Investor relations teams I work with know their story cold. They can land it in a lift, over bad coffee, in front of a fundie who made up their mind on the way in. The story gets told in rooms they control: the results call, the roadshow, the site visit, the one-on-one where the CFO leans forward. Then the room empties, and the story stays behind with the chairs.
This worked because the people who mattered were in the room. It stops working when a growing share of the people who matter are reading a machine's version of you instead.
The reader has changed
In December 2025 Brunswick surveyed 100 US institutional active equity investors, most from firms managing more than a billion dollars. Forty-six per cent said they were now more likely to skip the earnings call, or the transcript, and read an AI summary instead. Forty-four per cent said they trusted that summary as much as a sell-side note. Sixty-eight per cent said AI had changed how they approach earnings calls at all. A separate survey for the Center for Audit Quality in May 2026 found 60 per cent of institutional investors using AI always or often on earnings calls, and 41 per cent on filings.
This is not the machine replacing the market. It is something quieter and harder to argue with. More of the market reads a synthesised output every day, the number reading it first is 46 per cent, and it is rising.
The machine was reading your business before the language models arrived. Between 2003 and 2016, machine downloads of company filings from the US regulator's system rose from 39 per cent of the total to 78 per cent. Companies noticed, and measurably changed their words: firms with high expected machine readership made their filings easier to parse and dropped the terms that scoring algorithms punish. That paper was published in 2023 in the Review of Financial Studies and describes corporate behaviour from a decade before ChatGPT. Writing for machines is not a forecast. It is established practice, and nobody has told the board.
What the machine actually compiles
Here is the part IR teams underestimate. The machine did not attend your call. It does not read your results presentation, because your results presentation is a PDF, and it is reading the HTML version published by a third-party portal, which in one case I looked at last week labelled every US dollar figure as Australian. It assembles you from whatever is legible: the Wikipedia lead, a data vendor's paragraph written in another era, the finance portals that copy that paragraph, and your worst headlines.
Stale facts persist in that corpus the way old bottles persist at the back of a bar. One ASX 20 company I read this month has five different answers to "who is the CEO" across the profiles a model draws on, and its largest single investment narrative appears in none of them, except as a cash flow headwind in a valuation note. The story its executives told on the results call was coherent, specific, and confident. Almost all of it stayed in the room with the chairs.
Expertise is not being replaced. It is being overrun.
I started with a hypothesis that deep expertise in this field was giving way to expansive generalism. The evidence says something more useful. In a randomised trial across 776 professionals at Procter & Gamble, the control group, working without AI, split along their function as you would expect: technical staff proposed technical answers and commercial staff proposed commercial ones. The group given AI did not split. They produced balanced solutions regardless of background. Breadth, historically the scarce thing, was the first thing to be commoditised.
So the expert still condenses the first rounds. The expert still frames the question. Then the machine overruns the framing with its generalised biases, drawn from the corpus, and the corpus is the one above. Meanwhile, on the buy side, the funds with the budget are training their own weights on their own view of the market, and understandably not publishing the results. Everyone else reading you, the smaller funds, the brokers, the retail investor, the journalist, gets the frontier models, reading the same public surfaces, producing the same compression.
The most instructive study here is about satellite data, not language models, and it shows what happens to expertise when a machine starts covering the same ground. From 2011 a data vendor began selling satellite counts of cars in the car parks of 48 US retailers, a live read on foot traffic and therefore on sales. Before that, knowing how a retailer's quarter was going was the specialist's edge. Afterwards anyone could buy it. The study tracked 4,000 active funds and found that once a retailer was covered, fund managers' ability to pick that stock fell. The fall was steepest for the funds that had known the stock best: retail sector specialists lost ten to twenty times more stock-picking skill than generalist funds, because the data had replaced exactly what they knew. Their response was not to become generalists. They sold down the covered retailers and moved the same money, and the same depth, into retailers the satellites had not yet reached, where their knowledge still paid. That is the lesson. Expertise is not a store of value. It is the gap between what you know and what the machine now covers, and the machine's coverage keeps growing.
The sceptics are half right
The sceptics say these tools give everyone the same access to the same reports, so there is no edge in them. They are right that an edge everyone can see stops being an edge. The finance literature has measured this for decades. When academics publish a pattern that beats the market, and every fund can read the paper, the pattern loses about 58 per cent of its return because everyone starts trading it. AI makes that happen faster. A strategy that used GPT-4 to read news headlines and trade on them made outstanding returns in late 2021 and lost 81 per cent of that performance by mid-2024, and the authors put the loss down to everyone else adopting the same tool.
