700 AI Statements, 2 Measured Outcomes: A Review of Corporate Disclosure
Over seven hundred AI statements across twenty-two major firms in one reporting cycle. Two reported a measured business outcome. A review of what corporate disclosure quantifies, and what it does not.
Listen to the prepared remarks of a major earnings call today, and the focus on artificial intelligence is immediately apparent. Count the statements, and the volume is overwhelming: over seven hundred across twenty-two major firms in the latest reporting cycle. Count the measured business outcomes attributed to the firms' own use of AI, and the number is zero. The gap between how much is being said about AI and what is being quantified is vast, potentially highlighting the challenges of attribution, or perhaps the caution of legal review.
These figures come from a published MKAI study that coded every AI-related statement in the firms' prepared remarks, drawn from earnings calls and, where those remarks were thin, from investor days, annual reports, and filings. The study established specific boundaries for what constituted a measurable outcome:
- Capability: A statement that reported a measured business outcome directly attributable to AI.
- Adoption: A statement about where AI had been deployed and how its use was contained.
- Marketing: A statement about what AI was expected or built to do.
Against this coding, the inventory recorded 404 adoption statements, 308 marketing statements, and just 2 of capability. In fact, the only two statements reporting a measured result came from Philipp Schindler, Alphabet's chief business officer, on the company's earnings call of 4 February 2026. The first:
"For instance, Aritzia, Canada's premier fashion house, used AI Max to find new high-value customers that traditional strategies miss, delivering an 80% incremental uplift in conversion value for Q4."
The second:
"AI Max enabled the L'Oréal Group to maximize its presence across the full consumer journey, fuel its consumer growth, and increase revenue for DTC brands like NYX by 23%."
Each describes a customer using Google's AI Max advertising product rather than anything Alphabet measured from its own operations. The boundary between adoption and capability can be complex. For example, on Salesforce's call of 25 February 2026, Jason Lemkin of SaaStr told investors:
"On Agentforce alone, as a tiny organization, we've closed $2.7 million."
The statement was coded as adoption, because the agent's contribution could not be separated from the workflow around it. The hardest case in the sample comes from Meta's call of 28 January 2026, where Susan Li reported:
"Since the beginning of 2025, we've seen a 30% increase in output per engineer, with the majority of that growth coming from the adoption of agentic coding, which saw a big jump in Q4."
This is a measured gain, yet it was coded as adoption. The reason is the baseline. The two Alphabet statements cleared the threshold because each measures an AI product against an existing alternative. The Meta statement reports output rising as adoption rises, with nothing separating what the model contributed from the effect of more people using more tools. Output is also a quantity, and nothing in the statement indicates whether the additional output was good or how much checking it created downstream.
The count of two is conditional on that threshold, and the study makes this explicit. A looser threshold would admit more capability statements. The study reviews those boundary cases and records that their effect on the aggregate is small: a handful of statements against more than four hundred that count deployment.
Neither the adoption statements nor the marketing statements commit a firm to a result that a later quarter could show to be false. On Accenture's call of 19 March 2026, Julie Sweet said of a set of AI-enabled platform capabilities:
"These capabilities analyze large volumes of data, initiate routine actions, and support better decisions in real time."
No measurement is attached to the word better. Similarly, in JPMorgan's 2025 Annual Report, published on 6 April 2026, the co-chief executives of its commercial and investment bank wrote:
"Over 90% of our engineers now use AI code assistants, and more than 65,000 CIB colleagues actively use LLM Suite, our generative AI platform."
The statement counts users and seats. A measured outcome attributed to AI can be checked, and it can be wrong.
The study measures disclosure only; it makes no claim about reality. It records what these firms said about their AI. A result not reported and a result not occurring are different things. Several explanations could fit the same silence:
- Attribution may be difficult to establish. A firm unable to isolate AI's contribution from everything else moving inside it may decline to claim one.
- The measured return may be small so far. The AI now deployed may have yet to produce outcomes a firm would put its name to.
- The exposure question. A measured claim is the one statement a regulator, an investor, or a litigant could later use.
- Legal review. Prepared remarks are reviewed by lawyers and wrapped in disclaimers, and that review may remove a specific claim of measured outcome before anyone speaks it.
More than one of these factors may be at work, and nothing in the record indicates which. However, the study suggests a clear pattern: these transcripts say where AI has been put to work and what it is meant to deliver, but on what it has actually delivered, they are almost silent.
You can read the full report on SSRN or via MKAI.org/inquiries.