Real-world AI agent use cases for Quebec SMBs

At the end of a first meeting, the question comes up almost every time: "Concretely, what would an AI agent do here?" Definitions are easy to find (we wrote our own), but a definition doesn't tell you what a custom AI agent changes on a Tuesday morning, between two meetings, in a 10-50 person business.

So here are the use cases we actually see in production in Quebec SMBs in 2026, each with the economics that justify it: what the task costs today, what the agent takes over, and what stays human. No vendor fantasy list. Specific tasks, with the numbers that let you judge whether they apply to you.

The short answer, for the busy

In a Quebec SMB of 10 to 50 employees, the AI agents actually being deployed in 2026 cover five tasks: sorting and routing incoming email, extracting supplier invoice data into the accounting system, qualifying inbound sales inquiries and answering them fast, answering internal questions from company documentation, and generating quotes from a client brief. The common profile: high repetitive volume, clear business rules, and human review on ambiguous cases. Inference cost is marginal: with a model like Claude Haiku 4.5, priced at US$1 per million input tokens, reading a full month of incoming email costs a few dollars. The real investment is design and integration: in Quebec, a custom agent on subscription starts around $947 per month and pays for itself once it frees up roughly an hour and a quarter of work per day.

Sorting and routing incoming email

The typical case: a brokerage or services firm receiving 150 to 200 emails a day. Quote requests, claims, questions from existing clients, supplier follow-ups, solicitations. Two people spend half their day reading, filing and forwarding, and an urgent claim sometimes ends up buried under three quote requests.

The agent reads each incoming email, identifies the type, extracts the key details (file number, date, amount at stake) and routes it to the right colleague. Urgent items jump the queue; the rest gets filed without anyone touching it.

The economics. On the machine side, classifying one email costs a fraction of a cent: with Claude Haiku 4.5, listed by Anthropic at US$1 per million input tokens and $5 per million output tokens (pricing verified July 2026), 4,000 emails a month come to less than $15 of inference. On the human side, if triage was eating two hours a day, the agent hands back about forty hours a month. It's the most profitable use case we know, and often the first one we ship.

Extracting supplier invoice data

The typical case: a manufacturer or distributor receiving 100 to 150 invoices a month, in PDF, sometimes scanned, never in the same format. Someone opens them one by one and keys in supplier, amounts, GST/QST and account codes. It's the bookkeeper's least interesting work, and where she makes the most mistakes, precisely because it's repetitive.

The agent watches an inbox or a shared folder, extracts the structured fields from each invoice and pushes them into the accounting system. When something is ambiguous (unknown supplier, missing tax), it flags the invoice for review instead of guessing.

The economics. At 10 to 12 minutes per invoice, keying in 150 invoices is 25 to 30 hours a month. The agent brings that down to a few hours of validation. And a data-entry error in the books costs far more to find at month-end than to avoid.

Qualifying inbound inquiries and answering fast

The typical case: a services firm getting 20 to 30 inquiries a week through its website and LinkedIn. Quality varies: real prospects, students looking for internships, salespeople. The partners sort it themselves, and good prospects sometimes wait 48 hours for a callback. By then, they've already talked to a competitor.

The agent enriches each inquiry (company size, industry, contact's title), scores it, routes it to the right partner, and sends a personalized first reply within minutes, even at 11 p.m. Unqualified requests get a polite answer pointing them elsewhere.

The economics. We won't promise you a conversion percentage: nobody serious can, without your data. The mechanism, though, is solid: in a services sale, the first vendor to respond intelligently frames the conversation, and with an agent, response speed no longer depends on who happens to be at their desk.

Answering your team's internal questions

The typical case: a 25-person team where the same questions come back every week. "What's our refund policy again?", "Where's the latest version of the standard contract?" The answers exist, in a Notion, a Drive, an old email. Nobody knows where, so everyone interrupts their neighbour.

The agent indexes the internal documentation and answers in plain language, right in Slack or Teams, citing its sources and respecting permissions: an intern doesn't see HR documents. This is exactly what separates an agent connected to your documents from generic ChatGPT (we've covered the difference).

The economics. Run the numbers for your own team: if each employee asks one question a week that burns 30 minutes in total (the person searching plus the person answering), a team of 25 loses some fifty hours a month to information hunting. The agent doesn't absorb everything, but it absorbs every question whose answer is already written down somewhere.

Generating quotes from a client brief

The typical case: a custom shop (machining, 3D printing, events) where every quote takes 30 to 60 minutes: digging prices out of the spreadsheet, calculating margin, writing it up, adding terms. So the quote goes out the next day, sometimes the next week, and clients drop off in between.

The agent takes the brief (web form, email, call transcript), consults your price list and business rules, calculates, drafts the quote in your standard template and presents it for approval. You adjust if needed and hit "send".

The economics. Forty quotes a month at 45 minutes is 30 hours; the agent brings each quote down to a few minutes of review. An underrated side effect: prices stop varying depending on who writes the quote, and at what time of day.

The economics of an agent, unvarnished

Two numbers to keep separate. First, inference cost, the one that makes headlines: as of July 2026, the official pricing pages list Claude Haiku 4.5 at US$1/$5 per million tokens (input/output), Claude Sonnet 4.6 at $3/$15 at Anthropic, and GPT-5.5 at $5/$30 at OpenAI. A million tokens is roughly 750,000 words. Having a model read the equivalent of ten novels costs between $1 and $5. That is not where your decision gets made.

The real cost is design, integration and upkeep: connecting the agent to your systems, encoding your business rules, handling edge cases, monitoring quality over time. Our custom agents start at $947 per month on subscription (we've published the full cost breakdown). At a loaded cost of $35 an hour, break-even sits around 27 hours freed per month, about an hour and a quarter per working day. Every case above clears that bar comfortably when the volume is there. And when it isn't, say 20 invoices a month instead of 150, the honest answer is that an agent isn't justified yet. Keep doing it by hand, and come back when the volume starts to grind.

Where to start

What doesn't change, first: the agent takes the repetitive part of the work, never the part that requires judgment. The broker still calls the client on the delicate file, the bookkeeper still closes the books and interprets the numbers, the partners still lead the strategic conversations. The agent recovers the hours lost to sorting and data entry, and puts them back where a brain is useful.

If one of these cases made you think "we have exactly that problem", you're probably right. The hard part is choosing which agent to build first, starting from the task that costs you the most time. We have that conversation in 30 minutes, no commitment, and if the conclusion is that it's not the right moment, we'll tell you straight.

→ First conversation, no commitment