The buildout is easier to count than the payoff.

A board packet arrives with a tab labeled "AI value." Behind the tab are three sheets. The first is a purchase order for servers. The second is a license report: seats bought, seats active last month. The third is a service report: cases resolved, cases reopened, cost per case. The packet is invented for this dispatch, but the stapling is familiar. Three different kinds of evidence travel under one label.

Last Friday, July 17, three Federal Reserve Board economists, Paul Soto, Mason Thieu and Jeffrey Allen, published a note on public data for tracking the AI buildout. They sort the indicators into three layers: capabilities and costs; firm investment and adoption; and productivity and labor. They describe that order as the sequence often associated with general-purpose technologies, and they read the evidence so far as "consistent with a buildout phase rather than the onset of broad-based displacement."¹ The note is a measurement framework. It is not an estimate of what AI has done to the economy, and it is not a forecast.

My argument is that a board should keep the same layers apart. Investment and adoption indicators show that money was committed and that people are using what it bought. Demonstrated economic outcomes show that something about the service or the business changed. A strong reading in one layer is a good reason to look at the next. It is not a substitute for looking.

What each sheet establishes

The purchase order establishes a commitment of resources. The license report establishes activity under a definition, and the definition matters: "active" might mean one session in a month or work done every day. The service report establishes what happened to the work, measured some particular way.

The chain can break at each joint. Servers can be delivered and sit idle. Seats can be busy while case quality stays where it was. Quality can improve while integration and supervision costs absorb the savings. These are scenarios, not findings from the Fed note, but each describes a different failure and calls for a different fix.

Even the blue sheet does not settle attribution. Resolved cases might rise because the tool helps, because the case mix got easier, or because the team grew. To credit the tool, the board needs a comparison: a group that did not get it, a staggered rollout, or at least a before-and-after with the other changes written down.

Keep the denominator still

The note is candid about how hard its own investment layer is to scope. Hyperscaler capital spending has become a leading proxy for AI infrastructure, but it includes spending that has nothing to do with AI. One example in the note: at the end of 2025, only about a third of Amazon's gross property, plant and equipment was servers and networking equipment. Leased data center capacity can make the headline figures understate the buildout instead. The national-accounts series for computer equipment includes ordinary office computers.¹

The same discipline applies inside a company. A chart of all infrastructure spending does not become an AI total because someone retitled it. If last quarter's "AI spend" counted only model licenses and this quarter's adds the data platform, the growth is partly a change of definition. Write down the scope, and keep it fixed long enough to compare.

Use is not yet payoff

The note's adoption layer shows why. The Census Bureau's business survey shows the share of firms reporting AI use trending upward, but headline adoption says nothing about intensity, which other surveys suggest remains shallow.¹ Vendor telemetry has the same limit at higher resolution. Anthropic's June Economic Index report describes in fine detail how people use Claude, and its accompanying survey of users says plainly that it is "not representative of the general population."² Usage data describes a product's users. It is valuable for that, and it is still a use sheet, not an outcome sheet.

The reverse error is just as tempting. The note finds productivity trends across sectors relatively consistent over time, suggesting that gains measured in micro-level experiments are not yet adding up in aggregate. It adds that the absence of a large aggregate signal would not invalidate future effects, since gains from general-purpose technologies have historically lagged investment by years.¹ A flat national series does not prove a particular support team's tool failed. A national series and a team decision have different scopes, and each needs evidence at its own scale.

Objections, considered

Waiting for macroeconomic proof would stop useful projects. It would, and it is not the standard. A local decision can rest on credible local evidence. The point is to match the size of the claim to the reach of the evidence.

Adoption is a leading indicator. It can be. Then label it as one, and say which outcome it is supposed to lead.

Benefits take time. Often they do. Then state the expected lag and the intermediate changes that should appear along the way, so that "not yet" has a date attached.

A bigger buildout is evidence of a bigger commitment before it is evidence of a bigger payoff.

Before the next AI value review

• Which sheet supports this claim? Ask whether a statement about value rests on spending, use or an observed outcome.
• Did the denominator change? Ask for the scope and definition behind each number, and whether either moved since the last report.
• What comparison links use to the outcome? Ask what would show the tool, rather than a changed case mix or headcount, moved the result.

Back to the packet. The tab labeled "AI value" comes off. In its place go three tabs: commitment, use, outcome. The sheets are the same. Now each says only what it can show, and the board can see which one is still missing a comparison.