The AI Implementation Gap

About this report

An honest, evidence-based look at what AI adoption actually delivers.

This report was researched and written by the Avolis Research Group in mid-2026. It synthesizes more than forty sources, including U.S. Census Bureau data covering roughly 1.2 million business survey responses, peer-reviewed economic studies, and adoption research from McKinsey, RAND, BCG, Gartner, IDC, the U.S. Chamber of Commerce, Intuit, Goldman Sachs, and the Federal Reserve. Every statistic is cited. Where a widely circulated figure could not be verified against a primary source, we either excluded it or flagged its limitations in the text, including some of the most famous numbers in the AI conversation.

Our aim is simple: give owners and operators of small and mid-sized businesses an honest, evidence-based picture of what AI adoption actually delivers, why so many efforts stall, and how to run an adoption process that pays for itself.

Executive summary

Key findings at a glance

1. 01Small-business AI adoption looks like a boom, until you define "adoption."Self-reported use has roughly tripled in three years: generative-AI use grew from 23% in 2023 to 58% in 2025 ^[4], and 77% of U.S. SMBs told Intuit QuickBooks they use AI "regularly" as of January 2026 ^[5]. The Census Bureau, using a stricter definition, finds only about 20% of firms use AI in any business function ^[1], and the share with AI fully embedded in operations is just 9% ^[7] to 14% ^[8].
2. 02The returns are real for the businesses that get past experimentation.A peer-reviewed study of 5,179 support agents found AI lifted productivity 14–15%, with gains of 34% for newer staff ^[11]; a Science experiment found professional writing 40% faster ^[13]; AI demand forecasting cuts error 20–50% ^[14]. Among SMBs using AI, those reporting revenue gains outnumber those reporting declines roughly 20 to 1 ^[5].
3. 03Most AI initiatives still return nothing, and the cause is rarely the technology.A widely covered MIT project estimated ~95% of enterprise generative-AI pilots showed no measurable P&L impact ^[18]; S&P Global found 42% of companies scrapped most AI initiatives in 2025, up from 17% ^[21]; RAND attributes failures to leadership, problem-selection, and data ^[19]. BCG's rule of thumb: AI value is roughly 10% algorithms, 20% technology, and 70% people and process ^[20].
4. 04What separates the winners is consistent across every study.They pick one painful, measurable problem instead of "doing AI." They redesign the workflow around the tool, the single biggest driver of bottom-line impact in McKinsey's 2025 survey ^[22]. They measure baselines and track KPIs, a practice followed by fewer than 1 in 5 companies ^[22]. They put an owner on it. And they usually buy and adapt; partnered deployments succeeded about twice as often as internal builds ^[18].
5. 05The window for "wait and see" is closing faster than the adoption numbers suggest.The top reason non-adopters give, "AI isn't applicable to my business," is collapsing: from 81% of non-adopting firms in 2023–24 to 65% in 2025–26 ^[2,3]. Growing SMBs are far more likely to be investing in AI than declining ones (83% have adopted; 78% plan to increase investment, versus 55% of declining peers) ^[9]. The real question is how soon your workflows will be AI-assisted, and whether yours will be before your competitors'.

01 · The state of AI adoption

Adoption in operations-heavy small and mid-sized businesses

1.1Three numbers that can't all be true

Ask "how many small businesses use AI?" and you'll get answers ranging from 17% to 89%. All of them are, in their own way, accurate. Understanding why they differ is the fastest way to understand where the market actually is, and where you actually are.

• ~89% of small business owners said they use AI tools (Intuit/ICIC, 2025) ^[6]. This is the loosest definition: any use at all, including asking ChatGPT to draft an email once a month.
• 77% report using AI "regularly" (Intuit QuickBooks 2026, 34,000+ responses) ^[5].
• 58% use generative AI specifically, up from 23% in 2023 (U.S. Chamber of Commerce) ^[4].
• ~20% of all U.S. firms use AI in any business function (Census Bureau BTOS, the most representative measurement) ^[1].
• 9–14% have AI embedded in core operations, integrated into daily workflows with process change around it (SAS/IDC; Goldman Sachs) ^[7,8].

