Where AI pays off first, by business function

Among Canadian businesses using AI, Statistics Canada found the most common applications were data analytics at 36.6 percent, text analytics at 34.5 percent, and virtual agents or chatbots at 28.2 percent, in its Q2 2026 release published 11 June 2026. Those three are all versions of the same underlying capability: taking a large amount of text or numbers and turning it into something shorter a person can act on.

Start with an internal function rather than a customer-facing one, because a wrong output stays private while you learn how often it happens. Finance and operations tasks like categorising invoices, summarising supplier documents, or preparing recurring internal reports are common starting points. They repeat, they have clear right answers, and nobody outside the company sees the mistakes.

It can handle the repetitive top of the queue: hours, order status, and common how-do-I questions with settled answers. It handles complaints, exceptions, and anything requiring an apology badly, because those need judgement and accountability. The workable pattern is answering routine questions automatically while routing anything unusual to a person immediately, rather than making an unhappy customer argue with software.

It is useful for the preparation around selling rather than for selling itself: researching a company before a call, drafting a follow-up from your notes, summarising a long email thread, and keeping the CRM current. The conversation, the pricing judgement, and the relationship stay human. Automating outreach volume without improving relevance usually produces more messages and fewer replies.

Finance is one of the strongest starting points because the work repeats, the format is consistent, and correctness is verifiable against a source document. Categorising transactions, extracting fields from invoices, and flagging unusual entries for review are practical uses. Keep a person approving anything that moves money. The value is in preparing the work, not in authorising it.

It works well for first drafts, outlines, and adapting one piece into several formats, and poorly for anything requiring a real opinion or genuine expertise. Published content that reads as though nobody was behind it damages trust and rarely performs. The AI Marketing Playbook on this site covers this in depth, including where drafts should stop and a person should take over.

Anything where being confidently wrong is expensive and hard to detect: final pricing decisions, legal or medical judgements, hiring and firing, and anything you would have to defend to a regulator. Also avoid tasks that happen rarely, since there is no time saving to recover, and tasks so varied that describing the rules takes longer than doing the work.

For one week ask the team to note anything they did more than three times, with rough minutes. The list is usually short and surprising, and the biggest item is rarely what leadership expected. This costs a week of light record-keeping and it consistently produces better candidates than a workshop, because it measures the work rather than people's memory of it.

It is the most requested and one of the riskier first projects. A chatbot is customer-facing, so errors are public, and it needs accurate content about your business to answer from, which most sites do not have in usable form. Statistics Canada found 28.2 percent of AI-using businesses run virtual agents, so it is common, but consider whether an internal task would teach you more at lower risk.

Use it for drafting job descriptions, summarising policy documents, and answering routine staff questions about internal procedures. Do not use it to screen or rank candidates. Screening decisions affect people's livelihoods, the reasoning is difficult to audit, and inference about an identifiable individual counts as a collection of personal information under the Privacy Commissioner's generative AI guidance.

One. The failure pattern MIT documented, where 95 percent of generative AI pilots showed no measurable profit impact, is strongly associated with breadth: several half-finished pilots and nothing embedded in daily work. One task carried all the way into production teaches you more and produces a result you can point at when asking for the next budget.

Yes, in the office half of those businesses: quoting, scheduling, supplier document handling, and job report writing. Statistics Canada found urban businesses adopted AI at 21.0 percent against 9.9 percent for rural businesses, and sector variation is wide. The paperwork surrounding physical work is usually where the recoverable hours sit, not the physical work itself.