If the tools are affordable and available, why do so few businesses turn them into working capability? That question has shadowed this series from the first ChatGPT experiments of 2022, through the agent-building year of 2025, to now. In 2026 the evidence is consistent across multiple studies: adoption keeps rising, but the distance between using a tool and operating with AI remains large. This closing article looks at that implementation gap — and what sits inside it.

What the data says

Start with the most recent evidence. In April 2026 the OECD published Empowering SMEs in the age of AI, based on a sample of more than 2,000 SMEs across 12 countries. It reports rising AI adoption among SMEs but uneven strategic and secure integration — many businesses are using AI somewhere, far fewer have integrated it into how the business actually runs. One caveat matters before drawing conclusions: this sample is non-representative, so the figures tell us about the direction of travel among surveyed firms, not exact prevalence across all SMEs.

Earlier, the OECD's Generative AI and the SME Workforce (2024 publication) surveyed a representative sample of more than 5,000 SMEs across seven countries. It found 31% were using generative AI — but, importantly, that among those users only 29% reported using it in core business activities. In other words, even where generative AI had arrived, it had mostly arrived at the edges.

Eurostat's 2025 figures for the EU show the same stratification by size: among enterprises with 10 or more employees, roughly 20% used AI, broken down as 17% of small enterprises, 30.4% of medium-sized ones and 55% of large ones (Eurostat, 2025; EU; enterprises with 10+ employed persons). Larger organisations adopt first — they have the integration capacity, the data infrastructure and the people to own the change.

The frontier-versus-open-source squeeze

One 2026 development reshaped procurement as much as any product launch: for most business tasks, the best open-weights models — DeepSeek, Qwen, Kimi, Llama, Mistral — now sit close enough to the proprietary frontier that the choice is no longer "capable or not". It is a set of ordinary engineering trade-offs: per-unit cost, data residency, rate limits, fine-tuning control and vendor dependence. Local inference on owned hardware, and private AI cloud arrangements where the model serves only your organisation, became routine architecture rather than exotic projects. Two consequences follow for the gap below. First, access stopped being the bottleneck — if models were the constraint, 2026 would have solved the problem. Second, the differentiation moved entirely to integration, ownership and measurement: exactly the disciplines this series has traced since 2023.

Our reading of the gap

The interpretation that follows is Cooper Agile's, not any research organisation's: the opportunity extends beyond first-time adoption. Businesses also need help turning individual AI tools into dependable operational capabilities.

Four things sit inside the gap:

  • Integration. As this series covered in the 2025 article, the models learned to act — and acting exposed permissions, tooling and escalation problems that no licence purchase solves. A tool that does not connect to your systems produces demonstrations, not output.
  • Skills. Using an assistant for drafting and running an AI-assisted workflow across a team are different activities, with different failure modes. The second requires people who can specify, evaluate and challenge the system — skills that are built, not bought.
  • Operational ownership. Tools without owners decay. When the person who built the prompt leaves, or the vendor changes the model, an unmaintained workflow quietly breaks. Capabilities need a named owner, monitoring and maintenance budget, just like any other production system.
  • Measuring value. Many 2026 programmes still cannot answer "what did this change?" with anything better than anecdote. Without a baseline and a measure, improvements cannot be defended at budget time — and unmetered systems get switched off.

Illustrative example: a 60-person logistics firm buys AI seats for its operations team in January. By October, three workflows — enquiry triage, document extraction, depot notes — run daily, but only one has an owner, none log outputs, and nobody can say whether response times improved. The firm hasn't failed to adopt AI; it has failed to implement it. The fix is not more licences. It is wiring those three workflows into monitored, owned, measured capability.

Why this persists

The gap persists because buying is fast and implementing is slow. Procurement is a transaction; capability is an operating habit. It also persists because the industry narrative rewards adoption counts over operational outcomes — a vendor can report "customers using AI" without any of those customers depending on it.

The trade-off is real either way. Closing the gap properly takes engineering time, discipline and patience — real costs, especially in smaller firms. Leaving it open is cheaper this quarter and compounds: shadow workflows accumulate, quality drifts, and the eventual cleanup is larger than the incremental work would have been.

What 2026 leaves us with

The technology cycle from 2022 to 2026 took us from novelty to assistants to agents to infrastructure. The remaining bottleneck is not the models. It is the ordinary, unglamorous work of connecting them to real processes and keeping them honest in production. That work is where the value is, and it is the work most businesses have not yet done.

Action checklist

  • List every AI tool in use today, and mark which touch core business processes versus peripheral tasks.
  • Assign a named owner to each operational workflow, with monitoring and a maintenance budget.
  • Baseline the metric each workflow is meant to improve before tuning anything.
  • Audit tool permissions and data access, especially where agents can write to systems of record.
  • Invest in evaluation skills: teach teams to test workflows against fixed task sets, not impressions.
  • Prefer one dependable, measured capability over five demonstrations.
  • Review vendor dependence: know what breaks when a model or feature changes.

Closing the implementation gap is the core of how we work — focused, measurable steps through our FORT delivery cycles: FRAME → ORCHESTRATE → REALISE → TRANSFER. Start with a platform modernisation engagement, automate a CRM workflow that actually gets used, or bring in fractional engineering leadership to own the capability after we hand it over.