Your New Reader Is a Machine — and It Reads Indonesia Worst

An institutional investor researching your company this year may never read your disclosure. They ask a machine, and they read what the machine says about you.

That is not a forecast. Brunswick’s 2026 investor survey found institutional investors already using generative AI to summarise financial news (57%), to search for specific answers about a company (50%), to compare one firm against another (49%), and to replace internet search outright (38%). Buy-side adoption of AI sits somewhere between 70% and 95% across the industry studies. One research platform that sells generative search over filings and transcripts has crossed $600 million in annual revenue, and reports that three-quarters of its users run a generative search every month.

So the reader has changed. The analyst still exists, but increasingly the analyst reads you through a layer of software that has already read everything you published, pulled the sentence it thinks answers the question, and surfaced the negative signal buried on page 140. The human you spent a career learning to persuade now meets your company as a summary written by a machine.

The audience moved, not the job

Most of the anxiety about AI in IR is aimed at the wrong target. The fear is that the machine takes the IR professional’s judgment: that it targets the investors, sets the strategy, makes the calls. It does not, and the adoption data is blunt about it: only around 5% of IR teams use AI for investor targeting. The high-judgment work has barely moved.

What the machine has taken is not the job. It is the audience.

For as long as the function has existed, investor relations has been the practice of informing a human decision-maker, giving an analyst or a portfolio manager what they need to understand the company and take a position. That human is now downstream of a machine. The disclosure still matters, enormously, but it is read first by software and relayed to the person, who acts on the relay. The companion piece to this one argues that the live human moment, the executive under questioning on the call, is the part a machine cannot reproduce. This is the other half of the same shift: everything you commit to a document is now read by a machine before a human sees it.

That is survivable. The job did not disappear; the reader did. But how well the machine reads you is no longer a courtesy question. It is the question. And in Asia, the answer is decided by forces that have nothing to do with how good your IR is.

First, whether a machine can read you at all is decided by your exchange

Whether AI can read a company cleanly is largely set by the filing regime, not by IR effort. European issuers have filed machine-readable iXBRL under ESEF since the 2020 financial year. US filers have filed Inline XBRL since 2018. In those markets, the structured, tagged version of the accounts is the official one, and a machine ingests it without guessing.

Indonesia has the rails. IDX has required XBRL financial statements since 2015 and ran e-IPO from 2021. What is not confirmed is the depth: how much is tagged, whether the tagging extends to English, how the block-tagging breadth compares to ESEF. The Western conversation assumes the company controls its own machine-readability. In much of Asia, the exchange and the format decide it first, and the IR team inherits whatever the regime provides.

Why this matters is best shown by a single experiment. XBRL International gave a language model a set of structured reports and asked it to analyse them; it handled them correctly. Then it gave the model the same numbers as a PDF. The model pulled a figure from an adjacent column into a calculation that did not add up. The data was identical. Only the format changed. A machine reads structure; it misreads layout, and it misreads silently, in ways you never see and never get to correct.

Second, the machine reads your language worst

There is a penalty the Western corpus never has to carry. Language models underperform on non-English text, and the gap is widest for low-resource languages. Financial models are overwhelmingly English-centric, and the research literature notes the near-absence of high-performing models trained on Indonesian-language data.

The consequence is uncomfortable and specific. A company can disclose flawlessly in Bahasa Indonesia and still be misread by the model that now mediates its audience, not because the disclosure was poor but because the reader handles the language badly. The gap is closing, but slowly and unevenly, and no IR team can close it on the regulator’s behalf.

This reframes something IR teams have long treated as optional. Structured, English-language materials are not a politeness extended to foreign investors. They are the channel through which the machine actually reads you. In a market where the dominant new reader is weakest in the local language, the English version is not the translation. It is increasingly the original.

Third, thin coverage makes the machine the primary reader

Here is the inversion the generic take gets backwards.

In a heavily covered US large-cap, AI is a layer on top of a dense human market. Twenty or thirty analysts read the filing; if the machine hallucinates, a wall of published notes corrects it. The machine is one reader among many, and the others have credentials.

In an under-covered IDX small or mid-cap, that correction layer is thin or absent. Sell-side coverage is sparse. Ownership is concentrated (IDX has begun publishing high-concentration lists and is moving to a 15% free-float requirement by March 2027), and the register is often retail-heavy. In that setting the software may be the only thing that reads the full filing end to end. There is no analyst consensus standing by to catch its error.

The cost of being unreadable by a machine is higher in Jakarta than in New York, not lower.

The generic take treats AI-readability as a large-cap sophistication, a refinement for companies that have already solved everything else. It is the opposite. Machine-readability matters most for the small, thinly covered, non-English issuer, precisely the company least equipped to be thinking about it. And the regulator is already an AI reader: HKEX has reported that an AI-assisted ESG review expanded its scope by more than 500% over what manual effort had covered.

In New York the machine is one voice in a crowd. In Jakarta it is often the whole room, and it reads you in a language it barely speaks.

What to do about it

The instinct this should produce is not panic, and not a vendor contract. It is a short list of things an IR team can actually control.

Treat English-language, structured disclosure as primary distribution, not as translation overflow. If the model reads English best and reads your market most, the English version has to be first-class, complete, and timely, not a courtesy PDF posted a week late.

Get your numbers out of layout and into structure wherever the regime allows it. Tagged data and structured HTML are read correctly; a PDF is read by guesswork. The format you publish in now determines whether the machine gets your figures right.

Be your own source of truth. If your IR site is not the authoritative, machine-readable version of the company, something else becomes the answer instead: a third-party aggregator, or a stale figure it half-remembers and states with confidence. The third piece in this series takes up that audit directly.

And hold on to the half of the job that does not automate. The machine reads what you published. It cannot ask the follow-up that reveals what you did not. The inbound half of investor relations — the perception you gather, the relationship you build, the question an investor asks that tells you what they actually fear — remains entirely human. If anything, transparency matters more now, not less: the machine reads the unfavourable specifics too, so selective silence is more exposed than it used to be, not better hidden.

One honest caveat. There is, as yet, no OJK or IDX guidance on AI in investor relations. On this front you are ahead of the regulator, an unusual position in Asian capital markets, and one worth using while it lasts.

The reader you spent a career learning to persuade has hired a machine to read first. In New York that machine is one voice in a crowd. In the markets we serve it is often the whole room, and it reads you in a language it barely speaks. The work now is to make sure that when the machine reads you, it gets you right — because increasingly, the machine is who you are talking to.