We build in two stages. Not out of caution, not out of AI scepticism - but because any other order almost always burns money. Here's why.
AI is everywhere. In every pitch, in every demo, in every other LinkedIn post. And the temptation to simply start there is great.
We don't do that. With us, AI always comes second - after a first stage that looks much less spectacular: cleanly mapping your work into digital form. This order isn't a detail of our approach. It's the reason the tools work in the end.
Map the process digitally
Capture the workflow, structure the data, make bottlenecks visible.
Apply AI deliberately
Where it actually saves time - and only there.
If your daily work today consists of Excel spreadsheets, WhatsApp notes, slips of paper on the desk and an old program no one fully understands any more, then AI on top of that won't rescue anything. It will only veil it. Magic on top of chaos stays chaos - just less visible now.
The honest order is: first know what's actually happening. Which steps really exist, which data flows where, at which points time gets lost. That's stage one. It's more boring than AI - but it's the only one after which AI later has a chance at all.
A widely cited MIT study from summer 2025 ("The GenAI Divide," MIT NANDA Initiative) followed 300 AI implementations in mid-sized businesses. Result: 95 percent of pilot projects delivered no measurable business success. The most important reason wasn't the technology - it was the missing process foundations the AI could have docked onto.
And this isn't a passing fad. Back in the 1980s, Michael Hammer coined the phrase "Don't automate, obliterate" (Harvard Business Review, 1990): whoever digitalises a bad process gets a faster bad process. Bill Gates put it this way in 1999: "Automation applied to an efficient operation will magnify the efficiency. Automation applied to an inefficient operation will magnify the inefficiency." For AI it's doubly true - it's the most expensive automation we've ever had.
Many providers sell "digitalisation" by replacing an existing form with a web form. The printed order sheet becomes a PDF to upload. The Excel list becomes a longer Excel list. That's not digitalisation. That's a move.
What we mean by stage one: we map the workflow structurally. Who does what when, where the information comes from, where it goes, what states an order has, which steps are necessary, which are merely habit. Only that makes data usable - for you, for reports, for clean interfaces. And for everything that becomes possible in stage two later.
The term is process mining (Wil van der Aalst, TU Eindhoven, from 2004): the discipline of making real workflows in companies visible - not the ones described in the manual, but the ones actually lived. A Bain & Company analysis from 2023 shows: mid-sized businesses that do process mining before an AI project achieve on average three times the ROI of AI projects that start straight in.
This fits with an old computing classic: "Garbage in, garbage out" (George Fuechsel, IBM 1957). Translated: a machine can't make clear results from unclear inputs. With AI this becomes more painful - because it looks so convincing, even when the input was garbage.
AI needs two things: data and a clear problem. If your business has no structured data picture of itself, the first is missing. If no one can articulate what's actually meant to be saved or improved, the second is missing. Both have to be in place before the AI - otherwise it just guesses more elegantly.
That's exactly what happens in stage one: the daily chaos turns into reliable data points. Orders get a status. Appointments get dates. Customer enquiries get fields. And suddenly you can ask questions that weren't answerable before - which enquiries lead to which orders, on which weekday do most complaints come in, where is the bottleneck. These questions are preparation for any AI. Without them, it has nothing to say.
The keyword is data readiness - the question of whether a company is at all ready for AI in data terms. A BCG study from 2024 ("Where's the Value in AI?") found: only 26 percent of the companies surveyed produced measurable value with AI - and they split strikingly cleanly on one criterion: they had invested in a structured data foundation beforehand. The others had bought AI. The first had a tool. The others had theatre.
Andrew Ng, one of the most influential AI researchers, calls this "Data-Centric AI": in recent years, the leap in usefulness comes less from better models and more from a better data foundation. Whoever skips the foundation is buying cutting-edge technology for a building that will never stand.
When stage one is in place, you can ask honestly: where does AI belong, where doesn't it? The answer is pleasantly clear. AI is strong at pattern recognition in large volumes, at language and text, at classification, at suggestions a human checks at the end. It's weak anywhere precise logic, hard rules and complete correctness are required.
Appointments, master data, invoices, accounting - classical software does those better. It's reliable, traceable and cheap to run. AI, on the other hand, is the right choice when two important enquiries need to be filtered out of 50, when handwritten notes need pre-sorting, when long texts need summarising. Those are the real time-sinks - and that's exactly what stage two solves.
The Stanford concept for this is called human-in-the-loop (Stanford HAI, since 2019): AI delivers the suggestion, a human checks and approves. It's not the most spectacular use of AI - but the one in which studies like McKinsey 2024 ("The state of AI in 2024") consistently measure the biggest productivity gains, 30 to 50 percent depending on the sector.
Erik Brynjolfsson (MIT, later Stanford) sharpened this in his concept "the Turing Trap" (2022): whoever builds AI to fully replace the human gives away value. Whoever builds AI that amplifies human strengths creates it. We build by the second principle - and we build it only when it's clear which human strength is to be amplified. That only works once stage one is in place.
The most common request we hear is "We want to do something with AI now too." The pressure behind it is real - competitors talk about it, the board asks, the press covers it. We understand. And still, our answer is mostly the same: let's start three months earlier, at the place no one sees.
Reverse the order and you build twice. Once the AI, which doesn't work without a foundation. And then the foundation underneath, after you've noticed it's missing. A bad AI integration isn't "easier to throw away" than a missing process either - it's damage to the trust of employees and customers, and it lasts a long time. Quick shots are rarely cheap in the digital world.
Gartner's Hype Cycle describes the pattern: every new technology first runs through a "Peak of Inflated Expectations" - overblown hopes, many pilot projects - and then the "Trough of Disillusionment" - disappointment, cancellations, loss of trust. Only after that does the sustainable "Plateau of Productivity" arrive. Generative AI was at the peak of the hype in 2024. Whoever buys there buys expensively.
The second keyword: sunk cost fallacy (Arkes & Blumer, 1985). Whoever has once put money into a bad AI investment rarely admits honestly that it isn't working - and instead pushes more money in afterwards to justify the failure. That's human, but expensive. The best prevention is starting in the right order.
The order isn't caution. It's respect - for your money, your time, your people.
AI is a force multiplier. It amplifies what's already there. But it also multiplies by zero - and zero times a hundred is still zero. Whoever honestly wants to build good tools therefore starts where the substance comes from: at the invisible foundation everything else later rests on. AI on top of that is then no longer hype. It's leverage. And exactly then does it become what the advertising promises: real relief.
Sometimes that means: stage one is enough for now. Sometimes: here AI actually makes sense - and exactly at this one place. But we never decide that before the first conversation.
How we hold the first conversation and clarify the problem before building is on Before we build, we figure out what's broken. Why we look especially carefully at AI is on What AI can't do, even when it looks like it can. The quality standard we apply to everything we build is on When is software actually good.