evaluate honestly. Little at a time. Measure. And lay the foundation so the next process arises not beside it but on top of it. Hold to that, and over time individual quick wins grow into something that didn't exist in this form before: a real operating system for the business — one that gets smarter with every process, instead of more cluttered with every tool.
This is the first part of a series on adopting AI. The next part is about how we deal with the technology itself — where we leave the last word to the human, where governance begins, and why trust is the real scarcity.