The AI Implementation Gap

In the Census Bureau's Business Trends and Outlook Survey (BTOS), businesses that reported using AI were asked what changes they had made in the last six months in order to use it. Nationally, 64.3% selected “This business did not make any changes to use AI.” 15.4% reported developing new workflows, and 15.0% reported training current staff.

The headline is not that AI is failing. It is that, for most AI-using businesses, adoption had not yet been accompanied by the kinds of deliberate organizational change the survey lists. That distance between using a tool and reorganizing work around it is what this series calls the implementation gap — a descriptive label for a survey pattern, not a measured outcome.

“In the last six months, to use Artificial Intelligence (AI), what changes did this business make? Select all that apply.”

In the published tables this is Q7; in the official questionnaire PDF it appears as item 29. It was asked only of businesses that reported using AI, and respondents could select more than one option, so the percentages overlap and should never be added together.

The printed questionnaire numbers this question as item 29. The Census Bureau's AI supplement data workbook labels the same question Q7 and reports it for Scope 2 — businesses that reported using AI. The values here are the workbook's Q7 Scope 2 estimates.

Denominator

Only businesses that reported using AI — not all businesses.

Period

Changes made in the last six months.

Responses

Select all that apply; shares overlap.

“This business did not make any changes to use AI” is a literal response option. Selecting it does not mean integration failed, that a business is immature, or that AI produced no value, productivity or return. A business may use an off-the-shelf tool that needs no new process, may have made changes before the six-month window, or may have made changes the list does not name.

Industry groups follow the North American Industry Classification System (NAICS). They are broader than their names suggest:

• All U.S. businesses. The national all-industry estimate for AI-using employer businesses — not a sector.
• Real estate and rental and leasing (NAICS 53). Real estate lessors, agents and brokers, property managers, appraisers, plus businesses that rent or lease equipment and consumer goods. Broader than brokerages.
• Construction (NAICS 23). Building construction (residential and nonresidential), heavy and civil engineering, and specialty trades such as electrical, plumbing and roofing. Not only homebuilders.
• Finance and insurance (NAICS 52). Banking and credit intermediation (including, but not limited to, mortgage lending), securities and investments, and insurance carriers and agencies. Not only mortgage lenders.

Among AI-using businesses, finance and insurance stands apart: 54.6% reported no changes, while 22.0% developed new workflows and 23.4% trained current staff. Construction sits at the other end — 75.0% no changes, 10.0% new workflows and 7.5% staff training. Real estate and rental and leasing (65.2%, 14.1%, 14.8%) is close to the national figures.

Each percentage applies only to AI-using businesses in that group. Because the share of businesses using AI also differs by sector, these figures say nothing about how many businesses in a sector overall changed their operations.

The housing economy spans all three sectors shown here: property owners, managers and brokers (NAICS 53), builders and trades (NAICS 23), and lenders and insurers (NAICS 52). The survey cannot isolate residential firms within any of them, so the examples below describe the kinds of change the response options refer to — they are not drawn from the survey.

For housing businesses, whichever option applies, the practical questions are the same: who reviews AI output, how resident and borrower data are protected, and how fair housing and fair lending obligations are kept in view when text or analysis is generated.

Studies that follow workers after an AI tool is introduced offer a different kind of evidence: what happens where a tool is deployed, rather than how many businesses report change. None of them is a national implementation rate, and none is directly comparable to the Census estimates.

In a randomized study across 66 firms and 7,137 knowledge workers, treated workers who used an AI assistant spent about two fewer hours per week on email in the latter half of the study, with no detectable changes in the composition of their tasks (Dillon, Jaffe, Immorlica & Stanton (2025)). Several of the authors have ties to Microsoft, whose tool was studied, and the paper is an NBER working paper rather than a peer-reviewed article.

Among 5,172 customer-support agents, access to a generative AI assistant raised productivity — measured as issues resolved per hour — by about 15% on average, with substantial differences across workers (Brynjolfsson, Li & Raymond (2025)). That is an effect within one firm's support operation, not evidence about how widely businesses have reorganized.

A global survey of 1,491 respondents in 101 nations asked about workflow redesign and AI governance (McKinsey & Company (March 12, 2025)). It is a nonprobability survey of individuals, not a weighted sample of businesses, so its shares cannot be set beside the Census figures.

Read together, the experiments are consistent with the Census pattern in one narrow sense: a tool can change how time is spent without a business recording a formal change in how work is organized. They do not establish that this is what most “no changes” respondents experienced.

Issue 01 asks what AI does: drafting, extracting, retrieving and assisting. Issue 02 asks a different question: how does that output become part of an operation someone can depend on? A convincing answer in a chat window is only one step. Work also needs a trigger, trusted inputs, an accountable owner, a destination for the result and a way to recover when something goes wrong.

Consider three distinct designs, not stages of a maturity ranking. A task assistant helps a person produce an output. An integrated workflow connects that output to review, routing and a system of record. An operating system coordinates multiple workflows under shared permissions, monitoring and change control. A business may reasonably choose the first. More integration is not automatically more value.

Q7 records reported organizational actions, not the architecture behind them. Developing new workflows could mean adding a manual review checklist or connecting several systems; the response cannot distinguish the two. Training current staff records an action, not its quality or effectiveness. Nor can the separate workflow and training shares tell us how many businesses did both.

Integration obstacles are practical design questions, not additional findings from BTOS. Does an invoice reference the same vendor identifier as the ledger? Can a reviewer trace a proposed answer to the document version that produced it? Does the connection preserve the user's access permissions? A model can produce fluent text while these handoffs remain unresolved.

Exceptions deserve as much attention as the normal path. Missing fields, unreadable scans, conflicting policies and service outages need a queue with an owner and a manual fallback. A confidence score is not proof of correctness; test the actual fields or decisions on checked examples. Faster extraction can merely move the bottleneck to an overwhelmed reviewer.

