AI automation for small business has moved past the demo stage, but not evenly. Some uses save real hours every week for businesses with twenty staff. Others are still slideware that only works for the company that made the slide. I build both kinds of automation for a living, the AI kind and the plain kind, so this is a field guide to which is which, what a first project costs, and how to keep a person in charge of every decision that matters.
What actually works right now
Every AI automation delivering results for small businesses that I have seen or built has the same shape: the AI does the first pass, a person makes the call. Within that shape:
- Inbox triage. Reading incoming email and forms, classifying them, extracting the details, and routing them to the right person with a draft reply ready. Hours a day for any business with a shared inbox.
- Document extraction. Pulling fields out of invoices, receipts, applications and forms into your systems, with a review screen for anything the model is unsure about. Replaces re-keying almost entirely.
- Drafting. First drafts of quotes, follow-ups, summaries, reports and job descriptions from your data and your past examples. The person edits and sends.
- Answering from your own material. Internal or customer-facing assistants that answer questions from your documents, policies and product information, with citations, and hand off to a human when they cannot.
- Summarising. Meeting notes, long threads, call transcripts, support histories, turned into what you need to know and what to do next.
- Quality checks. Reviewing outgoing documents against a checklist before a person signs off.
The use-case explorer on the home page walks through eight of these interactively, including where the human sits in each.
What is still a demo
- Fully autonomous agents taking actions with money or customers and no review. The failure rate is low enough to demo and high enough to hurt.
- Replacing judgement. Pricing decisions, hiring decisions, anything with legal or safety weight. AI can inform them; it should not make them.
- "Train it on our data." Almost no small business needs a custom-trained model. Current models plus your documents plus a clear task is the whole recipe, and anyone proposing training as step one is selling.
- AI for its own sake. If you cannot name the hours it saves or the errors it prevents, it is a toy.
The most reliable AI automations I have built are boring. An email arrives, the system reads it, files it, drafts a reply and puts it in front of a person with one button. That is it. It saves the business a day a week and nobody would call it artificial intelligence at a dinner party. The impressive demos are the ones that break. The boring ones pay for themselves in a quarter.
Human in the loop, by design
Keeping a person in the loop is not a limitation you accept; it is the design that makes the system trustworthy. The pattern is simple: the AI prepares, a person approves, the system executes and logs. Approval takes seconds per item because the AI has done the reading. Over time you learn which categories the model gets right every time and can widen its autonomy there, with evidence. The automation guide applies the same principle to non-AI workflows.
What a first project costs
| Pilot | Typical build | Running cost |
|---|---|---|
| Inbox triage with drafted replies | CAD 5,000 to 15,000 | Tens of dollars a month in AI usage at small business volume |
| Invoice or form extraction with review | CAD 6,000 to 18,000 | Similar; scales with document count |
| Internal assistant over your documents | CAD 8,000 to 25,000 | Low to moderate depending on use |
| Drafting and summarising inside your existing tools | CAD 3,000 to 10,000 | Low |
These are Toronto-market orientation ranges. The right pilot pays for itself in saved hours within a few months; if the numbers do not show that before you start, it is the wrong pilot.
Your data: the questions to ask
- Which AI service, and under what terms? Business-grade services from the major providers do not train on your data by default. Consumer tools may. Know which you are using.
- What leaves your systems? Only what the task needs. Sensitive fields can often be masked or kept out of prompts entirely.
- Is it logged? Every input and output, so you can audit what the AI saw and said.
- Where is it hosted? Canadian data residency is available from the major clouds and matters for some sectors.
- Who can see the logs? Same access control as the rest of your data.
Direct tool or built workflow?
Use ChatGPT or Copilot directly for one-off tasks: a draft, a summary, a question. Have something built when the same task happens many times a day, needs your business data, has to hand results into your other systems, or needs an audit trail. That is the line between a tool a person uses and a workflow the business runs on, and it is the point at which a developer earns their fee.
Related
Practical AI systems for growing businesses are a large part of what I build, and I run my own AI marketing platform, so I know where the technology is genuinely ready and where it is not. If you have a process you suspect AI could take off your team's plate, describe it and I will tell you honestly whether it is a pilot or a demo.
Questions people ask
What AI automation actually works for small businesses?
Reading and classifying documents and emails, drafting replies and summaries for a person to approve, extracting data from invoices and forms, answering routine customer questions from your own material, and triaging requests. The pattern in every one is the same: AI does the first draft, a person does the decision.
How much does AI automation cost for a small business?
A well-scoped pilot, one workflow with a human review step, is typically a few thousand to the low tens of thousands of dollars to build, plus modest running costs for the AI service. The pilot should pay for itself in saved hours within months or it was the wrong workflow.
Is my business data safe with AI tools?
It can be, if the system is built that way. Use business-grade AI services with clear data terms, keep sensitive data out of prompts where it is not needed, and log what the AI saw and produced. A consultant who cannot explain where your data goes should not be building this.
Do I need a lot of data to use AI in my business?
No. The current generation of AI models works well with your existing documents and a clear description of the task. Training a custom model is almost never necessary for a small business, and anyone proposing it as the first step is overselling.
Should I use ChatGPT directly or have something built?
Use it directly for individual tasks. Have something built when the same task happens many times a day, needs your business data, has to hand results to your other systems, or needs an audit trail. That is the point where a tool becomes a workflow.