Loading today’s stories
Product1 distinct publisher3 min readPublished
An LLM read a decade of corporate filings and found production AI at under a quarter of the index, two thirds of it in technology. The non-tech comparison group is tiny.
The Product Desk · Product desk
Compiled by The Product DeskSomething wrong?How this is made
Run the arithmetic the authors leave implicit. One third of a number below a quarter is a number below 8%, which is the share of the S&P 500 that sits outside technology and has AI either embedded in its processes or working in the production of goods and services [2][3][1]. That is the real peer group a manufacturer, an insurer or a hospital system is measured against when someone in the room says everyone is doing this.
The stricter bar is narrower still. Fewer than two dozen non-tech constituents clear what the analysis calls full deployment, defined as AI being a core component of the firm's strategy and financial performance, deeply embedded across business functions and operations [4][5]. Four are named: Moderna, Mastercard, Bank of New York Mellon and GE Healthcare [4].
Set the pilot majority against the production minority and the gap is the story: at least a quarter of the index is running pilots that have not become process [7][2][2]. Fast Company's authors read that queue as adoption arriving steadily rather than at the pace the headlines imply [11]. An operator should read it as a crossing rate, because standing up a pilot is the cheap half of the work.
It also matters what was actually measured. An LLM scoured federal financial filings and scored each firm over ten years on the text describing its AI deployment [1]. That is a reading of what companies write about themselves. Disclosure language can be rewritten in a quarter; a demand forecasting system cannot.
The authors name their own binding constraint. In many cases models remain too expensive to run at the precision a task requires, so the set of tasks AI can theoretically automate stays much larger than the set it does [8]. Their progress series shows the shape of that. A 10 to 12 slide quarterly customer review deck, three to four hours of human work, was produced by LLMs at roughly a 50% success rate two years ago and 65% a year later, with 80% to 95% on most text tasks projected by 2029 [9]. From the 65% mark, that projection averages three to six points a year, against the fifteen points logged in the single year they measured [3]. The forecast built into the optimistic case is a slowdown.
Text is the easy end. The grocery example in the piece is 500 stores at 50,000 SKUs each, 25 million forecasts every two weeks [10], which works out to roughly 650 million forecasts a year [4]. Very little in the pilot column looks like that, and the non-technology firms that make up 90% of the US economy [6] are the ones who would have to build it.
Ranked by verification strength, evidence, and original report placement.
The analysis used a large language model to scour corporations' federal financial filings and score each firm over a 10-year period on the text describing its deployment of AI.
The model defines full deployment as AI being "a core component of the firm's strategy and financial performance, deeply embedded across business functions and operations."
The authors' reading is that corporate AI adoption is coming steadily and inexorably rather than at a blistering pace, and that most big businesses are currently in the AI slow lane.
Through the end of last year, less than a quarter of S&P 500 companies had AI either deeply integrated into their business processes or were using AI in the production of goods and delivery of services.
The technology sector accounts for two-thirds of extensive AI integration and use among S&P 500 firms.
Fewer than two dozen non-technology S&P 500 firms, including Moderna, Mastercard, Bank of New York Mellon and GE Healthcare, have achieved full AI deployment.
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One self-reported analysis, no disclosed methodology
Every quantitative claim traces to a single commentary describing its own LLM scoring of filings and its own unpublished task research. The source states its definitions but discloses no sample construction, scoring rubric, validation, or error rates, and there is no second publisher or independent dataset in the cluster to corroborate the index shares or the 50%/65% success readings. The derived arithmetic is internally consistent, which raises reliability of the inferences but not of the inputs.
Broad piloting, narrow production
On the source's own scoring, real production adoption is narrow: under a quarter of the S&P 500 at deep integration or production use, two thirds of that concentrated in technology, and fewer than two dozen non-technology firms at full deployment — under roughly 8.3% of the index outside tech. Pilot breadth is much wider (more than half the index), so intent is high and crossover is low. The score reflects that measured production state rather than pilot counts.
Deflationary framing, extrapolative forecast
The headline framing is anti-hype and matches the measured state: slow lane, pilots outnumbering production, cost-at-precision ceilings, partial automation as an end state. That part is aligned or even understated relative to typical coverage. The mild positive gap comes from the forward-looking layer built on the same unverified base — 'steadily and inexorably' certainty and an 80-95%-by-2029 projection — where the source's own success-rate series implies decelerating annual gains rather than the confident trajectory asserted.
Authors promoting their own unpublished model
The piece is a first-person commentary whose central evidence is the authors' proprietary filings-scoring model and internal task research, which gives a direct interest in the credibility and visibility of that work. The supplied text discloses no employer, client, or commercial relationship for the authors, so the assessment is limited to that visible self-referential interest; no vendor, fund, or product tie can be established from the material and none is inferred.
Directionally plausible, unverifiable specifics
The qualitative shape — wide piloting, narrow production, tech-concentrated deep integration — is coherent and internally consistent, and the derived arithmetic follows cleanly from the stated numbers. But there is one publisher, one undisclosed methodology, and no independent measurement in the cluster, so the specific shares (under a quarter, two thirds, fewer than two dozen) and the success-rate series should be treated as claims by interested authors rather than established figures.
product
ChatGPT Work's real ask is your Slack, and somebody has to say yes on everyone's behalf1 distinct publisher
product
OpenAI shipped a teen ChatGPT. The over-65 cohort doubled to 23% and got nothing.1 distinct publisher
invest
Only 34% of the C-suite knows who owns AI decisions. 57% of their subordinates do.1 distinct publisher
product
PayPal stopped saying no. Payments teams should now plan for a Stripe-owned checkout rail3 distinct publishers
Distinct publishers with included, body-backed reporting in this cluster.