The demo always ends the same way. The software has read the quarter’s peer disclosures, flagged what changed, assembled a briefing memo, and drafted a holding response to an awkward investor email. Then comes the final click, the one the vendor lingers on: a button that sends the reply on its own, with no human reading it first. That button is the whole product, and it is the reason this article exists.
Earlier in this series I asked you to let a machine answer your earnings call and to hope it did badly. That experiment had a tidy boundary. The machine generated text, a human read it, and a human decided whether to say it out loud. The judgment stayed where it belonged. The systems being sold to IR functions this year erase that boundary. They are described in the vendor deck, and more carefully in the regulators’ documents, as agentic: software that does not merely produce an output for a human to use, but takes actions toward a goal with some measure of independence.
This is a genuinely different thing, and it deserves a different question. The first article asked what a machine gets wrong. This one asks who is responsible when a machine acts.
Autonomy is two dials, not one switch
The instinct is to treat AI autonomy as a switch: the robot is either doing your job or it isn’t. The people who have thought hardest about this, including the authors of Singapore’s framework for governing agentic AI, do not use a switch. They use two dials.
The first dial is the action space: how much the system is permitted to touch. Reading your CRM is one setting. Drafting an email is another. Sending it, moving money, filing something: each is a wider aperture, and each carries a larger blast radius when the system is wrong. The second dial is autonomy: how far the system runs between human checkpoints. A tool that proposes and waits sits at one end. A tool that plans, executes, and reports back afterwards sits at the other.
Plot any use case on those two dials and you can see its real risk, which has almost nothing to do with how clever the model is and almost everything to do with what it can reach and how long it runs unattended. An IR drafting assistant with a human reading every word before it leaves the building is low-risk by this map, however impressive it sounds. An agent that watches the wire and auto-publishes a clarifying statement is high-risk, however dull it sounds. The cleverness is not the variable. The leash is.
The frameworks all converge on the same word
What is striking, reading across Singapore’s agentic framework, the MAS risk-management work, and the broader Asian governance material now accumulating, is how little they disagree. They differ on detail and emphasis. They converge, almost word for word, on a single requirement: meaningful human accountability.
Not human oversight in the weak sense, meaning a person somewhere in the building who could in principle have intervened. Meaningful accountability in the strong sense: a named human who owns the outcome, whose judgment is engaged at the points that matter, and who cannot offload the consequence to the model when it goes wrong. The frameworks are emphatic that a human who rubber-stamps whatever the agent proposes does not satisfy this. A person clicking “approve” forty times an hour is not providing oversight. They are providing a signature they have stopped reading, which is worse than no signature at all, because it manufactures the appearance of control.
That failure mode has a name in the literature, automation bias: the well-documented human tendency to defer to a confident machine. The frameworks treat defending against it as a design problem, not a training slogan. Stripped down, the whole apparatus of agentic governance is machinery for keeping a human signature real as the volume of machine action rises.
Deploying agentic tools in your IR function? An IR gap diagnostic maps where the human signature has to stay, and where the leash is already too long.
The residual moves under agentic AI, but it does not disappear
This is the part that should interest anyone who has run IR, because it is where the agentic shift quietly changes the job.
In the first instalment the residual, the thing the machine could not reproduce, was informational: the firm-specific answer to the unscripted question. With agentic systems the informational residual shrinks, and it is tempting to conclude the human residual shrinks with it. It does not. It relocates, from knowing the answer to standing behind the action.
Consider what is actually irreplaceable when an agent drafts and an IRO releases. It is not the drafting; the agent did that, probably faster and arguably better. It is the decision that this statement, at this moment, in this market, carries the company’s name and the IRO’s judgment, and the willingness to be the person who is wrong if it is wrong. A language model has no name to sign and no career to lose. Accountability can only attach to someone who can be held to it, which is why the signature is both the residual and, by construction, human.
Automation removes the typist, not the signatory.
Matt Levine put the boundary in one line when he wrote, in 2023, that you should not let a robot do your earnings call. His point was not that the robot would do it badly. It was that the earnings call is a place where a human commits the company to its words under questioning, and a system that cannot be committed to anything cannot do that, however fluent it becomes. Agentic AI does not change that. It raises the stakes, because now the machine can act all the way up to the threshold of the signature, and the temptation to let it cross is real. The frameworks exist precisely because someone has to say no.
So what, if you run IR
The practical posture is not to refuse the agents. They are useful, the dull high-volume work is exactly where they earn their keep, and a company that bans them will simply be slower than one that governs them. The posture is to be deliberate about the two dials and ruthless about the signature.
For every AI use case in the function, ask the two questions the framework forces: what can it reach, and how far does it run before a human owns the result. Keep the action space narrow wherever the consequence is irreversible: a published statement, a response to a regulator, anything that moves a price. Keep the autonomy short wherever the company’s credibility is on the line. And make sure the human at the checkpoint is exercising judgment rather than manufacturing approvals, which means giving them the time and context to actually read, not the volume that turns them into a bottleneck they will rationally start ignoring.
Do that, and the agents become what they should be: very fast typists who never sign. Because in the end the function comes down to a smaller and smaller human act: the decision to put the company’s name to a sentence and to be answerable for it. Everything around that act can be automated. The act itself cannot. The day it is, you no longer have an investor relations function. You have an outbox no one is responsible for.
Frequently asked questions
What is agentic AI in investor relations?
Agentic AI is software that does not merely produce an output for a human to use but takes actions toward a goal with some measure of independence. In an IR function that can mean reading peer disclosures, assembling a briefing memo, drafting a response to an investor email, and, at the far end, sending that reply on its own without a human reading it first. The further the system moves toward acting unattended, the more governance it requires.
How do you assess the risk of an agentic AI tool in IR?
Plot the use case on two dials rather than a single on-off switch. The first dial is the action space: how much the system is permitted to touch, from reading a CRM to drafting an email to sending it or moving money. The second is autonomy: how far it runs between human checkpoints. Risk has almost nothing to do with how clever the model is and almost everything to do with what it can reach and how long it runs unattended.
What does meaningful human accountability mean for AI in IR?
It means a named human who owns the outcome, whose judgment is engaged at the points that matter, and who cannot offload the consequence to the model when it goes wrong. A person clicking approve forty times an hour is not providing oversight; they are providing a signature they have stopped reading, which manufactures the appearance of control. The whole apparatus of agentic governance exists to keep that human signature real as machine action rises.
Advising listed companies representing over $50 billion in aggregate market capitalisation.
Deploying agentic tools in your IR function? Map where the human signature has to stay. We review your AI use cases against the two dials — action space and autonomy — and show you where the consequence has outrun the oversight. The diagnostic is free and carries no obligation.
Request an IR Gap Diagnostic