I've watched a lot of demos over the last two years. Most of them were impressive for about ninety seconds. Someone types a question into a box, a paragraph comes back, and everyone in the room nods, because the paragraph is fluent and confident and arrives faster than a human could have written it. Then somebody asks a second question, and the whole thing quietly falls apart.
I should say up front that I work for a company that sells software to investor relations teams, so discount what follows accordingly. But the pitch I keep hearing, that IR is about to be transformed and the function will look unrecognizable in three years, doesn't match what I actually see when I look at how this work gets done. Some things are genuinely moving. Most of the things people tell you are moving are not. It's worth being precise about which is which, because the confusion is expensive.
Start with what doesn't move
The calendar doesn't move. Earnings lands when it lands, set by the exchange and your own filing obligations, and no model shortens the gap between quarter close and the release. The disclosure process doesn't move either. Your CFO still decides what gets said, legal still reads it, and the committee still meets. If anything, the introduction of tools that can draft quickly puts more pressure on that review process, not less.
Regulation FD doesn't move. This one deserves saying plainly, because I've heard vendors imply otherwise by accident. Nothing about a model changes what you're permitted to say, to whom, or in what order. There is a real question about where the line sits between observing institutional behavior and possessing something material and non-public, and I'll come back to it in a later piece. But that's a question about disclosure law, not about technology.
The relationship doesn't move. A portfolio manager takes your call because of an accumulated history of you being straight with her, including the quarter you missed and didn't spin. That's earned in a way no system participates in. I've never seen an IRO win a long-only holder with better analytics. I've seen plenty win one by being the person who called before the news broke.
And the asymmetry doesn't move. The buy side has always had better data than you do, more of it, and more people looking at it. They have these tools too, and larger budgets for them. If your mental model is that AI will let IR catch up to the people trading your stock, adjust it. The gap is structural.
What actually moves is boring
Here's the whole shift, as far as I can tell: the cost of paying attention went down.
That's it. Not the cost of writing. The cost of reading. For twenty-five years, the binding constraint on an IR team has been the number of things a small group of busy people could actually read, listen to, and hold in their heads at once. You could not listen to all eleven of your peers' earnings calls. You could not read every sell-side note on the sector, every 13F amendment, every quarterly commentary letter from a top-thirty holder. So you triaged, and the triage was mostly reasonable, and the things you missed were invisible to you precisely because you missed them.
When attention gets cheap, three things follow.
Coverage becomes continuous instead of episodic. Humans monitor in bursts, and the bursts are scheduled around board meetings. The unglamorous truth is that most of what matters in a shareholder base happens on an ordinary Tuesday in August when nobody is looking. A system doesn't care that it's August.
Take the shape of it. A fund nobody has on the target list spends real time in the investor section of a company's website. On its own that means close to nothing. Institutions look at IR websites constantly, and most of it is an associate doing homework that never becomes a position. What would make it matter is the visit being read against everything else known about that firm, the combination clearing a bar worth interrupting a human for, and the IRO hearing about it in the same week rather than the following quarter. That is a phone call that gets made in August instead of November.
I want to be careful about the claim, though. A system like that does not find you a shareholder. It finds a reason to make a phone call on a particular afternoon, which is a much smaller thing and also the entire thing.
Explanation becomes cheap, which changes what you can defend. Targeting has always been a ranking exercise, and the rankings were mostly defensible but rarely explainable. "Because the screen said so" and "because I've known them twelve years" are both fine answers until your CFO asks why this name and not that one, in front of the board. A ranking that carries its own reasoning is a genuinely different object, and the difference shows up in a meeting rather than in a metric.
The floor rises before the ceiling does. Your best day of preparation doesn't get much better. Your average day gets a lot closer to your best day. That's a real improvement and it's almost impossible to demo, which is part of why demos are so misleading.
How to tell a demo from a working system
This is the part I'd want if I were sitting on the other side of the table. Five questions, none of them technical, all of which I've watched break a demo.
Ask it about something it can't explain. Demos are built on a clean quarter with a clean story. Bring your worst one. Ask why a specific holder cut the position in a month where nothing happened. A working system tells you what's observable, tells you what's inference, and tells you when it doesn't know. A demo produces a fluent paragraph either way, and the fluency is the problem, because it reads identically whether or not there's anything behind it.
Ask where the data ends. Every system has a horizon. What's the vintage of the ownership data? What's the settlement lag? What's the reporting threshold below which a position simply doesn't appear? If nobody in the room can answer that in specifics, you are looking at an interface, not an analytical system. I don't think this is a hard question, which is why I find it so revealing when it lands badly.
Ask it to distinguish observed from inferred. Then ask it to do so in the output itself, every time. Ownership and trading work is full of legitimate inference. That's the craft. But inference presented in the same register of certainty as a filed fact is worse than useless. It's the thing that will eventually put a wrong number in front of your board.
Ask who gets interrupted. A dashboard you have to remember to open is a report. Reports are fine. But a report doesn't solve the problem you actually have, which is that the thing you needed to know surfaced on a day you were doing something else. Ask what would cause the system to tell you something unprompted, and then ask what it would stay quiet about. If it can't stay quiet, it will train you to ignore it inside a month.
Ask the third question, not the first. Demos survive the first question by design. Most survive the second. Very few survive being asked, three times in a row, "and why is that?"
Where I'd actually start
Not with a tool. Start by writing down the decisions your team consistently makes late. Not badly. Late. The targeting refresh that happens quarterly because that's all anyone has time for. The peer read that gets done the week of your call instead of the week the peer reported. The holder that showed up in the register in a way nobody had context for until someone spent two hours digging.
Those are the places where cheap attention converts into something you can feel. Everything sitting underneath a late decision is gathering, cross-referencing and sequencing, which is exactly what these systems are good at. The decision itself stays with you. So does the relationship, and so does the call you have to make under pressure with incomplete information, which is most of the job and always has been.
I'd rather see an IR team automate three specific late decisions than adopt an AI strategy. The first is measurable by December. The second is a line in a budget deck.
Written in a personal capacity. The views here are my own.