AI in Investor Relations 2026: Earnings Prep Without the Fire Drill

AI in investor relations stopped being a conference panel topic somewhere around the start of 2026. According to Nasdaq's seventh annual Global IR Issuer Pulse survey, 51% of IR professionals have embedded AI into their processes, compared with 30% in 2024 and less than 10% in 2023. That is a small function, usually two or three people at a mid cap, changing how it works in about twenty four months.

What makes the IR case worth studying is that the constraints are unusually sharp. An IR team drafts under embargo, speaks under Regulation FD, and answers to a regulator that has started reading what companies say about their own AI. So the useful version of this article is not a list of tools. It is a map of which parts of the earnings quarter a model can honestly take, which parts it can assist with under supervision, and which parts stay in human hands for reasons that will not change with a better model.

1. The Quarter IR Stopped Doing It by Hand

The adoption curve in IR is steeper than in most corporate functions, and the reason is structural. IR runs a fixed calendar with four hard deadlines a year, produces mostly text and slides, and consumes enormous amounts of public information written by other people. That is a near perfect profile for language models, and the numbers show it.

The forward looking numbers point the same way. In the same Nasdaq survey, 72% of respondents said AI and technology advancement could deliver meaningful benefits in 2026, and 32% of the most innovative IR efforts already implemented or planned for the year were AI related, including investor engagement hubs, custom GPTs and sentiment scoring. The barrier list is equally revealing: data privacy and security sits at the top with 22%, followed by limited understanding at 15%.

That barrier ranking is the honest part of the story. IR did not slow down because the drafts were bad. It slowed down because the material in question is often price sensitive and unreleased, and nobody wanted to be the person who put next quarter's revenue into a text box before the release crossed the wire. Any serious workflow has to answer that question first, which is exactly what section five is about.

2. Where the Earnings Quarter Actually Goes

Research from IR Impact puts 25% of an IR professional's time on earnings preparation, ahead of roadshows and investor meetings at 19%. A quarter of the year on a process that repeats four times with the same shape is the clearest automation signal a function can send. In July 2026, the IR leader at Lear Corporation told IR Impact that AI tools shaved a week off the earnings process, applied to targeting, analysis and executive readiness rather than to the disclosure itself.

A week is a lot when the whole cycle is three or four. But the saving does not come from one heroic prompt. It comes from removing the small research and assembly tasks that pile up in the two weeks before the call and that nobody schedules because they are supposed to be quick.

Read that table as two different jobs wearing one name. The top half is research and assembly, which is public, repetitive and easy to check. The bottom half is disclosure, which is legally consequential and where the value of a human sign off is the entire point. Teams that got a real week back moved the top half and left the bottom half alone.

3. The Three Jobs AI Already Does Well

The Nasdaq use case ranking is unusually concentrated. Summarizing peer and market events leads at 81%, earnings preparation follows at 70%, and internal management reporting sits at 36%. Three jobs, and all three are the same underlying task: take a large pile of public text, compress it without losing the parts that matter, and hand a decision maker something short.

Peer and market summaries won first because the failure mode is cheap. If a summary of a competitor's call misses a nuance, someone catches it in the read through and nothing leaves the building. Doing it manually means five or six transcripts a quarter, each an hour, produced during the exact week the team has the least time. This is the same shift we described in our piece on enterprise market intelligence after analyst reports, arriving in a function that already lived on other people's disclosures.

Earnings preparation at 70% is the surprising one, until you separate preparation from disclosure. Preparation is the briefing book, the executive readiness session, the anticipated question list, the fact pack for the CFO. None of that is published. All of it is assembled from material the team already has.

Internal management reporting trails at 36%, and the gap tells you something about trust. Reporting up to the executive team means your name is on a number that someone will repeat in a meeting, so teams move slower here. The same caution shows up in our guide to AI board reporting, where the constraint is never the drafting, it is the traceability of every figure back to a source the board can check.

4. The Q and A Book Without the Invented Premises

The Q and A book is the single most useful place to put a model, and the single easiest place to hurt yourself with one. Useful, because generating a hundred plausible analyst questions from prior calls, peer calls and published commentary takes a model about a minute and a team about a day. Dangerous, because the same fluency that produces a good question produces a confident answer to it, and a confident wrong answer read aloud on a live call is a disclosure problem, not a writing problem.

