Where AI belongs in an operational workflow, and where it does not

Most disagreements about AI in business operations dissolve once you separate two very different tasks: interpreting unstructured information, and calculating a result that has to be correct every single time. AI is genuinely good at the first. It should not be trusted with the second.

What AI handles well

• Reading unstructured input such as emails, notes, forms and documents, then extracting the structured fields underneath
• Classifying and routing incoming work so it reaches the right person without a triage step
• Drafting a first version of a recurring document for a person to review and approve
• Summarising long records into the few points a decision-maker actually needs
• Answering internal questions from documented company knowledge, with a citation back to the source

What should stay deterministic

• Pricing, costs, totals and any calculation that reaches a customer
• Anything that must produce an identical result given identical inputs
• Approval thresholds and business rules with financial or compliance consequences
• Calculations you would need to reconstruct and defend six months later

This is not a limitation to work around. It is a design principle. A system can use AI to read a messy enquiry email and populate a structured intake form, and then hand those structured values to defined rules and approved reference data for the decision itself. Both parts do what they are good at.

Illustrative example: a request-handling workflow

Consider a hypothetical process. It could be a quote, a booking, an application or a work order. The request arrives as free text. AI extracts the type, quantities, timing and other details into a structured record, and prompts the person handling it for anything missing. From that point, every consequential figure is produced by defined rules against the company’s approved reference data. A manager approves anything unusual. AI never invents a number that a customer will see.

The result is faster than the manual version and more controlled than either a pure spreadsheet or a fully generative one. The value is not that AI was involved; it is that the process became repeatable by someone other than the most experienced person in the building, and that is true whatever the industry.

Adoption is not the hard part any more

BDC’s February 2026 survey of 1,500 Canadian business owners found that only 30% of SMEs were using generative AI, but that those who did were 24% more productive than those who did not. The same research put the wider prize at up to a 38% productivity increase if all Canadian SMEs reached top-tier digital maturity.

Read that carefully, though. It is a correlation drawn from a survey, not a controlled result, and the businesses already mature enough to adopt AI well are likely to have been better run to begin with. The useful signal is not "adopt AI and get 24%". It is that maturity, meaning how well technology is embedded in day-to-day operations, that separates the businesses which get a return from the ones that do not.

A test before you deploy

Ask what happens when the model is wrong. If the answer is that a person notices during review and corrects it, the risk is acceptable. If the answer is that an incorrect figure reaches a customer or a ledger, the task needs deterministic logic, not a better prompt.