Here is the honest synthesis: most small businesses have experimented with AI, while operational adoption remains rare. A large minority use it regularly for individual tasks such as drafting, summarizing, and marketing content. Only about one business in ten has done the harder thing: wired AI into an operational workflow (quoting, scheduling, fulfillment, underwriting, support) in a way that changes how the work gets done and shows up in the numbers.

That gap between using AI and getting operational value from AI is the subject of this report. We call it the implementation gap.

1.2The adoption curve is steep, and uneven

However you define adoption, the direction is unambiguous. The Chamber's tracking shows generative-AI use rising 23% → 40% → 58% across 2023–2025 ^[4]. QuickBooks' panel shows regular use climbing from 48% to 77% in the eighteen months to January 2026 ^[5]. But the growth is not evenly distributed:

• Size matters, and mid-sized firms are pulling away. As of May 2026, 37% of firms with 250+ employees and 32% with 100–249 use AI in at least one function, versus roughly 17% of firms with fewer than 20 employees, with no statistically significant growth among the under-20 group over the prior six months ^[1,30].
• Sector matters less than you'd think, but it matters. Information (39.7%) and finance & insurance (33.9%) lead; retail (~14%) trails the 19.8% national average ^[1]. Operations-heavy industries cluster near the middle, so competitive separation is still up for grabs.
• Employment-weighted, the picture flips. Because large firms adopt more, roughly 32% of employment is at AI-using firms even though only ~18% of firms use AI ^[2]. Your biggest competitors are disproportionately likely to already be adopting.

1.3What's driving adoption

Three forces come up in every credible survey. Labor pressure: 56% of small businesses believe AI can offset reduced or frozen headcount, and 53% believe it helps retain staff by removing drudge work (Verizon Business) ^[10]; adoption is not translating into job cuts, and 82% of AI-using small businesses increased their workforce over the past year ^[4]. Measured results: 84% of AI-using small businesses cite efficiency as the primary benefit ^[8]; 78% say AI improved productivity, up from 46% in July 2024 ^[5]. Competitive visibility: 83% of growing SMBs have adopted AI and 78% plan to increase investment, versus 55% of declining peers ^[9]. There is also a quiet trap: 80% of AI users believe AI use is common among peers, while only about a third of non-users think so ^[9]. If you're not using AI, you may be underestimating how many of your competitors are.

1.4What's blocking adoption, and the barrier that's collapsing

For two years, the top reported reason for non-adoption was simple: "AI isn't applicable to my business." In the Census Bureau's 2023–24 supplement, 80.9% of firms not planning to use AI said exactly that ^[2]. By the 2025–26 supplement, that figure had fallen to 65%, a 16-point drop in two years ^[2]. What remains are the practical barriers this report exists to address:

• Data privacy and security: the most consistent top-3 barrier across QuickBooks, Upwork, and PayPal surveys (27–49%) ^[5,17,24].
• Limited knowledge of AI capabilities: ~20% of non-adopting firms ^[2]; 73% of small businesses say training and implementation resources would help them most ^[8].
• ROI uncertainty: the #2 barrier to AI-agent adoption at 24% ^[17].
• Difficulty choosing tools: of thousands of vendors claiming to sell "agentic AI," Gartner estimated only about 130 offered the real thing ^[23].

1.5Where this leaves the operations-heavy SMB

If you run an operations-heavy business and haven't captured value from AI yet, the data says three things. First, you are in the majority: roughly 9 in 10 SMBs are in the same position. Second, the excuse inventory is shrinking: the tools have reached your industry and your software stack, and "not applicable to my business" is a position your competitors are abandoning at about 8 points a year. Third, the winners tend to be the better-implemented businesses. The rest of this report is about what that means in practice.

02 · The ROI case

What well-implemented AI actually delivers

2.1Separating the evidence from the hype

AI benefit claims range from sober to absurd, so this section is deliberately layered. Tier 1 is rigorous, peer-reviewed or controlled research: the numbers you can take to the bank. Tier 2 is large-scale survey data, self-reported but consistent across independent sources. Tier 3 is consultancy benchmarks and named case results, directionally credible and labeled clearly.

