AI in IR Roundtable event: what we learned

Here are some of our key takeaways:

1. Start with the problem you're trying to solve, not the technology.

One of the most useful examples was also one of the simplest. Rather than deciding to “build an agent”, start with the IR problem and ask AI itself how best to solve it. It may recommend an agent, Notebook, another Microsoft tool, or something much simpler. Better still, ask the model to interview you until it properly understands the problem. The skill is increasingly less about writing the perfect prompt and more about conducting a productive dialogue.

2. Use cases abound for efficiency around common IR pain points

One team started with repetitive ESG questionnaires: consolidate the source material, let an agent generate the answers, then review. The result was reportedly about 99% complete before human intervention. Other examples included: saving 30-40 hours by using AI to automate the process of updating the website from the annual report; consistency checking across the suite of reporting documents; sentiment analysis and peer results summaries. 

3. Use AI to create more time for judgement, not simply more output.

A helpful framework from the discussion was to think about three areas in which AI can be deployed: content creation, process efficiency and data synthesis. But the objective is the same across all three – to free up more time for the IRO to spend doing what they are paid for: exercising judgment.

4. AI may be a better editor than author.

There was healthy disagreement about AI-written content. Models can increasingly learn executive tone. One participant had trained a model to write in the voice of the CFO through several years of earnings transcripts, another used it for executive quotes in earnings releases. But there was also a warning which chimed with many: let AI create the first draft and you can become trapped inside its framing, tinkering with its answer rather than doing the thinking yourself. 

5. You are increasingly communicating with machines as well as humans.

Good writing still matters, but a beautifully nuanced outlook statement may not survive contact with an LLM summary. Neither will the distinction between the IR website, sustainability report, corporate press release and something published by another part of the organisation. The machine sees the corpus, not the org chart and this is leading some companies to be much more co-ordinated across functions in the production of company literature and messaging. Other ways in which IROs are adapting include: greater use of HTML, PDF metadata and machine-readable content formats; and a focus on information architecture to ensure LLM visibility.

6. Your investors may encounter - and trust - an AI interpretation of your story before they encounter yours.

Research cited at the event suggested 68% of investors say AI has changed their approach to earnings calls, with 40% preferring an AI summary to attending one. The room supplied the real-world evidence: an investor using a flawed LLM answer as the basis for challenging a company; a little-read and factually wrong article being recirculated and amplified by an LLM, causing a sharp share-price move; and an investor-facing AI tool confidently misrepresenting a company's exposure to a major risk. Companies have always competed with external interpretations of their story. What changes with AI is the speed, scale and apparent authority with which those interpretations can now be synthesised and propagated.

7. Your own data may be your biggest AI advantage.

IR holds an unusually rich dataset: CRM history, shareholder registers, meeting feedback, website behaviour, trading information and document engagement. No individual supplier necessarily sees the whole picture. One of the more intriguing opportunities discussed was using AI to synthesise those sources for targeting and perception analysis rather than relying on a supplier's black box. Another related idea was to have a ‘sentiment agent’ talking to an ‘equity story agent’ to stress-test investor messaging. 

8. AI governance needs to enable experimentation — and it is becoming an IR issue in its own right.

Several participants described a tension between employees wanting to use AI and internal policies that were either highly restrictive or poorly understood. In the most extreme case, blocking access simply pushed experimentation onto personal devices, arguably creating more risk rather than less. At the same time, governance is starting to move onto the other side of the IR desk. Participants reported the first investor questions about how companies themselves govern AI, with the AI Corporate Disclosure Initiative mentioned as one emerging reference point.

9. The more AI develops, the more some very human IR skills matter.

The discussion kept returning to judgement, trust, relationships and networks. An LLM can process vastly more information than an IRO. But it cannot have the private conversation that explains the missing context, understand every stakeholder relationship, or replace the credibility built between management, IR and investors over time.
That has implications for the Society too. Members asked for fewer generic exhortations to “go and experiment” and more practical help: real IR use cases, simple guides to building agents, live demonstrations and direct exposure to how investors are using these tools. That will be a major focus for the AI Steering Group going forward, including at our annual in-person session on 10 November. The technology will keep moving. Our job is to help IROs work out what is genuinely useful, what is changing around them, and what good IR looks like as a result.