On this page
AI projects fail in small businesses for the same reason automation projects always have: they start with the technology instead of the task. A chatbot on the homepage is not a strategy. Removing three hours of daily retyping is.
The shape of a good automation candidate
- It happens often — daily or many times a day.
- The rules are stable and the input is fairly consistent.
- A wrong output is annoying, not catastrophic.
- There is an obvious moment when a human should step in.
Anything involving a decision with financial, legal or clinical consequence needs a person in the loop. That does not rule out AI — it changes the design so that the model drafts and a human approves.
Three places it usually pays off
First, answering repeated questions. If a support inbox receives the same twenty questions, an assistant grounded in your own documentation can answer most of them and escalate the rest with context attached. The value is not the model; it is that your documentation finally gets used.
Second, documents. Invoices, purchase orders, delivery notes and forms are still retyped by hand in most small operations. Extraction into a structured record — with low-confidence fields flagged for review — removes a genuinely unpleasant task and reduces transcription errors.
Third, the handover between tools. A deal closes in the CRM, and then someone creates the invoice, files the documents and messages the delivery team. That chain is automation, and it does not need a language model at all — just a trigger, a few API calls and an alert if a step fails.
What to insist on before it goes live
- Logging: every automated run recorded, with the input and output.
- A confidence threshold, and a fallback when it is not met.
- A visible human handoff with the conversation or document attached.
- Cost controls: a spend limit and an alert when volume spikes.
- An owner inside the business who reads the failures weekly.
The last one is the difference between automation that survives and automation that quietly breaks. A pipeline nobody watches will fail eventually, and the team will quietly go back to the manual process — while paying for both.
Automate the work nobody wants to do. Keep the judgement where it belongs.
