Choosing an AI assistant for investor relations: the questions to ask any vendor

In short

Every demo in this market looks impressive, because fluent answers are the easy part. The differences that matter for a listed company sit underneath the fluency. What is the system allowed to answer from? What does it do when it does not know? And what does the IR team get besides answered investors? This is the question list we would put to any vendor, ourselves included, grouped by what each question reveals.

Key facts

• Fluency is table stakes. The real differences are the source boundary, what the system does when the information is absent, and citations.

• The most revealing minute of any evaluation is asking a question the company’s materials do not answer.

• Beyond compliance, question visibility, languages and the setup model decide the operational value.

• A serious vendor will also tell you what the product deliberately is not.

Which questions expose the source boundary?

Start here, because everything else depends on it. Ask what exactly the assistant can answer from, and who controls that set. The safe answer is a defined body of the company’s approved information, controlled by the issuer, with a clear route for material to enter and leave it.

Then ask whether it can ever reach outside that set, to the open web or its own training data, and under what circumstances. Any ambiguity here is the finding.

Then ask how the set stays current. Is it uploads only, or also a feed from your website and integrations with regulatory sources? And who verifies what entered?

The boundary is what determines the disclosure risk. The compliance logic behind these questions is covered in is an AI assistant on your IR website a disclosure risk.

What is the edge test, and why is it the most revealing minute?

Ask the assistant, live, a question you know your materials do not answer. A valuation multiple you never published will do, or a forecast you never gave. There are only two possible behaviours. Either it says plainly that the published information does not cover it, or it produces something plausible.

The second behaviour is the failure, however eloquent it sounds. A system that improvises under pressure in a demo will improvise in front of an investor. Run the test more than once, and rephrase the question each time. What you are checking is whether it refuses consistently.

Pair it with the citation check. Ask a question the materials do answer, open the citation, and verify that it lands on the passage that supports the answer, not just on a document. Citations that resolve to real passages are what make the system auditable after the fact.

Which questions reveal the operational value?

There are three areas.

Start with question visibility. Does the IR team see what investors ask, and in what form? The questions are a market signal. A system that answers investors without showing IR the questions leaves the more valuable half of the product on the table. Why that signal matters is covered in what are investors actually asking.

Next, languages. Can investors ask and get answers in their own language while the source documents stay as published? And does the citation survive the switch of language? For an international register this is daily value, not a feature checkbox.

Last, setup. What does implementation involve, who does it, how long does it take, and what does the issuer maintain afterwards? Distrust both extremes. “Instant” suggests nobody is checking what enters the source set. A months-long project suggests you are buying consulting, not a product.

What should the vendor say the product is not?

A serious answer to this question is a strong signal in itself. An investor-facing answering assistant is not a data room. It is not a document or engagement tracker, not a file-sharing system, and not an analytics suite bolted onto a chatbot. A vendor who claims everything is describing a roadmap, not a product, and roadmaps do not pass compliance reviews.

What a product refuses to be is usually where its discipline shows. It is the same discipline you test with the edge test, applied to the company itself.

For the category underneath all these questions, see what is a source-grounded IR assistant. It explains what a source-grounded assistant is and why the constraint is the point. MIRA is such an assistant, and we are happy to be evaluated against every question on this page.

The bottom line

Evaluate an AI IR assistant on what sits under the fluency. Look for a source boundary the issuer controls, honest behaviour when the information is absent, and citations that resolve to real passages. Check that the IR team sees the questions, that answers in other languages keep the citation intact, and that the setup sits somewhere between “instant” and “endless”. And look for a vendor who can say what the product is not. One live edge test tells you more than any deck.

Want to run the edge test yourself? Book a MIRA demo and bring a question your published materials do not answer.

About the author: Sander van Prooijen is Head of Commercial Strategy at Finvictum, where he leads the European rollout of MIRA. He works with IR teams across Europe on how investors, and more and more AI tools, find and read the equity story. Connect on LinkedIn.