The Fed Just Fact-Checked Every AI Earnings Call. The Results Aren't Good.

Listen to enough earnings calls and you start to notice the verb tense.

Will. Expects to.

Economists at the St. Louis Fed just tested that instinct at scale. Their study, published July 31, 2026, ran an AI classifier across roughly 490,000 earnings call transcripts from 5,198 public U.S. companies going back to 2000, sorting every sentence about productivity. Of the sentences that mention AI, 95% describe gains that have not arrived yet. For productivity talk with no AI in it, the figure is about 75%. Companies always talk up the future. On AI, they talk it up harder.

Then there is the number sitting underneath. Actual productivity growth measured in that same data came to 0.07% per quarter.

Behind those 2 figures is someone you know. An operations manager was told the new tools would give her team its evenings back. She is still at her desk on Thursday night, same as last year, reading a deck that says her function has been transformed. She is not behind. Almost nobody is ahead.

I spent 20 years running innovation programs inside a Fortune 100, and I approved budgets that read exactly like those sentences. The money was real. The intent was real. The payoff was scheduled for a quarter that kept moving.

The spending is not waiting for the proof.

On July 30, Microsoft disclosed $130 billion of new data center leases signed in a single quarter, with quarterly capex up 70% from a year earlier, and Satya Nadella said the company is "on track to roughly double its data center capacity over two years." The same day, Meta's filing showed $279 billion in future AI data center lease commitments, up 53% from the prior quarter, structured through special purpose vehicles so the obligations stay off the balance sheet until the leases begin.

Nobody in this story is bluffing. These are deliberate bets by operators with better data than most of us will ever see, and compute bought early is compute you own when everyone else wants it. The gap sits downstream of the purchase, in what an organization can absorb.

Factories have run this experiment before. The first plants to buy electric motors bolted them onto the design they already had: one big engine, one long driveshaft down the length of the building, belts dropping to every machine. The power source changed. The floor plan did not. Output moved decades later, when new plants were built in the order the work actually happened, and when the people on the floor were taught the new way through it. Paul David's research on the dynamo is the standard account of that lag.

Most of us have run a small version of it at home. You can replace every appliance in a kitchen and still take the same 6 steps to make dinner, because nobody moved the sink.

This is the say/do gap, and the St. Louis Fed has now put a number on the say side. 95% future tense. 0.07% a quarter of real movement. A company can close that gap 2 ways: rebuild the work around the tools, or keep buying capacity and hope the work rebuilds itself.

Capex is a signature. Productivity is a rebuild.

Which raises the question I would want settled before signing the next check. If the payoff lands 3 years later than planned, who carries the wait? Compute contracts run for years. Retraining budgets get approved one year at a time, and they are the first line cut when a quarter goes sideways. The half-life of the skills on your team keeps shortening either way. That is the AI leadership question of this cycle, and no vendor can answer it for you.

If you do not hold a budget, the question still belongs to you. When your company bought its AI tools, what did it buy for you? Ask it plainly, out loud, in a room with your manager in it. Companies worth staying at can answer.

Multiply these decisions across a few thousand companies and they settle things no board ever votes on. Whether the analyst who is 12 years from retirement gets a bridge to her next decade or a severance letter. Whether the gains show up as longer careers and steadier towns, or as one more round of hard news explained at a kitchen table. Economists will score this decade in output per hour. Families will score it by whether the work held.

This is the first thing I ask when I walk into a leadership room, and you are welcome to borrow it. Take one sentence about AI from your last board update. Try to rewrite it in the past tense, with a number attached and a name beside it. One workflow, changed, measured, owned. Write that sentence today and you are further along than 95% of the calls the Fed just read. Come up empty and you have a great deal of company, plus your first project. And if you read the data differently, I am easy to find.

Sources: Federal Reserve Bank of St. Louis, "AI and Productivity: What Firms Say on Earnings Calls," July 31, 2026 · Microsoft Q2 FY2026 earnings, July 30, 2026 (Bloomberg / Irish Times) · Meta Q2 2026 SEC filing, July 30, 2026 (Bloomberg / CryptoBriefing) · Amazon Q2 2026 earnings, July 31, 2026 (DataCenter Knowledge) · European Commission AI gigafactory call, July 30, 2026 (Washington Post).

Here is what makes Alex a credible voice on this topic: he approved AI and innovation budgets inside Cisco for 20 years, building the $1.1B portfolio that taught him exactly how long the gap runs between a spending announcement and a real result.

Take this question to your team. Book a conversation →