Perspectives · Data Commentary

Half of Small Businesses Now Use AI. Almost None Use It Strategically.

By Rishi Desai7 min readData as of mid-2026

The headline numbers on SMB AI adoption look like a solved problem. Surveys across 2025 put small-business AI usage at roughly 47–58%, up from around 23% in 2023 — a doubling in two years. One tracker reports 68% of small businesses now use AI "regularly," with 28% using it daily. By any headline measure, the adoption race is over.

Read one layer down and a very different picture emerges — one with a direct message for any founder deciding where AI sits on their priority list.

The gap between "using AI" and using AI

The most telling number in the recent data isn't an adoption rate. It's this: when researchers apply a stricter, production-based definition — AI actually embedded in business processes rather than individuals using chatbots — small-business adoption measured around 8.8% as of late 2025. Call it the survey gap: half of SMBs "use AI"; less than one in ten has operationalized it.

Both numbers are true. They're measuring different things. The ~50% measures whether anyone in the company has a ChatGPT habit. The ~9% measures whether AI is wired into the workflows that produce revenue and margin. The distance between those two numbers is, bluntly, where the competitive advantage currently lives.

The performance correlation the surveys keep finding

Two findings recur across the 2025–26 survey wave. First, adoption correlates with trajectory: 83% of growing SMBs report using AI, versus 55% of declining ones. Second, the self-reported returns are striking — 91% of AI-using SMBs say it boosted revenue, 86% say it improved margins.

The consultant's caveat is mandatory here: this is correlation soaked in survivorship and selection bias. Growing companies have more slack to experiment; optimistic operators both adopt tools and grow. A 91% "it boosted revenue" self-report should be read as sentiment, not measurement. But even after heavy discounting, the direction is consistent across every serious dataset: operationalized AI and business performance travel together, and the companies extracting real returns are the ones that connected AI to specific workflows with specific metrics — not the ones with the most subscriptions.

What this means if you run a $1M–$25M company

You are not behind on tools. You might be behind on integration. Buying licenses achieves survey-adoption. Returns come from the ~9% behavior: workflows, baselines, measured outcomes.
The window is real but closing unevenly. With production adoption still in single digits among SMBs, operationalizing even two or three workflows puts you in the top decile of your size class — likely ahead of every direct competitor you actually worry about.
Sequence matters more than speed. AI amplifies your existing strategy, coherent or not. Choose the workflows because the strategy names them, then automate. I've made the fuller argument in AI Won't Save a Bad Strategy.

Founder action: Ask one question of your business this week: "Which single workflow, if it ran 5x faster or cheaper, would visibly change our P&L this year?" That workflow — not the newest tool — is where your AI effort belongs.

Sources & notes — figures drawn from published 2025–26 survey data; methodologies and definitions vary across trackers, which is rather the point of this piece:

U.S. Chamber of Commerce–affiliated and industry trackers reporting ~47–58% SMB AI adoption in 2025 (vs. ~23% in 2023); regular-use figure of 68% from 2026 small-business surveys.
U.S. Census Bureau BTOS-based production-use measure: ~8.8% small-business adoption (Aug 2025) vs. ~10.5% for large businesses.
Growth correlation (83% of growing vs. 55% of declining SMBs) and ROI self-reports (91% revenue, 86% margin): Salesforce and related SMB survey waves, 2025.

Find Your Highest-ROI AI Workflow — Free About the AI Enablement Sprint

The structural advantage smaller companies have

Large enterprises have more capital, more data and dedicated teams. They are still frequently slower to get value from AI, and the reason is structural rather than a matter of competence.

No translation loss. In a large organisation an AI decision passes through several layers before reaching the work. Each layer reinterprets. In a thirty-person company the person who decides is in the room where it is used, often using it.

Fewer systems to integrate. Enterprise AI projects spend a large share of their timeline on integration with systems that predate the internet. A smaller company frequently has four or five SaaS tools with modern APIs.

Governance proportionate to the risk. A large enterprise needs a formal framework because the blast radius is enormous. A smaller company needs a clear, written position — which is a week of work, not a quarter of committee time.

Reversibility. If a workflow does not work, a small company stops using it. There is no programme to unwind, no vendor contract with three years left, no reorganisation to reverse.

Where smaller companies reliably go wrong

The advantages are real and routinely squandered in the same four ways.

Buying tools instead of changing workflows. Twelve subscriptions, no process redesigned. The spend is visible and the change is not, which makes it the most common pattern by a wide margin.

Starting with the most impressive use case. The demo-friendly application is rarely the one with a measurable baseline and a willing owner. Winning small and visibly buys permission for the harder attempt; failing publicly on the ambitious one costs the programme.

No baseline, therefore no proof. Without a before-measurement the value is anecdote, and anecdote loses to a budget question every time.

Leaving the old path open. Universal, and fatal. If the previous way still works, it will be used on the first bad day, and the first bad day is usually within a fortnight.

A sequence that works at this size

Weeks 1–2: inventory and honest audit. What are you paying for, who actually uses it, what overlaps. Consolidation alone frequently funds the rest of the programme, which is a useful thing to be able to tell a board.

Week 2: pick three workflows against real criteria. High frequency, well-defined, already measured, and with an owner who wants it. Not the most impressive three — the three most likely to survive.

Week 2–3: capture baselines and write the governance position. Two weeks of before-data. One page on data handling, human approval and logging. One page is genuinely sufficient at this size, and having it removes the objection that would otherwise stop you at scale.

Weeks 3–5: build and pilot with real users on real work. Not a sandbox. Sandbox pilots teach you almost nothing about adoption, which is the variable that actually decides the outcome.

Week 6: measure, close the old path, or kill it. Against the criterion written in week two. Killing one of the three is a good outcome, not a failure — it means the criteria were real.

What to do about it

Four things worth doing this week

List every AI tool you pay for and its actual usage. The gap usually funds the programme.
Choose three workflows that already have a measurable baseline. Measurability first, impressiveness later.
Write the one-page governance position now. Before a security question stops you at scale rather than at pilot.
Set the old-path closure date on day one. Adoption follows unavailability of the alternative more reliably than it follows quality.

Questions

On this topic

How much should a smaller company spend on AI?

Start by reallocating rather than adding. Most organisations of this size are already carrying redundant AI and SaaS subscriptions, and the consolidation frequently covers the cost of doing the work properly. Spending more before you can measure anything is how programmes end up unable to justify their own renewal.

Do we need a dedicated AI person?

At this size, usually not. You need a named owner for each workflow whose performance metric moves with it, and one person accountable for the governance position. Both can be existing roles. A dedicated hire before there are workflows in production tends to produce evaluation activity rather than deployed change.

What is the single most common mistake?

Leaving the old path open. It is unglamorous, it is not a technology problem, and it accounts for more dead pilots than model quality, integration difficulty and budget combined. People revert under pressure, and the first pressured week arrives quickly.

Reading is cheaper than a consultant. Up to a point.

When it stops resolving the problem, thirty minutes of structured framing usually will. Free, and useful either way.

Book a Call More insights

Half of Small Businesses Now Use AI. Almost None Use It Strategically.

The gap between "using AI" and using AI

The performance correlation the surveys keep finding

What this means if you run a $1M–$25M company

The structural advantage smaller companies have

Where smaller companies reliably go wrong

A sequence that works at this size

Four things worth doing this week

On this topic

Related

AI Enablement & Deployment

Most AI pilots die between demo and workflow

Agentic AI where it pays for itself

Reading is cheaper than a consultant. Up to a point.