If you run a mid-sized business and you have spent any time reading about how to "do AI properly", you have been quietly misled. Not by anything false: by something true that was written about a company that is nothing like yours. That is exactly the gap AI consulting for small business is meant to close, and most of what gets published does not close it.
The published advice describes a wheel diagram of specialists: a data scientist, an ML engineer, an MLOps engineer, an evaluation lead, a data steward, an AI product manager. It tells you to redesign your workflows around AI, to stand up a platform team, to build an evaluation harness as a quarter of dedicated work. It quotes research showing what separates the companies winning at AI from the ones stalling. All of it is sound. None of it was written for a company of forty people where the person who would "own AI" is a capable engineer with a full-time job already, and the person who decides whether to spend the money is the owner, who is also doing three other jobs this week.
This post is about that mismatch: why almost all AI advice is enterprise-shaped, what genuinely changes when you are the size of an actual mid-market business, and the smaller, sharper playbook that works at that scale.
The research is enterprise-shaped, and that matters
Start with where the advice comes from. The most-cited AI research of the last two years studies large organisations, because that is who can be surveyed and who spends conspicuously. MIT's widely-quoted 2025 work analysed a population dominated by big enterprise deployments. McKinsey's State of AI surveys large companies at scale. Gartner's prescriptions are built for organisations with the headcount to act on them. These are good studies and we have drawn on them ourselves. The point is not that they are wrong. The point is that their numbers and their prescriptions assume a scale most companies do not have, and when you apply enterprise conclusions to a mid-market budget, the maths stops working.
The adoption data, read carefully, makes the gap concrete, and then complicates it in an encouraging way. By early 2026, around 72% of large enterprises had at least one AI workload in production, and deployment among the very largest firms ran far ahead of smaller ones. But the more interesting trend is the reversal underneath. The U.S. Small Business Administration's tracking found the adoption gap between small and large firms narrowing from roughly 1.8x in early 2024 to about 1.2x by mid-2025, with small businesses adopting at a faster rate while large-firm adoption plateaued. That is not the usual pattern for a new technology, where big companies lead and small ones follow years later. Something about this technology is letting smaller companies move quickly.
The OECD, in a December 2025 report on AI adoption by small and medium enterprises, named both the obstacles and the workaround with unusual precision. The obstacles for smaller firms: difficulty finding vendors with solutions actually tailored to their needs, a lack of data quality and digital readiness, and the challenge of reshaping the business around new tools. The workaround, forced by resource constraints: most reach for off-the-shelf tools rather than building. And a striking finding on the upside: 39% of SMEs using generative AI that had recently faced a skills gap said the technology helped compensate for it. The mid-market is not failing at AI. It is doing AI under a completely different set of constraints, and the advice mostly ignores those constraints.
What actually changes at mid-market scale
Four things are simply different when you are forty or two hundred people instead of five thousand. They are not nuances. They change what the right answer is.
You cannot build the operational layer: you rent it. Enterprise advice frames build-versus-buy as a strategic choice. At mid-market scale it is barely a choice; it is a constraint, and the OECD data confirms it is the path most firms take. You do not have the engineers to build and operate a gateway, an observability stack, a cost-control layer, and a retrieval system, and even if you did, spending them on that infrastructure would be indefensible. The boring parts get bought. What you build is the thin layer of your actual business on top. The difference from the enterprise is that you have far less room to get this wrong.
There is no data team to do the unglamorous work. The single most repeated finding in serious AI research is that value comes from the data layer (readiness, integration, governance), not the model. Enterprises assign that work to a data team. You do not have one. Which means the data work either gets bought as a service, or it does not happen. And if it does not happen, your AI feature fails for exactly the same reason enterprise AI features fail: it is answering against data that was never made ready, only you discover it faster and it costs you a larger share of a smaller budget. This is the hardest truth for a mid-market company to sit with: the data work is not optional just because you cannot staff it.
The team-structure advice collapses. You cannot hire the wheel diagram. The data scientist, the eval lead, the data steward, the platform engineer: at your scale these are not six people, they are one or two people wearing all the hats, or a partner carrying the operational load so your people can carry the domain knowledge. Advice that says "hire an evaluation lead" is, at forty people, advice to take a third of your only capable engineer and point them at one task. Sometimes that is right. Usually it is not, and the honest version of the advice is: get the rhythm (look at what is failing, fix it, measure it) without the org chart. We described that rhythm, and the teams that actually have it, in what working AI teams actually look like.
Measurement matters more, not less, and it is harder. The enterprise can afford a few AI projects that go nowhere; the portfolio absorbs it. You cannot. One failed AI project is a meaningful fraction of your annual discretionary spend. So measuring honestly, against a real baseline, with all costs counted, is more important for you than for the enterprise, not less. And it is harder, because you have less instrumentation and messier baselines. The discipline has to be deliberate, because nothing about your setup produces it automatically.
