What the study actually measured
Carnegie Mellon University and workforce analytics firm Larridin scored roughly 500 public companies on a five-point scale based on how specifically each described its AI deployments in SEC 10-K filings. A score of one meant boilerplate language about "exploring AI opportunities." A score of five meant named use cases, adoption figures, and quantified business outcomes. The researchers controlled for industry, company size, and prior revenue growth, and they re-ran the analysis excluding AI-chip leaders like Nvidia, Broadcom, AMD, Micron, and Intel to make sure semiconductor demand was not driving the result. The conclusion held either way.
Companies scoring in the top disclosure categories showed an 8 percentage point revenue-growth advantage year-over-year over companies with minimal disclosure detail. More than 150 companies landed in the two highest scoring bands, but only five received the maximum score reserved for quantified business results, the kind of disclosure that names a metric and a number rather than a capability. That scarcity is itself a finding: even among companies confident enough to talk about AI in a regulatory filing, precision about business impact remains rare.
The limits of the finding matter as much as the headline
The study explicitly found no correlation between disclosure specificity and either operating margin improvement or stock performance. That is an important qualifier, because it means the finding is not "specific AI disclosure makes companies more profitable." It is closer to "companies specific enough to disclose real AI metrics tend to also be growing revenue faster," which could reflect selection as much as causation. Companies with genuinely working AI deployments may simply have more to say, rather than the disclosure itself driving the growth.
Larridin founder Ameya Kanitkar described the shift underneath the numbers as companies "moving beyond AI experimentation and are increasingly focused on identifying 'high-value' use cases." That framing is closer to what the data supports: disclosure specificity is a visible symptom of measurement maturity, and measurement maturity is plausibly what correlates with revenue impact. CIOs should read the study as evidence that rigor is measurable from the outside, not as a guarantee that better writing produces better numbers.
What good disclosure looks like in practice
Visa's filing, which scored at the top of the scale, reported that 26,000 employees use AI-powered chat tools internally, alongside 17 percent year-over-year revenue growth. Naming a specific headcount using a specific capability is the kind of detail that separates a five from a two and a half, because it is checkable, comparable across quarters, and impossible to write without an underlying measurement system that actually tracks adoption. Conagra Brands, by contrast, scored 2.5 for discussing AI in general terms without specifics, and posted a 2 percent revenue decline over the same period.
Neither example proves causation on its own, but the pattern across the full sample is consistent enough to take seriously. Companies that can name a metric in a public filing are, almost by construction, companies that built the internal reporting to produce that metric in the first place. The disclosure is downstream of the operational discipline, not a substitute for it, which is exactly why regulators and investors increasingly treat vague AI language in filings as a soft signal of an immature program.
Why this lands differently after eighteen months of AI pilots
Enterprises have spent roughly two years running AI pilots, and 2026 is the year a meaningful share of those pilots are supposed to reach production. That timing makes this study land at a useful moment. Boards and investors are no longer satisfied with disclosures that describe AI as a strategic priority; they want the same kind of quantified, auditable detail companies already provide for other capital allocation decisions. A CIO who cannot produce a number for what an AI deployment actually changed is, in this framing, indistinguishable from one whose deployment has not produced a measurable change at all.
That pressure will only increase as more companies clear the top disclosure bands and the bar for what counts as specific keeps rising. The five companies that hit the maximum score this year set a reference point that next year's filings will be measured against, whether or not regulators formalize any new requirement. Waiting for a disclosure mandate before building the underlying measurement capability is the wrong sequencing. The filing requirement, if it comes, will simply expose whoever has not done the work already.
The build-vs-report decision this creates for CIOs
The direct implication for CIOs is not about investor relations copywriting. It is about whether the organization has instrumented its AI deployments well enough to produce a defensible number when someone, a CFO, a board member, or eventually a regulator, asks for one. That requires adoption tracking, use-case-level outcome measurement, and a process for rolling those numbers up into language finance and legal can put in a filing without hedging every sentence into meaninglessness.
Organizations that treat AI measurement as an afterthought will keep producing the vague, defensible-sounding disclosure language that scored a two or a 2.5 in this study, the same language regulators, boards, and now researchers can increasingly detect and penalize. Building the measurement infrastructure now, ahead of any mandate, gives CIOs a genuine choice about what the next filing says. Building it only when compliance requires it means the filing will say whatever the incomplete data happens to support.
