What AI can and cannot do for a small business
AI is remarkably good at reading, sorting, summarising, drafting and extracting: the tasks that eat your team’s hours without using their judgement. It is unreliable at deciding alone, guaranteeing facts and handling situations it has never seen. That single distinction settles most buying decisions: automate the reading and the routing, keep the judgement human. The businesses that get value from AI are not the ones with the most tools, they are the ones that placed this line correctly. Our three levers describe how we combine diagnosis, automation and knowledge management around it.
LLMs in plain terms: a very well-read intern
A large language model is a system trained on enormous amounts of text that predicts, word by word, the most plausible continuation of what you give it. Treat it as a very well-read, very fast intern: excellent first drafts, useful summaries, surprisingly good analysis, and zero personal knowledge of your company. Everything it knows about your business is what you hand it in the conversation. That is why raw ChatGPT feels generic: it has read the whole web but none of your files. The fix is not a better model, it is connecting the model to your own knowledge.
RAG: plugging AI into your own documents
Retrieval-Augmented Generation is the technique that turns a generic model into your assistant: when someone asks a question, the system first retrieves the relevant passages from your documents, then lets the model answer using those passages, citing its sources. The quality ceiling is your documentation, not the AI. A company whose know-how sits in scattered emails gets scattered answers; a company with a structured knowledge base gets precise, citable ones. This is why knowledge management is one of our three levers, and why we train teams on it before deploying assistants.
Agents and automations: when AI acts instead of answering
A chatbot answers; an agent does. Reading incoming email and routing it, extracting the fields of an invoice into your accounting tool, monitoring a source and alerting the right person: these are AI decisions embedded inside a workflow that runs without anyone asking. This is where the measurable time savings live, because the work disappears entirely instead of merely going faster. It deserves its own guide: process automation for small businesses.
Hallucinations: why a human stays in the loop
A language model never says “I don’t know” on its own: when it lacks the answer it produces the most plausible-sounding one, with full confidence. That behaviour is structural, not a bug that next year’s model will fix. The consequences are manageable once you design for them: ground answers in your own documents with citations, keep a human approval step wherever a mistake is expensive, and let the machine act alone only where errors are cheap and reversible. Every serious deployment we ship keeps a human in the loop on sensitive decisions.
Confidentiality: what actually happens to your data
The honest answer has three tiers. Consumer tools may use your conversations to train future models: never paste client data into them. Business plans and API access from the same vendors contractually exclude training, which is acceptable for most SMEs. Self-hosted or French and European hosting keeps data entirely under your control, which matters for regulated sectors and sensitive know-how. The right tier is a business decision about your data, not a technical detail, and it should be settled during diagnosis, before any tool is chosen.
Choosing a model: the question that matters less than you think
GPT, Claude, Gemini, Mistral: the honest news is that for typical small-business use cases, every leading model is now good enough, and they leapfrog each other every few months. What separates a project that works from one that disappoints is everything around the model: the quality of your documents, the design of the workflow, the guardrails, the adoption by your team. Choose for data-hosting terms and cost, build so the model can be swapped later, and spend your attention on the surrounding system. That system is what an integration strategy is about.
Where to start, concretely
Not with a tool. Start by listing where your team’s hours actually go, and mark what is repetitive, rule-based and high-volume: that shortlist is worth more than any subscription. Then put honest numbers on the two or three strongest candidates and pilot the best one on a narrow scope. This is exactly what a diagnostic produces in one to three weeks, and what our pricing and ROI guide helps you budget. The worst starting point is buying licences first and looking for a problem afterwards.