2.2Tier 1: What controlled studies show

• Customer support: +14–15% productivity, +34% for newer staff. The landmark study of 5,179 agents at a Fortune 500 firm (Quarterly Journal of Economics) found AI raised issues resolved per hour 14–15%. Novice workers improved 34%, while the most experienced barely changed; AI effectively transferred top performers' know-how to everyone else ^[11].
• Professional writing: 40% faster, 18% better. A preregistered Science experiment found mid-level writing tasks completed 40% faster with 18% higher quality, with the weakest writers gaining most ^[13].
• Knowledge work: +12% completion, 25% faster, inside AI's competence zone. A Harvard/BCG experiment with 758 consultants found gains on suitable tasks, but a 19-point drop in accuracy when people trusted AI beyond its capability ^[12]. Knowing where AI is strong is itself an implementation skill.
• Whole-economy reality check. St. Louis Fed research estimates generative AI saves the average using-worker about 2.2 hours per week ^[15]. A University of Chicago study of 25,000 Danish workers found chatbots handed out without workflow change produced time savings of only ~3% ^[16]. Tools without implementation produce rounding-error results.

2.3Tier 2: What small businesses report

Survey data is self-reported, but the pattern across four independent large samples is hard to dismiss. Revenue: 43% of AI-using SMBs say it increased revenue and 2% say it decreased, roughly 20-to-1 positive ^[5]; 91% of AI-using SMBs say AI boosts revenue ^[9]. Productivity: 78% report improved productivity, up from 46% eighteen months earlier ^[5]. Margins and scale: 86% report improved margins; 87% say AI helps them scale ^[9]. Enterprise benchmark for context: a Gartner survey of 822 leaders found generative AI delivering, on average, 15.8% revenue increase and 22.6% productivity improvement among companies that got it working ^[23]. One caveat: successful adopters are more likely to be growing anyway, so treat Tier 2 as evidence that well-implemented AI correlates strongly with performance, short of a guarantee.

2.4Tier 3: Function-by-function operational benchmarks

For operations-heavy businesses, the most useful numbers are process-level. These are consultancy and case benchmarks: directional ranges, offered as guides.

The Klarna caveat, and why it matters for you

Klarna's AI assistant famously did the work of ~700 agents in its first month, cutting resolution times from 11 minutes to under 2 ^[25]. A year later, Klarna publicly rehired human agents after cost-driven over-automation damaged service quality ^[25]. Klarna kept the AI and still attributes tens of millions in annual savings to it, so the real lesson is about implementation judgment. Knowing which conversations to automate and which need a human is what separates a headline win from a public walk-back, and that judgment is precisely what most first-time adopters lack.

2.5What "ROI" realistically looks like for an SMB

• 20–30% efficiency improvement in the targeted process, consistent with the controlled-study range and practitioner benchmarks.
• Payback within months. Enterprise research found average deployment under 8 months, value within 13, and $3.70 returned per $1 invested in generative AI ^[31]. SMB deployments using off-the-shelf tools are typically faster; the paths in Section 5 target results inside 90 days.
• Compounding second-order gains. AI lifts less-experienced staff most ^[11,13], which means faster onboarding, less key-person risk, and more consistent quality.

03 · The implementation gap

Why most AI efforts return nothing

3.1The failure numbers, honestly presented

In August 2025, a report from MIT's Project NANDA went viral with a single statistic: despite $30–40 billion in enterprise investment, 95% of generative-AI pilots were producing zero measurable P&L return ^[18]. Honesty requires a caveat the coverage mostly skipped: the study was preliminary, non-peer-reviewed, and defined success narrowly. Critics reasonably called the figure directional. We cite it for its diagnosis, which matches everything else in the literature, even though the headline number itself is contested.

These studies are enterprise-weighted; no rigorous SMB-specific abandonment rate has been published. But the SMB-adjacent evidence rhymes: a PayPal survey of ~1,000 U.S. small businesses found 82% call AI essential to staying competitive, while only 25% have integrated it into daily operations, and 51% remain self-described "explorers" ^[24].

3.2Why AI efforts fail: five patterns, all organizational

Across RAND's interviews, MIT's deployment analysis, McKinsey's global surveys, and BCG's capability research, the same failure patterns recur so consistently they amount to a checklist of what to avoid.