Responsibilities also change. Someone must maintain source documents, approve tool access, investigate errors, train staff and decide when an update needs retesting. These requirements help explain why workflow design and staff training matter operationally, but the survey does not establish that any particular obstacle caused its response pattern.

Illustrative design · Hypothetical property-management office, not an observed firm or measured result

Before: a bookkeeper reads a vendor invoice, enters the fields and sends it for approval. Using an assistant to suggest the amount and due date can help the same task without changing the approval process. It could fit a business reporting no listed changes.

Proposed workflow: an approved inbox receives the invoice; extraction proposes vendor, invoice number, date, amount and account code with references to the original. Deterministic checks compare totals, vendor identifiers and potential duplicates. The bookkeeper verifies fields; missing information, new vendors and suspected duplicates go to a named exception owner. An authorized approver decides whether to post the invoice to the accounting system. Payment authorization remains separate and human-controlled.

Every handoff should leave a record: original document, proposed values, corrections, reviewer, approval and final posting. Bank-detail changes always require independent verification outside the incoming invoice. A failed connection goes to a manual queue, never silently into an unapproved payment. None of this asserts that the hypothetical firm improved its results.

Explore invoice capture and its pilot checklist →

Start with a checked sample and a baseline for the existing process. Run the proposed workflow in shadow mode, then a bounded pilot. Compare similar invoice types and volumes; where feasible use a comparable group without the change. Record differences in staffing, seasonality and document mix. A before-and-after comparison alone does not establish causality.

Use a balanced scorecard. Throughput: invoices fully verified and approved per staff hour, including review and exception time. Quality: incorrect fields divided by checked fields, and incorrect postings divided by reviewed postings. Flow: median elapsed time from receipt to approval and the age of unresolved exceptions. Control: duplicate or unauthorized postings, and overrides with a documented reason. These are proposed KPIs, not observed performance figures or promised targets.

Set acceptance and stop criteria before the pilot with the process owner. Do not scale if faster handling comes with more consequential errors, unresolved access risks or a growing exception backlog. Log staff effort and tool and integration costs alongside the scorecard. An apparent saving in one task can disappear when review, maintenance and recovery are counted.

Define the recurring problem and the output that someone will actually use. If a template or deterministic rule solves it reliably, start there. If the work involves varied language or documents, test whether AI adds something useful on a checked sample rather than treating integration as the goal.

Buy when a standard assistant fits the task and its access and data terms are acceptable. Configure when existing software can support the required templates, review rules and permissions. Integrate when a reliable handoff into an existing system is essential. Consider a custom build only when the requirements are genuinely distinct and an owner can support testing, monitoring and maintenance. These are decision criteria, not recommendations for a particular vendor.

Before proceeding, answer five questions: Are the source data trustworthy and permitted? Who approves the consequential output? Where do exceptions go? What would count as a meaningful improvement without worse quality? Can the team reverse the change and keep operating? If these answers are missing, narrow the pilot or keep the work manual.

The investigation ends with a distinction, not a verdict on adoption. BTOS shows which organizational actions AI-using businesses reported. Workplace studies show effects in specific deployments. A reader's workflow still needs its own evidence. Use the AI Opportunity Explorer to choose a task and examine its data needs, human checkpoints, safeguards and pilot measures before deciding how far to connect it.

The pattern is consistent with several explanations, which this data cannot distinguish:

• Early or light use: tools that slot into existing tasks without process redesign.
• Timing: changes made before, or planned after, the six-month window.
• Unlisted change: adaptations that do not fit the response options offered.
• Sector structure: regulated industries such as finance and insurance may formalize training and procedures more often.
• Firm size and composition: sectors differ in their mix of small and large businesses.

The Census Bureau has also published research on business AI use (working paper CES-WP-26-25, listed in the sources). This issue does not summarize its findings; readers should consult it directly.

1. BTOS Core and AI Content (official questionnaire; Q7 appears as item 29). https://www.census.gov/hfp/btos/downloads/BTOS%20Core%20and%20AI%20Content.pdf
2. U.S. Census Bureau press release, May 21, 2026 (AI supplement release). https://www.census.gov/newsroom/press-releases/2026/btos-may-21.html
3. U.S. Census Bureau, Center for Economic Studies working paper CES-WP-26-25. https://cdn.www.census.gov/library/working-papers/2026/adrm/CES-WP-26-25.html
4. BTOS AI Question Wording Updates (series break notice). https://www.census.gov/hfp/btos/downloads/AI%20Question%20Wording%20Updates.pdf
5. North American Industry Classification System (NAICS), U.S. Census Bureau. https://www.census.gov/naics/
6. Dillon, E. W., Jaffe, S., Immorlica, N., & Stanton, C. T. (2025). Shifting Work Patterns with Generative AI. NBER Working Paper 33795. https://www.nber.org/papers/w33795
7. Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at Work. Quarterly Journal of Economics, 140(2), 889–942. https://doi.org/10.1093/qje/qjae044
8. McKinsey & Company. (2025, March 12). The state of AI: How organizations are rewiring to capture value. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-how-organizations-are-rewiring-to-capture-value

Census values are the author-validated workbook Q7 Scope 2 estimates from the BTOS 2026 AI supplement, released May 21, 2026. Outside studies are summarized only as far as verified and are not comparable to the Census rates.

• If your business uses AI, what — if anything — changed in how work is assigned, reviewed or documented?
• Who is accountable for checking AI output before it reaches a customer, resident or borrower?
• Has anyone been trained on what data may and may not be entered into AI tools?
• Would “no changes” describe your business — and is that a deliberate choice?

Back to The AI Shift