The Machine on the Other Side Reinforces; It Doesn’t Reason

A buy-side analyst pastes your quarterly filing into an AI assistant and types a single question: should I be worried? The answer comes back in three well-structured paragraphs, sourced to your own numbers, fluent and fast. It reads like analysis. It is something closer to an echo, and the difference is the most important thing IR has to understand about its new audience.

The second piece in this series argued that your reader is now a machine: in thin-coverage markets, software often reaches your story before any human does, and being unreadable to it is most expensive exactly where IR least expects it. That piece was about access, meaning whether the machine can find and parse you. This one is about what the machine does after it reads, which turns out to be the more uncomfortable question. Because it does not stop at reading. The buy-side professional is not asking the assistant to read. They are asking it to interpret: summarise the quarter, flag the risk, tell me whether to worry. And the research on how these assistants actually behave should unsettle anyone whose job is to change an investor’s mind.

The assistant returns a disposition, not a verdict

The mechanism is not mysterious, and it is not a bug the next model release will patch. One study fed real investor due-diligence posts to an AI assistant and compared its replies with the human replies the same posts received. The machine’s answers were systematically more bullish than the humans’, by a wide margin, and stayed more bullish even on posts that were themselves bearish. Instructing the model to behave like a professional investor narrowed the gap by only a fraction. The disposition was baked in, and prompting barely moved it.

That is the uncomfortable shape of the thing. The assistant is not weighing your filing and reaching a verdict. It is returning a confident, well-cited answer shaped by what it was trained to produce and by how the question was framed, not by independent analysis of your company. Ask with a worried framing and the worry comes back, sourced and structured. Ask with an optimistic one and the optimism returns in the same costume. Either way the investor gets the feeling of rigour. They “did the research.” The polish makes the loop harder to notice, not easier.

For an investor this is a quietly dangerous tool, because it dresses a disposition in the costume of analysis. The old way of confirming a prior was lazy and felt lazy. You read the bull case, skipped the bear case, and some part of you knew it. The new way feels like work, produces a sourced three-paragraph memo, and is no more likely to tell you something you did not already believe.

The wrong prior is the whole problem

Now add the second force, because it compounds the first. Institutional investors are increasingly equipped, running structured-data pipelines over standardised disclosure, the kind XBRL was meant to enable and that AI now makes usable at scale. Research on AI-equipped investors finds that the firms they own file measurably more machine-readable disclosure, so the structured channel is real and it is widening. But that machinery does not interpret in any human sense. It pattern-matches against priors and peers. Feed it more standardised data and you get a faster, broader version of the same prior, not a wiser one.

Put the two forces together and you have the situation that should change how IR thinks about its audience. The investor arrives with a prior, formed from your sector, your geography, your last surprise, the reputation of your management, the country-risk discount the machine applies before it has read a word. They consult an assistant that returns a confident, biased reflection rather than a challenge. They run structured tools that amplify whatever signal is already there. At no point in that chain is there anything built to update the prior against the grain of the framing that set it.

This is where thin-coverage Asian markets get hit twice, the thread running through the whole series. A company in a market the models know little about does not start from a neutral prior. It starts from a discount the machine inherited from the average of its neighbours, and every downstream tool the investor uses tends to confirm that discount rather than test it. The under-followed issuer is not merely harder to find. It is easier to be wrong about, and the wrongness is now self-reinforcing.

The residual is the frame the buy-side machine can’t supply

So what actually moves a prior that an AI assistant and an army of structured tools are busy confirming?

Not more data. The investor already has the data, and the machine is swimming in it. What is missing is a frame: a way of seeing the company that reorganises the existing facts into a different conclusion. “Our margins fell” is a data point any tool will surface and any worried prior will absorb as confirmation. “Our margins fell because we are pricing for a land-grab in a market that will consolidate to three players, and here is the cohort data that shows it working” is a frame. It hands the machine no new numbers. It hands the human on the other side a reason to override the machine’s reflex, a story specific enough that the investor stops asking the assistant to reassure them and starts asking it to check something.

AI on the buy-side doesn’t change minds — it hardens them.

That is the residual, and it is the same one the series has found everywhere. The machine produces fluent reflection; the human produces the frame that makes reflection insufficient. A model has no thesis. It has a disposition and a prompt. The job that survives, and becomes more valuable as the tools get better, is the construction of the one thing the machinery cannot manufacture: a reason to think again.

So what, if you run IR

The instruction that follows is almost the opposite of the AEO advice in the third instalment, and both are true at once. Yes, make yourself machine-readable; that advice was not wrong. But do not mistake legibility for persuasion. The machine that parses you cleanly will still hand its user the prior they walked in with, beautifully formatted.

The work that changes an outcome is upstream of the machine and aimed at the human behind it: the deliberate construction of a frame that the existing facts support but the default prior misses, made specific enough to survive translation into the investor’s tools without dissolving into the optimistic-or-worried mush the assistant would otherwise produce. Give the machine clean data so it can find you. Give the human a thesis so the machine’s reflection stops being enough.

That is where the series has been heading for six instalments, and it is worth saying plainly at the end. Every layer of IR that a machine can do — produce the draft, read the filing, optimise the page, write the disclosure paragraph, run the agentic workflow, summarise for the buy-side — the machine will do, soon, and adequately. What is left over each time is small and human and the same: the judgment, the accountability, and the frame that makes someone think again. That residual is not a consolation prize for a profession being automated. On the evidence, it was always the only part the market was paying for. The machines have just made it impossible to keep pretending otherwise.