Yet the decision can be systematised. There are four levels at which AI enters a company, and each comes with clear signals.
The four levels
Level 1: ready-made assistants. Claude, ChatGPT and friends with a business licence. Cost: in the range of 20 to 40 euros per user per month. Immediately capable of: drafting, summarising, analysing uploaded files, research. The decisive detail is in the contract, not the brochure: on business plans, your inputs are not used to train the models. For company use, that’s not a footnote — it’s the ticket in.
Level 2: AI inside software you already own. Copilot in Office, AI features in your CRM or ERP, receipt recognition in your accounting package. The charm: no new infrastructure, the data is already there. The limit: you get what the vendor envisioned — nothing more.
Level 3: configure instead of code. Load your own documents into an assistant and make it company-specific — without a line of code. For many knowledge use cases this is the underrated sweet spot: roughly 80 percent of a custom project’s value at a fraction of the effort. Its limits show up around data volume, access control and traceability.
Level 4: a custom project on your own data foundation. Pipeline, central data model, and where needed a curated knowledge base with source-bound answers. The only way when company-specific numbers must be exactly right, or when knowledge needs permissions, curation and citations. And the only level that leaves a foundation behind.
The triage questions
Five questions, answered honestly, sort almost every case.
Does the answer require company-specific numbers that must be exact? Then there’s no way around Level 4 — more precisely, around the data foundation beneath it. A language model without clean data guesses.
Does it require company-specific knowledge from documents? Try Level 3 first. When permissions, freshness and source obligations start to matter: Level 4.
Is the task generic — drafting, summarising, translating? Level 1 or 2. Today. Commissioning a project for this is a waste of money.
Is a daily-running process attached? The more repetitive and high-volume, the sooner Level 4 pays off. Interestingly, MIT’s “GenAI Divide” study found the greatest returns not in the sales and marketing tools that absorb more than half of AI budgets, but in unglamorous back-office automation.
Who will run it the day after tomorrow? If the answer is “nobody”, any level above 2 is currently the wrong one.
Why this goes wrong so often
The same MIT research produced the number that has haunted every conference since 2025: 95 percent of GenAI pilots deliver no measurable return; only 5 percent reach production. What matters is the reason — not the models, but missing integration into processes and data. And a detail rarely quoted: projects with external partners scaled successfully about twice as often as pure in-house builds. Not because outsiders are smarter, but because they make the sizing mistake less often — they’ve seen the comparison cases.
The two classic mis-purchases we keep encountering are mirror images: the 80,000-euro custom project for a need the existing ERP’s AI feature would have covered; and the chat subscription that was supposed to “automate our reporting” — which it fundamentally cannot do, lacking the numbers underneath, after which “AI doesn’t work for us” enters company folklore.
The minimum set of rules — at every level
Wherever you start, three things belong in writing before the first licence is handed out. Which data may go into which tools (customer data? design data? never without clearance). Who reviews AI output before it leaves the building. And how new tools get approved, so that individual departments don’t run their own shadow AI. It fits on two pages and prevents most accidents.
Where you stand, and which level fits your case, is what our Data Readiness Check sorts out in five minutes — including the possible verdict that you don’t currently need a project at all. If so, that’s exactly what it will say.