How Cresques uses AI to find distributors — and where a person still decides

"AI-powered" says nothing about what a tool actually does. This page says where Cresques uses a language model, what it sees, what it cannot do, and which decisions stay with you.

Putting it simply: Cresques uses AI for three narrow jobs — researching whether a company named in public contract award notices is a distributor or a manufacturer, reading your materials to propose procurement categories, and drafting a first message in the company's language. Every company on a list comes from a public notice you can open. You confirm the categories, decide whom to approach and send every message yourself, from your own mailbox.

Where the data comes from

Every list in Cresques starts with a public document, not with a model's memory.

When a public buyer in the EU awards a contract above the EU thresholds, it must send a contract award notice within 30 days. The notice names the successful tenderer with its address, says whether it is an SME, gives the value of the winning tender and carries CPV codes (Directive 2014/24/EU, Article 50 and Annex V Part D). These notices appear on TED, the online Supplement to the Official Journal of the EU — more than 3,000 notices every working day.

CPV is the Common Procurement Vocabulary: one EU classification, eight digits plus a check digit, mandatory in procurement notices. So one notice tells you who bought, who won, and in which category.

What the models do

1. Distributor or manufacturer?

A company that wins a furniture tender is either a maker (your competitor) or a reseller (your potential partner). The notice rarely says which. When you open a company profile, a large language model researches what is public about the winner — the notices, the company name, its website — and proposes a role: independent distributor, manufacturer, manufacturer's subsidiary, or both. The role comes with verbatim quotes, and each quote is checked on its source page; if no quote can be confirmed, the role stays "unknown".

2. Products to categories

A language model also maps product wording to CPV categories, so that "height-adjustable desks" lands in the codes public buyers actually use. This is the materials reader below.

The materials reader

You can hand Cresques a brochure (PDF) or your website address. The reader extracts the text and an AI model (Claude) proposes the matching CPV categories — "office chairs" in the brochure proposes the category "Office chairs". It is free, and it is a proposal. You tick what is right and remove what is not before a market is built.

What the AI does not do

It does not invent companies — every company has a notice behind it. It does not send anything. The first message it drafts, in the company's language with an English version and for the channel that is lawful in that country, is a draft: you edit it and send it yourself.

What a person confirms

• The categories. You confirm the CPV categories for your market, from the materials reader or from search.
• The country. A market is one country plus your categories.
• Whom to approach. You read the evidence behind each company and decide. The label is a sorting aid, not a verdict.
• The first contact. You review the draft and send it yourself, in the channel that is lawful in that country.

How accurate is the classifier?

Honestly: not perfect, and we do not publish an accuracy figure we have not measured on the current method. What we do instead is make every label checkable: it comes with two to four verbatim quotes, each found again on the page it came from, a date and a confidence level — and when no quote can be confirmed the label stays "unknown" (see our methodology). That is why the label sits next to the notice and its quotes rather than replacing them.

NIST describes confabulation as "the production of confidently stated but erroneous or false content" by generative AI, and a 2024 paper in Nature notes that large language models "fluently make claims that are both wrong and arbitrary". We limit the damage by giving the model small tasks — propose a role, never invent a company — by keeping a role only when its quotes are confirmed on the source pages, and by linking every result to its source, so a wrong label is easy to spot.

What it costs

Every AI call is recorded: the model used, tokens in and out, and the cost in US dollars, attributed to the action and to your organisation. Each action has a fixed price in credits that you see before you run it — 4 credits to open a company profile (once per workspace), 1 credit for a drafted first message; the materials reader is free — with the real cost underneath on our side. No hidden AI spend.

Personal data and sub-processors

Award notices often name a contact person with a direct e-mail and phone number. Under the GDPR, information about an identifiable natural person is personal data (Article 4(1)). Cresques removes contact-person names, personal e-mail addresses and phone or fax numbers before a notice is stored. Generic company mailboxes such as info@ stay, because they are company data.

Our privacy policy names Anthropic (US) as the AI sub-processor, with standard contractual clauses or the EU–US Data Privacy Framework for the transfer. We do not send our users' account data to AI providers.

Where the law stands

The EU AI Act takes a risk-based approach. Its list of high-risk uses (Annex III) covers biometrics, critical infrastructure, education, employment, access to essential services, law enforcement, migration and justice. Sorting public companies by public evidence is not on that list, and the Commission's guidance is that minimal-risk systems carry no specific obligations. We still tell you where AI is used.

GDPR Article 22 gives individuals the right not to be subject to a decision based solely on automated processing that produces legal or similarly significant effects for them. Cresques labels companies, not people; a person decides whom to contact.

What this means for a manufacturer

• Ask any AI tool what the evidence is. If the answer is "the model", treat the list as a draft.
• In Cresques the evidence is a public notice with a date, a buyer, a value and a category. Open it.
• Expect the distributor/manufacturer label to be wrong now and then. Check it before you invest time.
• You confirm categories, choose companies and send messages — by design, because that is where your market knowledge counts.
• AI costs are measured per action and paid in credits. Budget them like any other line.