"AI agent" has become one of those phrases that means everything and nothing. Vendors use it for chatbots, for email tools, for software that books appointments. Owners hear it and reasonably ask: is this something I need, or something I am being sold?
Here is how we think about it, without the hype.
What an agent actually is
Strip away the marketing and an AI agent is software that can read something, decide what kind of thing it is, and take a step based on that. It reads an email and recognizes a purchase order. It pulls out the customer, the items and the quantities. It checks them against your system and drafts the order.
The key difference from older automation is that it can handle messy input. A rule-based tool needs the order to arrive in exactly the right format. An agent can cope with a PDF, a forwarded email and a customer who writes "same as last time but double the filters."
Where agents earn their keep
Agents are worth it when three things are true: the work is frequent, it follows a recognizable pattern, and a mistake can be caught before it does damage.
Good fits we see in small and mid-sized businesses:
- Reading incoming email and documents. Purchase orders, quote requests, invoices from suppliers, service requests. The agent sorts them, pulls out the details and files them in the right place.
- First drafts. Replies to common customer questions, quote summaries, follow-up emails. A person reviews and sends.
- Chasing. Quotes nobody answered, invoices that are overdue, certificates about to expire. The agent keeps the list and drafts the reminder.
- Searching your own information. "What did we charge this customer last time?" or "What does our policy say about this?" answered from your own documents and systems, not the open internet.
Where they do not belong
Agents are a poor fit when the work depends on judgment, relationships or context nobody has written down:
- Pricing a complicated job
- Handling an upset customer
- Anything with legal, safety or financial consequences if it is wrong
- Decisions where the right answer depends on "how we've always done it with this client"
That does not mean AI is useless there. It can prepare the information so a person decides faster. But the person decides.
Keep a person in the loop
Every agent we build has a clear point where a human approves, edits or rejects what it did. Early on, that is most things. As trust builds and the error rate is clear, the agent handles more on its own and only flags the exceptions.
This is not just caution. It is how you find out where the agent is wrong before your customers do.
The questions to ask before you start
Before you buy or build anything, answer these:
- What exact task will it do, and how often does that task happen today?
- Where does the information come from, and where does it need to end up?
- Who checks the output, and how will they know when it is wrong?
- What happens when it does not know the answer?
If you cannot answer the first question in one sentence, you are not ready for an agent yet. You are ready for a conversation about which task to pick, and that is a good place to start.
Agents work best on solid ground
One last thing that rarely gets said: an agent is only as good as the systems it connects to. If your customer records are duplicated, your price list lives in three spreadsheets and your CRM does not talk to your accounting, an agent will just make mistakes faster.
That is why our agent work often starts with integrations. Clean, connected data first. Then the agent has something reliable to work with.
Planning work like this? Part of it may qualify for Canada's SR&ED tax credits. Here is how SR&ED applies to custom development.