AI in Business Development: A Tool, Not a Solution

“We can get the same thing out of AI.” What we say to that

We hear this about AI in business development in first meetings more and more often, and it is a fair objection. A prospect opens ChatGPT, dictates three sentences about the company, and a minute later has a structured analysis of the Polish market: segments, competitors, channels, risks. It looks solid. It looks exactly like the thing consultants charge for.

We will not argue that the document exists, or that it is free. We will say something else. It contains no fact that anyone verified in that specific market, and no sentence that anyone in that market confirmed. It is not market analysis. It is the arithmetic mean of everything written on the internet about markets in general.

AI in business development

Using language models to prepare business development work: research, structuring public data, language versions of materials, and analysis of buyer feedback. A tool that speeds up preparation, not a source of the decision to enter a market.

Why almost everyone uses AI and only a third sees money from it

The McKinsey global survey run in May and June 2026 (1,719 respondents, 97 countries) puts two numbers side by side. First: nearly nine in ten organisations report regular use of AI in at least one business function. Second: only 37% attribute any EBIT impact to it, unchanged from 2025. The group McKinsey calls high performers is 6% of the sample.

The difference is not the models. Everyone runs roughly the same ones. The difference is that 73% of high performers fundamentally redesigned their workflows around AI, against 25% of everyone else. Value shows up when you change how decisions get made, not when you start generating documents.

There is a second number worth holding if Poland is your target. Eurostat reports that in 2025, 20.0% of EU enterprises with 10 or more employees used AI technologies. In Poland the figure was 8.4%, third lowest in the Union. That matters in a specific way for your outreach: the person who will read your first email is reading it personally, has not automated anything, and will recognise generated text immediately because it does not look like anything else in the inbox.

AI in business development: where the generic answer comes from

A language model answers the shape of your question with whatever appeared most often in text. Ask about “entering the Polish market” and you get an average entry into an average market: check the competition, find a distributor, account for local specifics. Formally correct. Practically empty.

The problem runs deeper than prompt quality. What decides your deal is simply not in public text: who actually signs inside that company, how long the procurement cycle really takes in that industry, why the previous foreign supplier left, what margin the market tolerates this year. That knowledge sits in twenty people’s heads and moves by conversation.

So it does not surprise us that models fail systematically even on public facts. The EBU and BBC study published on 21 October 2025 (over 3,000 responses, 22 media organisations, 18 countries, 14 languages) found that almost half of AI assistant answers had at least one significant issue, a third had serious sourcing problems, and a fifth contained major accuracy errors including hallucinated or outdated information. That is on a subject with thousands of verified sources. Your niche B2B market has none.

What AI in business development is genuinely good at

We use AI daily, several models, and deliberately a mix of them. It is a strong tool. Here is where it actually buys us time:

• First-pass desk research. Building a list of companies in a segment, structuring public data, getting a rough picture of an industry in hours instead of a week.
• Language. Fast drafts of materials in Polish, Ukrainian and English, then read by a person who sells in that language.
• Structuring feedback. After twenty-five conversations, a model is good at surfacing repeated patterns across notes.
• Meeting preparation. Hypotheses, likely objections, the questions worth asking.
• Fact-checking in reverse. The most valuable use is not “generate an analysis” but “here is what the market told us, find where it contradicts public data.”

The order is what matters. AI at the input gives you a draft the market then verifies. AI at the output, where the draft becomes the action plan directly, is where companies lose money.

What AI does not see

It was not in the room. It did not hear the tone change when the price came up. It does not know that the technical director nodded and procurement decides. It did not see the competitor’s equipment already installed on their floor. It has no access to what someone told you in the corridor after the conference, and that is usually where the answer to whether this market is yours actually sits.

We travel. We do not sit at a computer producing reports. We work alongside the client’s team, approach their future buyers in the client’s name, ask the questions, and bring back what we heard together with a conclusion. Sometimes the conclusion is negative, and that is also a result. Finding out after three months beats finding out after two years.

Why the generated email stopped working

Direct communication deserves its own point. Client work conducted through AI now ends at a very early stage, because people have learned to recognise generated text. The same structure, perfectly even paragraphs, an opening compliment, “I noticed that your company.” The recipient sees it in three seconds and draws one conclusion: this sender spent no time on me at all.

This is arithmetic rather than ethics. You are spending your first and only contact with a reachable person on a message that invalidates itself. In B2B, where your market is 200 companies and not 200,000, burning a contact is expensive.

