In 1999, every company needed a website. Nobody quite knew what for. Agencies were booked out, budgets were approved on a feeling, and a domain name was worth more than a business plan. Then the bubble burst.
The websites didn’t disappear. What disappeared were the projects without a purpose. AI today feels very similar — and the lessons from back then are surprisingly practical.
The web survived. The projects without a purpose didn’t
After the dot-com crash, the internet did not go away. It grew. What survived were websites with a concrete job: the shop that actually sold, the catalogue customers actually used, the form that replaced a fax. What vanished were the projects that existed because everyone else had one.
1999: the website
Every company needed one. Few knew what for.
Today: AI
Every company needs it. Few can say what it does on Monday morning.
AI will stay, just like the web did. The question for every company is not whether to use it — it is whether its own AI project is one of the survivors. And that is decided much earlier than most people think.
Three months — then it has to work
If your AI project is not used productively after three months, you have not built a product. You have burned money.
Not because AI doesn’t work. Because nobody decided what “works” means. The pattern is always similar:
- A pilot that only runs in the demo — with hand-picked data and a presenter who knows which questions to avoid.
- A new prompt every week — but no real users who depend on the result.
- “We just need more data” — a sentence that can justify any delay, forever.
- A growing budget with a vague goal — every month costs more, and nobody can say what would count as done.
The fix is boring: one task, one user group, one measurable result — and a fixed date.
Stopping is not failure. Continuing without a result is. A stopped pilot leaves you with a clear lesson and the budget for the next attempt; an endless one leaves you with neither.

AI can calculate. It cannot steer
AI is very good at code. It is very good at mathematical relationships: formulas, unit conversions, variants, test cases, the hundred small checks nobody enjoys writing. In these areas it is faster than any of us — and we use it every day.
What it cannot do is steer. It does not know which customer matters more, which risk you are willing to take, which rule is a law and which one is just habit, or where your company wants to be in five years. AI answers the question you ask. Choosing the question is still your job.
“Upload everything and let AI decide”
Whoever says this has already lost.
Nobody can give AI everything. Not the experience of your sales engineer. Not last week’s phone call with the customer. Not the reason why a rule exists in the first place — the failure ten years ago that nobody wrote down. So AI decides on what it got, and it sounds just as confident about what it didn’t get.
That is the dangerous part: the answer looks complete. The gap is invisible. A wrong decision that is obviously wrong gets caught. A wrong decision that looks well-founded gets signed.

- Decisions stay with people — AI prepares, compares and checks; a named person decides.
- Rules stay explicit — product logic belongs in data that can be read, tested and versioned, not in a model nobody can inspect.
- Results stay traceable — every number in a quotation must be explainable to the customer, without “the AI said so”.
How we use AI at digirom
We build selectors and configurators for manufacturers. AI makes us faster: it writes code drafts, checks formulas, generates test cases and finds inconsistencies in large product data sets. The product logic, the priorities and the final “yes” stay with people — ours and our customers’.

And every project has a date by which it runs in production. Not a pilot, not a demo: a tool your sales team and your customers actually use.
AI project without a result? We build configurators that are in production within months — with AI where it helps, and clear rules where it matters.