The advantages nobody writes about
Here is the part the enterprise-shaped advice never mentions, because the people writing it do not have these advantages and may not know they exist. The mid-market has real, structural edges in AI, and they are why the adoption data is quietly reversing.
The decision-maker is in the room. There is no AI steering committee, no three-month procurement cycle, no cross-functional alignment workstream. The owner who knows the business cold can decide on Tuesday and have something running by the end of the month. At a five-thousand-person company, that same decision is a quarter of meetings.
The scope is naturally narrow. An enterprise feels obliged to build an AI portfolio across every function. You do not. You can pick the one or two decisions that actually matter to your business and ignore the rest, and a narrow, well-chosen use case beats a sprawling portfolio almost every time. Focus is a luxury the enterprise has to manufacture; you get it for free.
You are close to the actual problem. The distance between the person who understands the business problem and the person building the solution is, in a mid-sized company, often zero: they are the same person, or they sit next to each other. That proximity is worth more than a research team. Most enterprise AI failures are failures of translation between people who understand the problem and people who understand the technology. You can shortcut that translation entirely.
And there is less to undo. You are not redesigning workflows around AI across forty departments with entrenched politics. You are changing how a handful of people work, most of whom you can talk to directly. The change-management problem that consumes enterprise AI programmes is, at your scale, a series of conversations.
The mid-market playbook, in plain terms
None of this means AI is easier for a smaller business. It means the playbook is different, and shrinking the enterprise one to a tenth of its size is not the playbook, because the enterprise playbook does not shrink linearly. The version that works:
- Pick one decision that matters. Not a portfolio. One decision your business makes regularly that better information would improve: pricing, which customers to focus on, where margin is leaking. Start there.
- Buy the boring infrastructure. The operational layer is not where you compete and not where you can afford to build. Rent it from people whose job it is to run it.
- Get the data right for that one use case, or have someone do it. This is the part you cannot skip and probably cannot staff. Either buy it as a capability or accept that the feature will not work. There is no third option where the data stays messy and the AI is fine.
- Measure against a real baseline, with all costs counted. What did this decision cost you in time and money before? What does it cost now, all-in? If you cannot answer both, you cannot tell whether it worked.
- Expand only when the first one pays. A second use case earns its place when the first has produced something you can point to. Not before.
There is a sharper version of all this in markets like ours. The OECD's first obstacle (difficulty finding vendors with solutions actually tailored to your needs) is wider outside the big technology hubs, where most tools are built for a different market, priced in a different currency, and supported in a different time zone. A mid-market business in South Africa reading advice written for a venture-funded company in San Francisco is reading about an even more distant planet than the average mid-market reader. That is not a disadvantage so much as a reason to be more sceptical of imported playbooks and more deliberate about what actually fits.
The companies in the mid-market that are pulling ahead on AI are not the ones following the enterprise playbook faithfully at a smaller scale. They are the ones who read it, recognised that it was written about somebody else, kept the few principles that are universal (get the data right, buy the boring parts, measure honestly) and threw out the rest. The advice was never written for you. The good news is that, read with that in mind, you have advantages the people who wrote it can only envy. If you want a second pair of eyes on which of those principles applies to you, that is exactly what AI consulting for small business should look like.
References
- OECD: AI Adoption by Small and Medium-Sized Enterprises (December 2025). https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/12/ai-adoption-by-small-and-medium-sized-enterprises_9c48eae6/426399c1-en.pdf
- U.S. Census Bureau Business Trends and Outlook Survey / SBA Office of Advocacy: small-business AI adoption tracking (the 1.8x to 1.2x narrowing of the small-to-large gap), reported in Capsule, Small Business AI Adoption Statistics for 2026. https://capsulecrm.com/blog/small-business-ai-adoption-statistics/
- McKinsey: The State of AI (Q1 2026; generative AI in at least one function, deployment vs. impact), reported in Medha Cloud, 67 AI Adoption Statistics for 2026. https://medhacloud.com/blog/ai-adoption-statistics-2026
- JPMorganChase Institute: Understanding the Use of AI Among Small Businesses (April 2026). https://www.jpmorganchase.com/institute/all-topics/business-growth-and-entrepreneurship/understanding-ai-use-by-small-businesses
- AdAI Research: Small Business AI Statistics 2026 (U.S. Chamber of Commerce, Thryv and SBA figures). https://adai.news/resources/statistics/small-business-ai-statistics-2026/
- ColorWhistle: Artificial Intelligence Statistics for Small Business (2026) (QuickBooks and cross-market adoption figures). https://colorwhistle.com/artificial-intelligence-statistics-for-small-business/
- WRITER: Enterprise AI Adoption in 2026: Why 79% Face Challenges Despite High Investment. https://writer.com/blog/enterprise-ai-adoption-2026/
Written by JP, Sixees Labs. Last reviewed June 2026.