3.3The shadow-AI economy: adoption is happening with or without you

MIT found that while only ~40% of companies had purchased an official AI subscription, employees at over 90% reported regularly using personal AI tools for work ^[18]. Microsoft and LinkedIn's 31,000-worker study found 75% of knowledge workers using AI at work, and 78% of them bringing their own tools without company involvement ^[32]. For a small business, the realistic choice is between managed and unmanaged AI: customer data pasted into personal chatbot accounts, inconsistent quality, and no capture of the learning. Doing nothing leaves AI in your business, ungoverned.

3.4What the successful minority does differently

Flip every failure pattern and you get the success playbook, which is consistent across the research:

• Start from one painful, measurable, recurring problem, and commit to it for a year instead of a demo cycle ^[19].
• Redesign the workflow around the tool: McKinsey's top driver. The unit of adoption is the process ^[22].
• Measure a baseline first, then track relentlessly: the highest-leverage practice, and one almost nobody follows ^[22].
• Give it an owner with authority: an operator or the owner, never "whoever's youngest" ^[19,22].
• Spend on adoption as well as acquisition: the 70% in 10-20-70 ^[20].
• Buy and adapt before you build: partnered deployments succeeded ~67% of the time versus ~33% for internal builds ^[18].
• Aim at the back office too: MIT found the best-documented ROI in unglamorous automation, well away from the sales-and-marketing spend that soaked up half of budgets ^[18].

04 · On the ground

Use cases across the adoption journey

The examples below are organized by adoption stage, since the stage you're at matters more than the sector you're in. Sources are labeled: independent research, trade press, or vendor-published case study.

4.1Stage 1 · First workflows: automating a single choke point

The pattern: one high-volume, rules-plus-judgment task; off-the-shelf vertical software; measurable inside 90 days.

• Quoting & estimating (construction & trades). An independent University of Kansas study found AI-assisted takeoff software delivered up to 76% time savings ^[26]. NC Painting went from 19 to 60 bids per month within 60 days; Total Flooring cut a high-rise takeoff from two weeks to under two days and won the contract (Togal.AI, vendor-published) ^[26]. Estimating hours were the bottleneck on bid volume, so cutting them raised the number of at-bats.
• Customer conversations (e-commerce & services). Small merchants report 30–56% of routine tickets resolved automatically within two months (Gorgias, vendor-published) ^[33]. The realistic first-quarter target is the routine 30–40%: "where's my order," "what are your hours," "can I reschedule."
• Documentation & admin (healthcare & professional services). A randomized controlled trial in NEJM AI found ambient AI documentation saved clinicians 30 minutes per provider per day and reduced burnout (UW Health) ^[29]. A 13-lawyer firm reports AI drafting a response to a 50-page expert report in under an hour, previously half a day ^[34].

4.2Stage 2 · Operational integration: AI inside the core process

Here AI becomes part of how the operation runs, where the 20–30% process gains live and where implementation discipline starts to matter more than tool choice.

• Freight quoting (logistics). Ward Transport & Logistics integrated an AI rate engine; carriers on the same platform auto-rate 98% of shipments with no human intervention ^[35]. Enterprise broker C.H. Robinson now turns quote requests around in about 32 seconds, versus 17–20 minutes manually ^[35].
• Underwriting (lending). FORUM Credit Union estimates it can process up to 70% more loans without adding staff ^[28]. Eagle Community CU cut underwriting time 65% on non-W-2 income files ^[27]. What's automated is the processing, and the credit decision stays with people.
• Leasing operations (real estate). Communities using an AI leasing agent outperformed local-market occupancy by 2 points over 12 months, validated against independent third-party data ^[36]. Two occupancy points on a 200-unit portfolio is real money.
• Inventory & forecasting. The McKinsey ceiling is 20–50% error reduction ^[14]; the floor is that only 23% of SMBs use AI for inventory at all ^[37]. Forecasting is one of the highest-ROI and least-adopted operational use cases right now.

