Short answer: AI helps most with unstructured input — drafting, summarising, classifying, extracting information from documents and messy text. It helps least with anything requiring guaranteed accuracy, consistent identical output, or accountability for a decision. As a rule: use AI where a competent human would produce a first draft, and automation where a competent human would follow a rule.
Where AI genuinely earns its place
- Summarising. Meeting transcripts into notes and actions, long email threads into a position, research into a briefing. Reliable, immediately valuable, low risk.
- First drafts. Proposals, replies, documentation, job specs. Not final output — a starting point that removes the blank page.
- Classification and routing. Sorting inbound enquiries by type, urgency or topic. Good accuracy on well-defined categories with a human catching the uncertain cases.
- Extraction from unstructured sources. Pulling structured fields out of PDFs, emails and forms. This is often the single highest-value AI application in an operations context, because it removes genuine hours.
- Search across your own material. Answering "what did we agree with this client about scope" from your own documents.
- Translation and tone adjustment. Strong, and immediately useful for businesses operating across languages.
Where AI is the wrong tool
- Calculation and reconciliation. Use a formula. A model can produce a plausible wrong number; a spreadsheet cannot.
- Anything requiring identical output every time. Contract clauses, compliance statements, regulated disclosures. Use templates.
- Deterministic decisions. If the rule is expressible — over £5,000 needs approval — encode the rule. Don't ask a model to apply it.
- Decisions with legal or safety consequence. Hiring, credit, anything affecting someone's rights. This is also where the heavy regulatory obligations sit.
- Facts you can't verify. Models produce confident, well-formed, occasionally invented answers. If nobody checks, nobody notices.
The prerequisite everyone skips
AI applied to a business's own data is only as good as the data. Duplicated client records, inconsistent statuses and undocumented process produce fluent, authoritative, wrong analysis. The order is: clean the data, define the process, automate the deterministic parts, then apply AI to what remains. AI first is how businesses spend money on a very impressive way of being wrong.
How to evaluate an AI use case before spending
Four questions:
- What's the failure cost? If a wrong output goes unnoticed, what happens? Low cost — good candidate. Legal or financial consequence — needs a human check, and possibly shouldn't be AI at all.
- Is there a deterministic alternative? If a rule or template does it, use that. Cheaper, faster, auditable.
- Can you verify the output? If nobody can tell whether it's right, don't deploy it.
- Does volume justify it? Three documents a month doesn't need an extraction pipeline.
What about AI agents?
The current generation is genuinely capable for bounded tasks with clear success criteria and reversible actions. They're much less reliable for open-ended multi-step work involving external systems. For a small business in 2026, the practical position is: use agents for research, drafting and analysis where a human reviews the output, and be conservative about giving them write access to anything that matters.
Frequently asked questions
Should a small business build custom AI or use off-the-shelf tools?
Off-the-shelf, almost always. Custom builds make sense when you have proprietary data and a repeated, high-volume task — which is rarer than the market suggests.
How much should we budget?
Start with tool subscriptions and internal time. Meaningful gains usually come from applying existing tools well rather than from spending more.
What about compliance?
If you use AI at all, you have AI literacy obligations, and if you deploy anything customer-facing you likely have transparency obligations. Neither is onerous, but both require documentation.
AI is the last step in our process, not the first — guardrailed, specific, earned. See how we sequence it.
