Access is no longer the problem. Goldman Sachs’ 10,000 Small Businesses Voices survey reports that 76% of the small businesses it polled use AI, yet only 14% have fully integrated it into core operations. McKinsey’s 2025 State of AI research shows the same pattern at every scale: 88% of organisations use AI in at least one function, but only around a third have begun scaling it.
The gap between those numbers is where the value sits. A tool that helps one person write an email faster is useful. A process that runs differently because AI is built into it is worth far more, and far harder to copy.
Start from the bottleneck, not the tool
Most AI projects in small firms begin with a product demonstration. The better starting point is a single, specific problem that costs the business something every week: quotes that go out late, customer requests that wait in a shared inbox, reports assembled by hand from three systems, follow-ups that never happen.
Name the problem, then name the outcome you would accept as proof it is solved: hours saved, errors avoided, response time cut, revenue kept. If you cannot state the outcome, the project is not ready. Small businesses have little room for technology that looks impressive and changes nothing.
Redesign the workflow, not just the task
Individual use of chat assistants produces individual gains. They are real but scattered, and they rarely show up in the numbers. The larger return comes from changing how a piece of work moves through the business. McKinsey found that the organisations getting the most from AI are nearly three times as likely as others to have fundamentally redesigned their workflows.
Take inbound customer requests. Using AI only to draft replies saves a few minutes per email. Redesigning the flow means the system classifies each request, judges its urgency, pulls the relevant order or account history, proposes the next action and drafts the response, while a person approves anything that commits the business. The same people now handle more requests, faster, with fewer dropped.
This is also where the choice between fixed rules and judgement matters. Some steps are best left to a deterministic bot, others need a model that can read and interpret. We cover that split in From RPA bots to AI agents.
Guardrails people can actually follow
Staff are already using AI, often through personal accounts the business has never approved. Banning it rarely works. Clear, short rules do. Every small business deploying AI should be able to answer three questions in writing:
Which tools are approved, and under which accounts.
What information must never go in: client data, personal data, pricing, anything under confidentiality.
When a person must check the output before it leaves the building or changes a record.
The same Goldman Sachs survey found that 73% of small businesses using AI would benefit from more training and resources, with data privacy, lack of expertise and tool choice among their main obstacles. In the EU this is also a legal matter: the AI Act expects organisations to take measures to support their staff’s AI literacy. Our note on AI literacy after the Omnibus sets out what a proportionate programme looks like.
Measure what changed
Take a baseline before anything is switched on: how long the task takes, how often it goes wrong, how long customers wait. Measure the same things again after a few weeks of real use. Without the baseline, every AI project looks like a success in the meeting and nobody can say what it returned.
Measurement also tells you where to stop. Some workflows gain a great deal from AI, others very little, and a few get worse. The evidence decides which ones earn further investment.
Integration beats accumulation
The firms that gain most will not be the ones with the most AI subscriptions. They will be the ones whose tools connect to the systems where their data already lives, so that work flows from one step to the next without someone copying and pasting between screens. Before adding a tool, ask where its data comes from, where its output goes and who maintains the connection.
Questions worth asking this month
Where do we lose the most time each week?
Which decisions wait too long for information?
Where are customers waiting on us?
Which repetitive work could software handle safely, with a person checking the result?
Answer those honestly and the first AI project usually picks itself.
Sources: Goldman Sachs 10,000 Small Businesses Voices survey on AI adoption; McKinsey & Company, The State of AI (2025). Figures as reported by the publishers.
