What AI Agents Actually Do for a Business (and What They Don't)

An AI agent is software that reads incoming work, decides what to do with it, and does it in your systems, with a person checking anything that matters. That sentence covers most of what you need to know. The rest of this guide is the detail: where agents earn their keep, where they don't, and how to start without getting burned.

Why is everyone asking about AI agents now?

Interest in Canada jumped this year. Comparing Google search trends for March to August 2026 with the six months before, searches for "ai agents for business" roughly doubled, and so did searches for "how to build an ai agent". Open-source agents you run yourself, such as Hermes Agent and OpenClaw, were the fastest-rising AI search topic of 2026 in the same data.

Part of that is real progress. Models have become much better at following multi-step instructions and at using software tools, which is exactly what an agent needs. Part of it is hype. Search interest tells you people are curious, not that agents are paying off. Whether one pays off in your business depends on the job you give it.

What is an AI agent, in plain terms?

Picture a new hire with a clear job description and logins to only the systems that job needs. Their manager checks anything important before it goes out. An AI agent is the software version of that hire.

Under the hood there are usually two parts. A language model, such as Claude, an OpenAI model or an open-weight model like Llama, does the reading and deciding. A workflow tool, such as n8n, Make.com or Zapier, runs the steps: watch an inbox, call the model, write to QuickBooks, send the approval request, record the result. The model brings judgment. The workflow brings structure, permissions and a record of what happened.

The word "agent" gets stretched. Some vendors use it for any chatbot. In this guide it means software that takes actions in your systems. Software that only talks is a chatbot.

How is an agent different from a chatbot or an automation?

• A chatbot answers questions in a window. It may do that well, but it usually can't create an order, update a CRM record or book a technician.
• A workflow automation moves data between apps on fixed rules: when a form is submitted, add a row. It's reliable and cheap, and it breaks the moment the input arrives in a shape it doesn't expect.
• An agent handles messy input the way a person would, then acts through the same tools a workflow uses.

In practice, a good agent is mostly workflow with a small amount of AI in the right places. If a step can be done with plain rules, it should be. AI is for the parts that need reading and judgment.

Which business jobs do AI agents handle well?

The jobs below share a pattern. They happen often, the input is messy, there's a clear right answer, and a person can check the result quickly. Each one is an example of the kind of agent a business could build, not a report on a specific client.

1. Turning emailed orders into system entries

A distributor's customers send purchase orders as PDFs, in the body of an email, or as spreadsheets. An agent reads each one, matches customers and part numbers to the ERP, and creates a draft sales order. A coordinator approves the batch instead of typing every line.

2. Processing supplier invoices

The agent reads each invoice, checks it against the purchase order and the receiving record, and enters it in QuickBooks as a draft bill. Anything that doesn't match, like a price change or a short shipment, goes to a person with the difference spelled out.

3. Sorting a shared inbox and drafting replies

An info@ or service@ inbox gets sorted by type and urgency, and each message is routed to the right person. For routine questions, the agent drafts a reply from your own policies. Someone reads it and clicks send.

4. Answering the phone

A voice agent built on Retell AI or Vapi answers overflow and after-hours calls. It tells callers it's an AI assistant, answers common questions, books appointments, and passes everything else to a person with a written summary.

5. Answering staff questions from your own documents

Your team asks "what's our return policy on special orders?" or "which torque spec applies to this unit?" and gets an answer drawn from your manuals and policies, with a link to the source. This can be a low-risk first agent, because it only reads and never acts.

6. Following up on quotes and missed calls

The agent checks your CRM for quotes with no reply after a set number of days and sends a short follow-up from the estimator's email, or drafts one for approval. Missed calls get a text back with a booking link.

7. Preparing quotes

From an RFQ or a site-visit note, the agent pulls out the scope, applies your price list and rules, and fills in your quote template. The estimator reviews, adjusts and sends. The agent does the typing; the estimator keeps the judgment.

8. Moving data between apps that don't talk to each other

A closed deal in HubSpot becomes a customer in QuickBooks and a project in your job tracker, spelled the same way everywhere. Plain automation handles the clean cases. The agent handles the ones where a person would normally have to look something up.

9. Writing scheduled reports

Every Monday morning, the agent pulls open orders, overdue invoices and the week's pipeline from your systems and writes a one-page summary in plain language, pointing out what changed since last week.

Which jobs do AI agents handle badly?