But access was never the asset. In a field experiment, 640 Kenyan entrepreneurs were given a GPT-4 business mentor. The average effect on profit was nothing. High performers gained 15 per cent, low performers lost 8, and the mechanism is the whole point: both groups asked similar questions and received comparable advice. The divergence came from which suggestions they chose to act on. Same tool, same answers, opposite outcomes.
The value has always gone to whoever connects known basis points. That is not a metaphor. It is the mosaic theory, and the US regulator wrote it into the record in 1999 when it drafted Regulation FD: a skilled analyst may piece "seemingly inconsequential data together with public information into a mosaic which reveals material non-public information", and doing so is lawful. The regulator's own position is that assembly is where the value sits.
Rory's bees
Rory Sutherland tells a story about honeybees. A significant minority of foragers ignore the waggle dance and go looking for nectar the hive does not know about. In the short run the hive would be better off if they fell in line. In the long run the hive that fell in line starves when the known flowers die. He calls them the colony's research and development function.
The science is on his side, and improves the argument. Field studies put the scouting share at 5 to 35 per cent of foragers depending on how much forage is about. A 2012 paper in Science found scouts are not rogues at all. They carry a distinct gene expression signature, over a thousand differences from non-scouts, and a bee that scouts for nest sites is more than three times as likely to scout for food. The colony is not tolerating randomness. It is staffing a small, permanent exploration cadre, at a proportion that would look indefensible on any efficiency measure.
Every IR team knows its explore bees. They are the few analysts and portfolio managers who still ring up, still come to site, still ask the question that is not in the model. IR teams know them by name and tell them the story beautifully. What IR teams have not spent time on, beyond the annual report, is how the story reaches anyone else. And "anyone else" now includes the machine that 46 per cent of the register reads first.
Two readers, one requirement
The compiler wants structure, consistency, and unambiguous language. It rewards a sentence said the same way on every surface, and it punishes the drafting reflexes that legal review installs. The explore bee wants what is not already in the corpus, and there is less of that every quarter, because everything in the corpus is now priced faster and held more uniformly than at any point in market history. Both readers want the same thing, which is the story stated once, clearly, and held consistently everywhere it appears.
That is brand discipline. It has never reached IR because IR reports to counsel and the CFO, brand reports to marketing, and nobody owns the sentence that has to survive both. The accounting literature has been telling CFOs for a decade what that costs: a one standard deviation improvement in annual report quality is associated with a 77 basis point reduction in the cost of equity, roughly the same as an equivalent improvement in earnings quality itself. How you say it is worth about what the numbers are worth. The average annual report is still written at a reading level four years beyond the Wall Street Journal editorial page.
So the prescription is short. Put the story you are comfortable with somewhere the machine can read it. HTML beside every PDF. One sentence, one word order, on every surface you control. A plain-language account of the strategy on the investor site, in your own terms, so the retriever has a primary source to prefer over a data vendor's paragraph from another era.
The alternative is that the machine writes your story for you, from your worst headlines and someone else's boilerplate, and 46 per cent of your register, and the ones considering joining it, read that first.
Plan B Machine Legibility
A story variance report on your company. Your last results set, annual report and transcript on one side. On the other, what three frontier models say you are when nobody hands them a document, and the surfaces they compile from: the data vendors, the portals, the forums, the headlines. Four pages. The variance, why it exists, and three moves to close it inside ninety days. Then we run it again and hand you the distance.
If you run investor relations for an ASX company and want to see yours, write to b@planb.works.
Sources. Brunswick Group 2026 Investor Survey (fielded 20 November to 4 December 2025, n=100). Center for Audit Quality with KRC Research, May 2026, n=100. Cao, Jiang, Yang and Zhang, "How to Talk When a Machine Is Listening", Review of Financial Studies 36(9), 2023. Dell'Acqua et al., "The Cybernetic Teammate", NBER w33641. Bonelli and Foucault, "Displaced by Big Data". McLean and Pontiff, Journal of Finance, 2016. Lopez-Lira and Tang, arXiv 2304.07619. Otis et al., "The Uneven Impact of Generative AI", SSRN 4671369. SEC Regulation FD proposing release, 28 December 1999, quoting Elkind v. Liggett & Myers. Seeley, Behavioral Ecology and Sociobiology, 1983. Liang et al., Science 335, 2012. Heflin, Moon and Wallace, Journal of Financial Reporting, 2016. Li, Journal of Accounting and Economics, 2008. Sutherland, Alchemy, 2019. No withdrawn or retracted paper is cited.
