AI Readiness Isn't What Your Adoption Rate Says It Is

The Census Bureau's Business Trends and Outlook Survey produces two very different AI adoption numbers depending on one thing: how the question is worded. Ask whether a firm uses AI in "any business function" and the reported rate climbs. Ask whether AI is deployed in production — meaning integrated into an actual operating process — and the number drops substantially. Same firms. Same period. Different question, different answer.

That gap is not a measurement artifact. It is the whole problem.

What founders count as readiness

When a 35-person firm's marketing coordinator uses ChatGPT to draft email sequences and the operations manager runs meeting summaries through a transcription tool, the founder sees AI adoption. The staff is using the tools. The tools are working. The mental model that follows is that the company is ready for deeper AI investment — maybe a CRM integration, maybe automated reporting, maybe something more ambitious.

The OECD's D4SME 2024 survey draws a hard line between exploratory deployment and production deployment. Exploratory means tools are in use. Production means AI output feeds into a business process that depends on it. Most small firms sit firmly in the first category and believe they are in the second. That belief is what drives the capital allocation error.

The maturity distribution nobody talks about

SAS and IDC maturity data show that most SMBs cluster at early, non-integrated stages. Pertama Partners' work on mid-market firms in Asia finds the same pattern: only a small minority reach what the research calls integrated or AI-native states. These are not firms that failed to try. They are firms where active tool use did not convert into process integration, and where that conversion stalled because the data foundations were not there to support it.

The mechanism is worth being precise about. A production-grade AI deployment needs clean, consistent data that flows across functions. A 40-person firm where customer records live in the CRM, invoice history lives in the accounting tool, and support tickets live in a separate inbox has three different counts of the same customer and no reliable way to aggregate them. An AI tool sitting on top of that structure does not become more capable because the team uses it enthusiastically. It produces outputs that are only as coherent as the inputs.

This is where the Census BTOS question-wording effect becomes diagnostic rather than just interesting. The firms that report high adoption under the broad question and near-zero production deployment under the narrow one are not lying. They are accurately describing two different things. The broad question captures genuine tool use. The narrow question captures something the firm has not built yet.

When tool enthusiasm doesn't close the production deployment gap

The strongest version of the counterargument runs like this: leadership commitment and worker skills are the decisive variables for AI success, not data infrastructure. Academic studies from Thailand, Indonesia, Bangladesh, the Czech Republic, and Austria cited in OECD D4SME commentary reach this conclusion repeatedly. If a founder's team is engaged and experimenting actively, the argument goes, that firm is building the organizational capability that matters most. Calling this a capital allocation error misidentifies where the problem actually lives.

This argument is correct about what drives learning. It is wrong about what converts learning into operational outcomes.

If leadership commitment and worker skills were sufficient to bridge the gap between tool use and production deployment, the SAS/IDC and Pertama maturity distributions would not show most SMBs stuck at early stages. These are not disengaged teams. The OECD D4SME data explicitly notes that small firms with active experimentation still fail to reach integrated deployment at scale. Enthusiasm and culture are necessary. They are not sufficient when the underlying data structure cannot support what the tools need to function at production level.

A founder who sees her team's active AI use as a readiness signal is not misreading the culture. She is misreading what readiness requires.

The four-stage diagnostic

The readiness framework built from Census BTOS, OECD D4SME, and SAS/IDC data sorts firms into four stages using three observable dimensions: firm size, integration depth, and data maturity.

Stage one firms use AI tools in isolated functions with no data flowing between systems. Stage two firms have begun connecting outputs from one tool to inputs in another but rely on manual transfers — someone exports a CSV and uploads it somewhere else. Stage three firms have automated at least one data connection and AI output feeds a repeating business process without human intervention. Stage four firms have AI integrated across functions with consistent data structures that allow cross-functional analysis.

The Census BTOS firm-size data show that production AI use remains low even among firms reporting broad tool adoption, and that this pattern holds across size classes in the 5–50 employee range. The SBA's work on SME digital tool access confirms that barriers are not primarily about tool availability — the tools are accessible. The barriers are about the data plumbing that makes tools useful at production scale.

Where to put the next dollar

If your firm is at stage one or two, adding more tools does not move you to stage three. What moves you to stage three is picking one process where AI output needs to feed a downstream step, identifying why that connection does not exist yet, and building it. Usually the answer is that two systems hold the same data in incompatible formats, or that nobody owns the data quality for a field both systems need.

Stage three firms face a different problem. The single automated connection works, but it is fragile — built around one person's knowledge of how the systems talk to each other. The investment question at stage three is not more tooling. It is documentation and redundancy for the connections already in place.

The OECD D4SME 2024 data frame data governance and foundation work as prerequisites for production AI deployment, not as nice-to-have infrastructure projects. That framing is useful because it reorients the question founders are actually asking. The question is not whether to invest in data infrastructure or tooling. The question is which stage you are at, because the answer determines which investment produces a return and which one produces a more sophisticated version of the same fragile experiment.