AI will write the email in ten seconds. But you get the reply for knowing something about that company their next supplier does not know, not for speed.

Dima V. Nechyporenko

Your client has AI too, and is analysing you with it

This side of it gets forgotten. The same tool sits on the other side of the table. Your proposal, your price list and your website now reach the buyer’s chat window before they reach a human desk. The buyer asks the model to compare you against two competitors, pull the weak points out of your terms, and draft the questions for the meeting.

The consequences are concrete. First, vague advantages disappear. “Individual approach” and “high quality” get scored at zero, because everyone wrote the same thing and the model compares on what is comparable. Second, anything checkable will be checked, and any gap between your website, your proposal and what you said in the meeting will surface. Third, you walk in to meet someone who already holds a list of your weakest points.

The practical conclusion is equally concrete. A proposal has to carry what does not reduce to an average: numbers, timelines, terms, the names of projects you actually delivered. And your public materials have to be written so a model reads them correctly, because that is what it will build the buyer’s picture of you from. We treat that as a separate job and do it for clients the same way we do it for our own site.

Commercial confidentiality: what you cannot upload

There is one more limit, rarely mentioned from conference stages. The most valuable material in your preparation is exactly what cannot go into a public chat: cost base, supplier terms, live contracts, customer data, technical solutions.

The scale of this has been measured. The LayerX Enterprise AI and SaaS Data Security Report of October 2025 found that 77% of employees using AI chats paste data into them, and 82% of those pastes go through unmanaged personal accounts, outside any company control. Around 40% of uploaded files contained personal or payment data.

Our practical rule is simple: split the circuit. Public data and hypotheses go to the model. Commercial information goes nowhere outside your perimeter and the people covered by an NDA. That is how we work, and part of our work with a client physically cannot be done by a model for this reason.

Language decides more than people expect

Yes, English will get you a conversation with almost anyone today, and for a first meeting it is usually enough. But we see the difference every time. When the conversation runs in Polish with a Polish buyer, or in Ukrainian with a plant director, you hear noticeably more, and you hear it in different words. In particular you hear the things nobody says to a foreigner: what is genuinely wrong with the incumbent supplier, and how much room there really is in the price.

So we work where we speak. Our working languages are Polish, Ukrainian and English, and we do not take markets we are not present in. Machine translation does not close this gap, because the gap is not in the words. It is in how candid someone is willing to be with you.

AI is a system of fine settings, not generic ones

This is how we frame AI in business development internally. A model produces value when you tune it to one specific market and one specific buyer, iteratively:

1. State the hypothesis: who needs you in this market, and why.
2. Test it in live conversations, not in search.
3. Bring the market’s answers back and rewrite positioning, materials and the target list around them.
4. Only then scale what was confirmed.

AI in business development speeds up steps 1, 3 and 4. Nobody speeds up step 2, and step 2 decides whether the rest means anything.

Where we are most useful

We tell clients plainly that we add the most value at two stages: market entry discovery and entering a new market. That is when you are still working the market out and are supposed to be listening to it, not when you need to scale a sales machine that already runs.

At those stages our value, and the value of AI in business development around it, is applied: verifying what you generated in your own office before you run at the market and make the mistakes many companies are making right now on unverified information. In practice it looks like this. A target company list agreed jointly with you, approach in your name in the local language, feedback collected and handed over, first meetings, and then a decision. Adapt the offer, or leave this market honestly.

What this looked like in practice

On the Voltage Group project we ran this cycle across four EU markets: research, first market contact, matchmaking, first meetings. The outcome was more than 65 MW of projects in 12 months, with the first contract signed three months after the start. No model would have written that result, because it was made of specific people, specific conversations, and decisions taken on what those people said.

The second example is a different shape. Ostapiv Dachy was already operating but wanted better sales processes and new segments. We ran the market research, chose the strongest customer segment, built the materials, and launched the B2B sales development process. Again, the segment choice came out of conversations, not out of a report.

What to do with this

If you are looking at a new market and already have a generated analysis, do three things. Pull ten specific claims out of it. Mark the ones you can attribute to a named person who told you. Delete the ones nothing supports. What remains is your real knowledge of that market today.

If very little remains, that is normal. It means you are at the discovery stage, and the next step is not writing a better prompt. We describe the process itself on the market expansion page.