There is a second failure worth naming. Models are good at inventing the premise of a question. Ask for likely analyst questions and you will get several that assume a product launch that never happened, a guidance range you never gave, or a margin trend from a peer's business. Those are not harmless. They pull rehearsal time toward scenarios that do not exist, and they occasionally survive into the book, where an executive rehearses an answer to a fact pattern that is not theirs.

The fix is boring and it works: require a citation on every generated question. If the question came from a real analyst note, a real prior call or a real peer transcript, the source is attached. If it came from the model's imagination, it is labeled as such and treated as a stress test rather than a forecast. This is the practical version of what we cover in human-in-the-loop AI, and it is also the difference between a briefing book and the polished emptiness we called out in the anti workslop playbook.

5. What Never Gets Delegated

Three things stay human, and none of them are about model quality.

Material non public information. Unreleased results, draft guidance and merger discussions belong in the controlled disclosure workflow the company already runs, with its access list and its audit trail. The question to ask about any AI tool in the earnings process is not whether it is smart, but where the text goes, who can read it, and how long it is kept. If a tool cannot answer that, it does not touch pre release material. That is a governance rule, not a vendor preference, and it applies to the sanctioned tools as much as to the ones that arrive quietly, a pattern we mapped in our piece on shadow AI.

Regulation FD judgment. Deciding whether a response to an investor question crosses into selective disclosure is a legal call about materiality and audience. Models are fluent about securities law and unreliable about the specific facts of your company, which is the worst possible combination for a judgment where being confidently wrong is the failure.

What you say about your own AI. This one caught companies by surprise. The SEC's Division of Examinations named AI a priority area and said it will review the accuracy of registrant representations about their AI capabilities, and the staff has been issuing comment letters asking for detail on development, validation, third party dependencies and the real operational role of the technology. On the enforcement side, the agency has already charged companies over false AI claims, including Presto Automation in early 2025 over statements about its AI product. The lawyers at Norton Rose Fulbright summarize the expectation simply: statements about AI usage or proprietary AI technology have to be precise, and vague or exaggerated claims are the exposure.

A guest analysis in The D and O Diary on the 2026 reporting season lists what a well built AI disclosure now addresses: what the company means by AI, board oversight, internal operations versus customer facing products, data provenance and usage rights, IP and licensing risk, and vendor governance. Notice the irony for an IR team. The same quarter you start using a model to help write the script is the quarter the script itself has to describe your AI accurately. If you are working through that language, our guide to AI due diligence covers the evidence you need before the claim goes on paper.

6. An IR Workflow That Survives an Audit

The workflow that holds up is not complicated. Every generated artifact carries its sources. Every draft that reaches an executive has a named human owner who read it end to end. The pre release perimeter is drawn once and enforced by policy rather than by good intentions in a busy week. And the outputs the team actually needs, which is a brief, a Q and A book, a deck and a page the whole team can read, get produced in one place instead of six.

The deck deserves its own mention. Investor presentations are rebuilt four times a year from a narrative that changes slightly each time, which is why they consume so many hours for so little creative gain. We went through that specific problem in detail in AI enterprise presentations, and the IR version is the cleanest case for it, as long as the numbers on the slides come from the released figures rather than from the model's memory.

7. A One Quarter Rollout

IR runs on a calendar, so the rollout should too. One quarter is enough to get real value without touching the disclosure path, and it gives you a before and after on the same process rather than a vague sense that things feel faster.

The metric that matters most is the last one. Track the corrections, not just the time saved, because the correction log is what tells you whether the drafts are getting better or whether the reviewers are getting tired. A team that stops finding errors has usually stopped looking. That distinction sits at the center of how we think about enterprise AI ROI metrics, and it applies with particular force to a function where one bad sentence is a regulatory event.

One more thing worth watching this year. Investors and analysts increasingly reach your disclosures through AI tools rather than by reading the transcript, which means the clarity of your prepared remarks now determines how a model summarizes your quarter to someone who never opens the filing. Writing plainly, defining your terms and putting the number next to the claim used to be a courtesy. It is turning into distribution, and it rewards the same discipline the finance function has been building in our guide to AI for finance.