This list matters as much as the last one. We'd steer you away from agents for:

• Judgment calls about people: hiring, discipline, or a dispute with a long-standing customer.
• Anything where one mistake is expensive and hard to undo, such as releasing a payment or signing a contract, unless a person approves every time.
• Rare jobs. A task that comes up four times a year won't repay the build. A checklist will.
• Processes nobody agrees on. If three people do the job three different ways, an agent will faithfully automate the confusion. Settle the process first.
• Exact arithmetic across many numbers, done by the model itself. Language models can slip on maths. A well-built agent hands the calculation to a spreadsheet or your accounting system and checks the result.
• Relationship work. A model can draft the email to your biggest customer. It shouldn't decide what that email says.

What does it really cost to run an AI agent?

There are two kinds of cost: building the agent once, and running it every month. The second is the one people forget to ask about. We won't quote generic figures here, because running costs depend almost entirely on your volume, but here is what goes into each.

• The build: a one-time cost for scoping, connecting systems, writing instructions, testing and handover. It grows with the number of systems involved, how messy the inputs are, and how many exceptions your process has.
• Model usage: hosted models charge by the amount of text they read and write. A one-page email costs a small fraction of what a fifty-page contract does, and larger, more capable models cost more per request. Usage scales with volume, so ask for an estimate based on your own numbers.
• Platform and hosting: a subscription for the workflow tool, or a small server if you self-host n8n. A private model needs a cloud server with a GPU, billed whether it's busy or not, or hardware you buy.
• Voice: phone agents add per-minute charges from the voice platform, plus the phone number.
• Maintenance: connected apps change, your prices and policies change, and model providers retire older models. Someone has to notice and adjust, either on your staff or through a support plan.
• Your team's time: each approval takes moments, but they add up. A well-built agent makes approving much faster than doing the work by hand.

A useful test: before you sign, ask any builder for a written estimate of monthly running costs based on your volume. Our own build prices are on our pricing page.

What are the risks, and which controls manage them?

None of these controls is exotic. They're the same ones you'd put around a new employee with access to your systems, written down and enforced by software.

Should you self-host an AI agent or use a hosted model?

Hosted models, such as Claude and OpenAI's models, are the most capable available and the quickest to start with. You pay per use and the provider runs everything. The trade-off is that your data is processed on their servers, which may be outside Canada, under their terms.

Self-hosting means running an open-weight model such as Llama, Qwen or Mistral, with a tool like Ollama, on a Canadian cloud region or your own hardware. Your data stays where you put it. The trade-offs are real: a GPU server is a meaningful cost whether you rent it or buy it, someone has to patch and back it up, and models that fit on affordable hardware are weaker than the best hosted ones at long, multi-step work. For focused jobs like sorting documents or pulling fields from forms, they're often good enough.

The rise of open-source agents like OpenClaw and Hermes Agent adds a twist. They're easy to install on a laptop and can run commands, read files and send messages on your behalf. That's useful, and it's also the kind of access you'd never give a new hire on day one. If staff are already experimenting with them, bring it into the open: give the agent its own account, an allow-list of tools, approval steps and a log, or switch it off until you can.

A hybrid often makes sense: a private model handles the sensitive documents and a hosted model handles low-risk drafting. If your data isn't sensitive and no contract requires Canadian hosting, start hosted. You can move to a private model later, as long as the agent is built so the model can be swapped. There's more detail on our private agents page.

How do you pick your first agent?

Pick a job that is boring, frequent and easy to check. Then work through these steps:

1. List a normal week's repetitive tasks, with a rough count of how often each one happens.
2. Cross off anything rare, anything that needs senior judgment, and anything your team does differently every time.
3. From what's left, pick the one where a mistake is cheapest to catch. Draft orders beat payments. Internal summaries beat customer emails.
4. Write down what "done" looks like: the input, the output, and how a person checks it.
5. Name one owner on your team, and one approver.
6. Start with approvals on everything, and loosen them only when the log shows the agent gets that action right.

A good first agent is a little dull. That's deliberate. You learn how agents behave in your business on work where the stakes are low, then move on to the jobs that matter more. Our operations agents and customer agents pages show what typical first agents look like.

A short checklist before you start

• The job happens at least daily, or weekly at high volume.
• You can pull together a few dozen real examples of the input.
• There's a clear right answer that a person can check quickly.
• You know which systems the agent needs, and you can give it its own account in each.
• You've decided which actions need a person's approval.
• You know where the data will be processed, and whether it has to stay in Canada.
• You have a written estimate of the monthly running cost.
• Someone owns the agent after launch, on your team or through a support plan.
• You can switch the agent off, and your team still knows how to do the job by hand.

If you can tick most of these, you have a good candidate. Our free 30-minute agent audit goes through the same list with you, for your business.