4.3Stage 3 · Compounding: multiple workflows, shared data

Few SMBs are here yet: only ~9–14% have AI embedded even in one core process ^[7,8]. The businesses that are here share a trajectory: they ran the Stage 1 to Stage 2 loop repeatedly, and each workflow made the next easier because the data, the habits, and the trust were already in place. NIST's work with small manufacturers offers the clearest picture: one automotive deployment cut unplanned downtime from 42 to 22 hours per month, with ROI inside 90 days ^[38,39]. NIST's guidance doubles as a summary of this report: "find the pain, find the gain" ^[39].

4.4What these examples have in common

• The automated task was high-volume and widely disliked. Nobody mourned manual takeoffs or bank-statement annotation.
• A human stayed at the judgment layer. AI processed and people decided. The one famous counterexample, Klarna's over-automation, became famous precisely for breaking this rule ^[25].
• The tool was bought and configured: vertical software fitted to the workflow ^[18].
• The metric existed before the tool did. Bids per month, minutes per file, occupancy, downtime hours.
• The first win funded the second. Every Stage 3 business was once a Stage 1 business that measured its results.

05 · The playbook

A readiness scorecard and a 90-day path

This section turns the research into something you can run on Monday. It has three parts: a readiness scorecard to locate yourself honestly, a use-case selection filter, and a 90-day implementation plan with decision gates.

5.1The Avolis Operational AI Readiness Scorecard

Research consistently finds that readiness predicts outcomes more reliably than enthusiasm: SAS/IDC found ~70% of SMBs stuck in the two earliest maturity stages ^[7], and RAND and BCG locate failure in exactly the dimensions below ^[19,20]. Score your business 1–5 on each dimension, where 1 means "not true of us" and 5 means "demonstrably true."

5.2Choosing the first workflow: the four-filter test

Score each candidate process against four filters. Candidates that pass all four are rarer than you'd expect, which is the point.

Classic first workflows that pass: quote and estimate prep, inbound customer questions, document intake and extraction, scheduling and dispatch, invoice follow-up, demand forecasting, job documentation. Buy vs. build: buy configurable vertical software and spend your effort on integration and adoption (partnered builds succeed ~67% of the time versus ~33% for internal) ^[18]; expect the true first-year cost at 2.5–5x the subscription line once integration, cleanup, and training are counted ^[40]. MIT found back-office automation out-returned the sales-and-marketing spend that soaked up half of budgets ^[18].

5.3The 90-day implementation plan

Day 90 decision gate: three honest outcomes.

Success targets: >70% voluntary weekly usage · 20–30% improvement on the baseline metric · one written SOP that survives the owner going on vacation. Hit those three and you have what almost none of your competitors do: a proven, repeatable adoption process.

5.4The honest role of outside help

Most of this playbook is DIY-able, and businesses scoring 20+ can run it themselves. The research is also clear about where outside expertise changes outcomes: partnered deployments succeed about twice as often ^[18], 73% of small businesses say implementation and training support is what they most need ^[8], and the failure patterns in Section 3 are exactly the kind an experienced implementer has already seen. Whether that help comes from a vendor's onboarding team, a fractional operator, or a partner like Avolis, one rule holds: pay for workflow change and measurable outcomes, and be wary of "AI strategy" sold by the slide.

Conclusion

The gap is the opportunity

Strip away the hype and the counter-hype, and the 2026 evidence base says something clarifying for the operations-heavy SMB.

The technology works: controlled studies keep finding double-digit productivity gains, concentrated exactly where SMBs need them, in making average performers very good ^[11,13]. Adoption is spreading quickly in name and slowly in substance: three-quarters of small businesses "use AI," while roughly one in ten has made it operational ^[5,7,8]. The difference between those groups is implementation, one measured workflow at a time, with an owner, a baseline, a redesigned process, and a 90-day verdict.

That is the implementation gap. It is also, for a while longer, an open competitive window. The "not applicable to my business" era is ending at a measurable rate ^[2], and the businesses crossing the gap now are compounding advantages (faster quotes, faster answers, cleaner books, better forecasts) that are differentiators today and will be table stakes in a few years.

You now have the same playbook they're using. The scorecard in Section 5 takes twenty minutes. The first baseline takes two weeks. The verdict takes one quarter.

Sources

Numbers correspond to bracketed citations in the text.