What AI can and cannot do for exporters

Where an assistant helps, and where you need a second source

An assistant is useful to an exporter for some jobs and risky for others. The line runs between work you can read and judge yourself, and facts about the outside world that you cannot judge from the text.

OpenAI's safety guide for developers recommends, wherever possible, having a human review outputs before they are used in practice. It adds that the reviewer should have easy access to the information needed to verify the output.

What the providers' own documents say

Anthropic's Claude Help Center says Claude can occasionally produce responses that are incorrect or misleading, which it calls hallucinating. The same page says Claude may not have been trained on up-to-date information in some subject areas, can show quotes that look authoritative but are not grounded in fact, and should not be relied on as a singular source of truth.

Google's Gemini help pages give the same warning: Gemini can make mistakes, so double-check its responses. Its page on responses adds that Gemini may give inaccurate or inappropriate responses about people. A buyer's name and job title are that kind of fact.

A paper by researchers at OpenAI and Georgia Tech, “Why Language Models Hallucinate”, says language models sometimes guess when uncertain and produce plausible yet incorrect statements. In its opening illustration an open-source model was asked for a researcher's birthday and told to answer only if it knew; across three separate attempts it gave three incorrect dates. When three popular models were asked for the title of the same researcher's dissertation, without web search, none produced the correct title or year.

Web search does not remove the risk. The paper argues that search reduces hallucinations but is not a cure, because binary grading, which gives no credit for “I don't know”, still rewards a guess whenever search fails to give a confident answer. Anthropic's help page tells users who work with web search results to review the cited sources, since original websites may hold context that the summary leaves out.

A separate Anthropic help page says Claude can produce links that do not work and can falsely claim that it has sent emails or produced external documents. According to that page, Claude has no access to tools that are not explicitly integrated, email included.

Buyer search: what to check, claim by claim

Verify one company by hand

1. Company, website. Type its name into a search engine yourself and find the official website; ignore any link the assistant gave you. On the products or services page, check that it actually buys, processes or distributes what you sell.
2. Company, details. Compare the name, address and phone on the About and Contact pages with what the assistant said.
3. Company, register. Search the business register of its country for the name. Do the registered name, address and status (active, in liquidation, dissolved) fit the website?
4. Company, directory. Look for it in a business directory such as Europages, or in the exhibitor list of a trade fair you would attend. Do the address and domain match?
5. Person. Check the decision maker on the company site or on the person's public profile. If the title or company does not match, drop the contact.
6. Email. Its domain should match the website's domain. If an address is published on the site, use that one; if only a guessed address exists, confirm it from a second source before you send.
7. Result. If any check fails, remove the row. Do not write to a company you could not verify.

For the register check, use the business register of the buyer's own country. The United Kingdom's Companies House is one public example: it offers a search by company name, company number or officer name, and its page states that Companies House does not check the accuracy of the information filed. A register entry is therefore a good first sign that a company exists, not proof of everything else, so set it beside the website and a directory.

Check market numbers at the source

Start from your product's Harmonized System (HS) code and look the market up in Trade Map and UN Comtrade. If an assistant has quoted an import or demand figure for that market, compare it with what you find there. If they differ, do not use the assistant's number; if it cannot show where its figure came from, keep it out of your report and your emails.

The overview page of Trade Map, from the International Trade Centre (ITC), says the tool provides indicators on export performance, international demand, alternative markets and competitive markets. UN Comtrade publishes imports and exports by product and trading partner, as reported by countries and areas to the United Nations and classified the same way so they can be compared.

Drafting and translation without invented details

Anthropic's documentation says that explicitly allowing the model to say it does not know can reduce false information, and that asking it to cite quotes and sources makes its answer auditable. The same document says these techniques do not eliminate hallucinations and that critical information should always be validated. So tell the assistant not to invent details, and to mark what is missing.

Where Alvara AI fits

Alvara AI is a tool for manufacturers and exporters that finds buyers abroad: enter your website address and it lists matching overseas companies with their decision maker and a verified email. It gets buyer details from each company's own website, from data providers, and from inferences it draws from both; next to every fact there is a link to its source, so you can check it yourself. It does not send email on its own; it writes drafts, and the decision to send and the sending are yours.