A reader who moves trillions does not read the advertisement. They read the filing. Every business day the SEC takes in the record of American business, filing by filing, and the business day after, Edgar reads it. Three lines in that record now tell the story of this piece. Artificial intelligence appeared in 848 annual-report documents in 2021 and in 3,904 so far this year. AI governance appeared in 20 two years ago and 226 so far this year. And print, the oldest line in the media budget, is four cents of every advertising dollar the big agencies place, and nobody reports it any more. Writing from the record costs a cent on any machine you pick, and millionths of a cent to put in front of a reader. What costs money now is the executive who cannot say how the machine is governed. Insurers have started writing that person out of the policy, regulators have started asking for the training record, and the credentials that answer both cost less than one renewal. The four cents on the dollar that print still takes is the four cents that moves next, to AI insurance and AI training. Paper was the last line item of the old media budget. Governance is the first line item of the new one.
The four cents still on paper
The three largest agency groups have not banned paper. They stopped funding it, and then they stopped counting it. Print advertising worldwide was $45.5 billion in 2025 against $1.08 trillion of advertising, by WPP Media's own mid-year forecast. Read that again: forty-five billion dollars, worldwide, for all of print advertising on earth. One company, Netflix, took in $45.2 billion last year. The sheds Americans rent for the boxes they never open, the self-storage trade, took about the same. That is the whole of print advertising: about four cents on the dollar, down 3.1% in a year. Digital is 73 cents, and 82 once streaming, digital billboards and the digital editions of papers are counted. Magazines fell 7.2% in 2025 and are forecast to fall another 10.7% in 2026, the steepest drop of any channel. Publicis, the group growing fastest, does not report a print line at all; its 2025 net revenue is Connected Media 60%, Intelligent Creativity 26%, Technology 14%. In WPP's year-end forecast, newspapers do not get their own figure. They sit inside a category called content.
The industry under those four cents is in run-off. The local paper closes at more than two a week, 136 last year, and 213 counties now have no local news source. The Blade in Toledo, printing since 1835, stops on the last day of this year unless somebody buys it. The survivors are a handful of national titles that sell digital: the New York Times added 1.4 million subscribers in 2025 and passed two billion dollars of digital revenue. Commercial printers lead every trade in bankruptcy filings. The presses come out and the plants are sold: the Chicago Tribune's was demolished for a casino, Tampa's and Roanoke's are becoming apartments, the Star Tribune's went to a data centre developer. In the whole record there is no case of a press finding new work. The press makers are another matter. Komori, independent and listed in Tokyo since 1983, built an expert system for press control with the Japan Printing Academy in 1997, put a self-learning version, KHS-AI, into its presses in 2002, and brought it in at the industry's pace rather than ahead of it, standard across the range by 2020. Its subsidiary Seria now builds the printers that make circuit boards and advanced chip packaging, and in August 2026 Komori agreed to buy a maker of temperature-control equipment for micro-semiconductor plants. The press did not find new work. The press maker did.
What a thousand words costs, on every machine
The same article, about 3,000 tokens in and 1,400 out, costs between a tenth of a cent and six cents at the published prices of 28 October 2026, whichever provider you pick. Anthropic's and Google's figures are from their price pages; OpenAI's page could not be read by a machine today, so its rows are as reported by price trackers. Thinking tokens can double any row, batch runs halve them, and Google has announced rises for December and January. The choice of machine moves the cost by a few cents. None of it is where the money is.
This house writes through the Anthropic API from inside Cloudflare Workers, most pieces on Haiku 4.5. The Workers, the filing readers, the pipelines and the pages are already built and running, so one more article costs only its tokens: under a cent. Serving a read costs thirty millionths of a cent at Cloudflare's rate of thirty cents per million requests; a day's 11,416 desk reads come to about a third of a cent. In thirty days 75 readers arrived from an AI assistant's answer. Every figure in every article comes from what a company filed, linked to the document, and every forecast range is hashed before the outcome is known. That is the data safety a buyer wants: a number that can be checked, not a number that was said.
What the people who carry the risk are focused on
The security chiefs have already moved. In Gartner's 2026 survey of more than 1,600 CISOs, enabling and protecting AI is the first priority, having been absent from the top five a year earlier; 44% plan to put money into AI this year, up from 35%. In Proofpoint's 2026 survey, 88% of CISOs call safe AI use a top priority. KPMG finds only 24% of organisations have AI fully integrated into their security; the rest are part way. Their list is short: know which AI is in use, govern it in writing, keep customer data out of public tools, be able to answer for a model's output, and be able to show all of that to an insurer.
