AI is useful to a startup as an accelerator for groundwork: texts, document structure, market research, code, customer support.
But it does not analyse or take responsibility for results — it produces the most probable, averaged answer.
So quality depends on how the task is set and on the expert who checks the numbers and conclusions. Without oversight, AI multiplies templates and errors.
Artificial intelligence was one of the most discussed topics at the European Business Summit in Alicante.
In almost every session — on launching a project, promotion, costs or scaling — AI came up either as a universal recipe or as a threat.
We decided to look at it without illusions: as a tool whose limits matter as much as its capabilities.
Below: what mass AI changes, why its answers are averaged, where it really helps a startup, how to set it tasks and what investors notice in documents written by a neural network.
What mass AI changes
- Ideas lose value. An idea used to be seen as an asset; now any neural network produces one in seconds. That does not make ideas worse, but it makes them common: if an idea is available to everyone, it stops being a competitive advantage. More in How to prepare a project for investment.
- Copywriting gets cheaper. Generated texts have no errors or lapses in logic, but no style, authorship or recognisability either. More and more clients ask: if the text is equally unconvincing, why pay for it?
- Content becomes uniform. Decks, websites, “about us” blocks, template value propositions — all recognisably “AI” and less effective than before. User attention drifts away.
- Live formats matter more. The more available automatic text becomes, the higher the demand for the original: video, live conversation, real-time interaction. AI is arriving there too, with a “plastic presence” effect.
- Digital sameness sets in. Texts and images are sterile and smooth, with no individual trace, and it is ever harder for a business not to look like everyone else.
Why AI answers are averaged
A neural network is neither a conversation partner nor an expert. It is a language model: it continues text in the most probable way based on a vast body of examples.
Hardly anyone said this directly at the summit, and most users do not realise it.
Most probable is not the same as best
The model gives not the deepest answer but the most typical answer to that kind of question.
That is why even a complex, precisely worded question often gets a template, shallow, predictable answer: depth requires atypical moves, and by default the model takes the beaten path.
Documents are read unevenly
When you attach a file or a spreadsheet, the model does not check every line as an analyst would.
In long documents it can miss details, contradictions and exceptions — and fill the gaps with a plausible guess.
Hence factual inaccuracies and confident-sounding errors, especially in finance, law, project management and corporate structure, where details decide.
The answer is no smarter than the question
If the question is vague and casual, so is the answer. There is a fair observation that a well-posed question already contains half the answer.
Most users do not structure the task, give no context or data and set no criteria or limits — and get an “averaged” answer: literate, logical, safe and useless in practice. That is how the illusion of a result is born.
Mid-level content irritates experts
The model lowers complexity so that most people can follow: it makes no bold assumptions, forms no hypotheses and does not go deep.
An expert sees generalisations, ignored details and repeated truisms — rounded wording instead of analysis, something like a school essay instead of an answer.
That is fine for a draft or a discussion; on a professional’s desk such text causes rejection.
Caution instead of assessment
Models are tuned to avoid sharp wording and offend no one.
As a result they are reluctant to call things by their names: a weak solution becomes “an alternative approach”, a mistake “a matter of interpretation”.
For a business this is dangerous: vague wording makes it impossible to draw conclusions and decide.
If you need an assessment, ask for it explicitly — with criteria and a request to name the weak points.
The conclusion is simple: neural networks can and should be used, but you need to understand how they work.
AI is not knowledge or thinking but a technology for generating plausible text. In analytical tasks without expert oversight it is not a tool but a source of distortion.
Where AI helps a startup and where it hurts
| Task | Where AI helps | Where a person is needed |
|---|---|---|
| Market research | A list of competitors, segment hypotheses, interview questions | Checking sources and numbers, talking to customers |
| Concept and business plan | Structure, section drafts, editing | Business logic, choice of segment, positioning |
| Financial model | A structure template, formulas, typo checks | Assumptions — prices, volumes, timing, costs — and their justification |
| Marketing and copy | Headline options, adapting to channels, translations | Brand voice, the author’s position, final editing |
| Product and development | Prototypes, code, tests, documentation | Architecture, security, responsibility for the result |
| Sales and support | Answers to typical questions, first qualification of enquiries | Complex negotiations, exceptions, key customer relationships |
| Legal and tax questions | A first overview of the topic, a list of questions for the specialist | Any decision — only with a lawyer and tax adviser |
How to set AI a task
- Role and goal. Who the model should act as and what you need: “assess as a sceptical seed investor whether this is worth investing in”.
- Context. Project stage, market, country, customer, constraints.
- Data. Real numbers and documents, not a request to “make up plausible ones”.
- Criteria. What to assess by, what counts as an error, what must be checked.
- Format and verification. A table, a list of risks, links to sources; a request to list assumptions and what the model is unsure about separately.
And the main rule: check every number and every fact in the answer against the original source.
What investors notice in AI-written documents
- Smooth text with no specifics: “an innovative solution for a wide range of users” instead of a named segment and numbers.
- Market figures without sources, or with sources that do not exist.
- A model whose assumptions are not linked to the sales plan and team: revenue grows while customer acquisition costs stay flat.
- Identical template sections — SWOT, risks, competitors — that would fit any project.
- Inconsistencies between the deck, model and business plan because each was generated separately.
Investors are not against AI — they are against documents that show no team thinking behind them.
So documents prepared with a neural network are worth an independent check before the meeting: do the numbers match, are the assumptions justified, is the business logic visible.
Promising applications: three examples from the summit
The summit featured commercial solutions already in operation. They show that with a well-set task a neural network becomes not just an assistant but part of the business model.
- Digital teachers. Lectures by leading teachers are digitised and avatars are created with their pace, intonation and manner. The video feels like a “live” teacher — the education product scales and costs fall.
- Sales chatbots. Bots identify needs, handle objections, adapt their style, work round the clock and stay in sync with the catalogue — sales grow without adding staff.
- Subscriptions to expert bots. Paid assistants in legal, medical, educational and other fields, tied to specific protocols and knowledge bases: expertise, trust, monetisation.
Frequently asked questions
Can you write a business plan with AI?
A draft — yes: AI helps with structure and text. But the business logic, segment, financial model assumptions and market figures must be yours and checked against sources, or an investor will quickly spot the template and the inconsistencies.
Why are AI answers shallow?
A language model produces the most probable continuation of text — that is, a typical answer. Depth appears when you provide context, data and criteria and ask it to name assumptions and weak points.
Can you trust numbers from AI?
Not without checking the original source. A model can confidently give a plausible but wrong figure or a non-existent source, especially in finance, law and market statistics.
Where does AI help a startup most?
In groundwork and repetitive tasks: document structure, text options, translations, prototypes and code, answers to typical customer questions. Decisions, numbers and responsibility for results need a person.
Will an investor notice documents were written by AI?
Often, yes: smooth text without specifics, numbers without sources, template sections and inconsistencies between documents. AI itself is not a minus; the minus is the absence of the team’s thinking behind the documents.
Key points on AI for startups
- AI speeds up groundwork but does not think or take responsibility for results.
- A model gives the most probable answer, not the best — hence the templates.
- Answer quality depends on the task: role, context, data, criteria, format.
- Check every number and fact against the original source.
- Investors look for the team’s thinking behind the documents — a neural network cannot replace it.