What the insurer now asks
The insurance market rewrote its forms this year. The ISO generative-AI endorsements for general liability took effect in January 2026, among them CG 40 47, which removes injury and advertising injury tied to generative AI. W. R. Berkley filed an absolute AI exclusion across directors-and-officers, errors-and-omissions and fiduciary lines. Chubb covers some AI incidents but excludes losses that hit many policyholders at once. In Delinea's survey, 42% say their cyber policies already carry AI exclusions, and 77% of insurers require a formal internal review before issuing or renewing. Aon finds 63% of insurance professionals say existing policies do not explicitly cover AI threats; Gallagher finds only 53% of those who had an AI-related loss were fully covered.
The same market will sell the cover back to a company that can document its governance. Brokers describe the split plainly: AI that is controlled and monitored earns a discount; AI that is unmanaged earns an exclusion. A standalone AI liability policy from Armilla and Chaucer at Lloyd's now carries limits past $25 million, for hallucination and model drift, and it is underwritten on the governance record.
For the directors, the exposure is personal. The SEC's Cyber and Emerging Technologies Unit has named AI washing an enforcement priority. Securities class actions over AI disclosures went from 7 in 2023 to 15 in 2024 to 16 in 2025. Delaware has confirmed that officers, not only directors, owe oversight duties. WTW's Lawrence Fine put the problem in one sentence: most directors and officers will not understand the nuts and bolts of AI. The underwriter now asks about it in the D&O meeting.
The slate
This is what a media executive should be able to put on the table at renewal. Ten questions the carrier now asks, the evidence that answers each, and the credential or framework behind the evidence, who should hold it, and what it costs.
The watch: what Edgar counts from here
The disclosure studies agree on the direction and differ on the sample. Among the S&P 500, the share flagging AI as a risk factor went from 17% for fiscal 2022 to 61% for 2024 and 77.6% for 2025. SEC staff say 59% of annual reports from large accelerated filers made an AI disclosure this year, up from 27%. Across more than 7,000 US filers, mentions of AI risk went from 4% in 2020 to over 43% in 2024. RSM read 1,098 cybersecurity disclosures and found AI-related mentions up 5.2 points in a year, the largest move in its data, with the finding that AI has entered the disclosure mainstream faster than AI governance has. That gap, between the word and the governance of it, is exactly where the insurer and the regulator are now looking.
So this house will watch three lines in the SEC record through the next two reporting seasons, and publish the counts on the first business day of each month as a Tidepool Note, sealed with a receipt: the phrase AI governance in annual reports, 226 so far this year; the phrase AI literacy, 17 so far this year against 3 in all of 2025, the EU training duty arriving in American filings; and AI inside the Item 1C cybersecurity disclosure, which is where the carrier reads. The baseline is above. When the governance line catches the AI line, the four cents will have finished moving.
The steal, for a product brand
Stop buying paper unless a statute makes you. Move the money to writing from the record: a cent an article, millionths of a cent a reader, every figure checkable to its source. Then spend the first few hundred dollars of what you saved on the slate. Name the AI owner and send them for the AIGP. Write the policy and the inventory on the ISO 42001 shape. Run every system through the NIST framework once and keep the sheet. Keep the training record, contractors included. Put the AI clauses in the vendor contracts. Vet every AI claim before it is published. Extend the incident plan to models. Walk into renewal with the slate and ask for the discount instead of the exclusion.
The actuary's view
Strip the insurance story to what a carrier actually prices and it is three numbers: how often the thing goes wrong, how much it costs when it does, and how many policyholders it hits at once. The third is the one that frightens an actuary, because a flaw in one widely used model is not a thousand independent claims, it is one claim a thousand times, the way a hurricane is. That is exactly what Chubb's wording excludes, losses that hit many policyholders at the same time, and it is why affirmative AI cover is sold with limits and underwritten on the governance record rather than on the balance sheet. Governance is the evidence that moves the first two numbers: a named owner, an inventory, a tested incident plan and a training record lower the frequency and cut the severity, and the carrier's rate follows the evidence. An executive who can produce the slate is a different risk, in the arithmetic, from one who cannot.
Where the four cents goes next
Follow the money and the replacement is already in the forms. The insurer now asks for a formal review before it renews, 77% of them by Delinea's count, and sells affirmative AI cover underwritten on the governance record, with limits past $25 million at Lloyd's. The regulator in Europe now requires a training record, contractors included, as of 2 August 2026. The frameworks are free and the credentials are a few hundred dollars a head. A media budget that moved its last four cents off paper has the same four cents to spend, and the two lines that now earn it back are the premium and the training. One lowers the other: a documented, trained operation is the one the carrier discounts. That is the spend that replaces print: not another channel, the proof that the channels are governed.
The position
Print is four cents on the agency dollar, falling, and no longer reported. The machine is a cent. The exposure is the person. A media business cannot carry an executive who cannot explain how its AI is governed, because the insurer will not carry them either, and the regulator now asks for the training record. The slate is public and cheap. The cost of not having it is the policy. Print's four cents go to AI insurance and AI training, and that is the spend that pays for itself.