Choosing where AI needs human approval
Set approval boundaries for AI-assisted content, customer replies, forecasting, pricing and recruitment before connecting the tools to business workflows.

An AI tool can suggest text, a classification or an action. The business still needs to decide who reviews it and whether the tool may change anything.
Map the approval boundary before connecting customer records or operational systems. The consequences of a wrong answer differ between a draft headline and a purchase order.
Content and design
Give drafts and visual concepts a named reviewer. Check facts, asset rights, accessibility and brand requirements before publication.
A concept image can show a product that can’t be built safely or economically, so engineering and design checks stay part of the process.
Use the customer-content pilot guide to test review effort on a limited task.
Customer replies and personalization
Start with staff-reviewed answers drawn from reference material you’ve checked. Define requests that must go to a person, such as complaints, refunds or changes to an agreement.
Before personalization, check the purpose and authority for using customer information. The Canadian generative AI privacy principles address these responsibilities.
Don’t grant a response tool authority to commit the business merely because it can send a message.
Forecasts, stock and maintenance
Distinguish language generation from predictive modelling. A well-written explanation can still come with a bad forecast.
Compare results with historical data and a simple baseline. Include missing records, seasonal demand and the cost of a wrong suggestion.
Leave orders, inventory changes and maintenance decisions with the people who make them now. Safety-critical work needs qualified review and established procedures.
Pricing and financial changes
Set who can approve prices, discounts and payment actions. Use verified inputs, and check suggested prices against your customer contracts and consumer-protection rules.
A suggested price may rely on stale information or assumptions the business hasn’t accepted. Test with fictional records before connecting live transactions.
Require logs and a way to reverse changes for any automation you switch on. Keep the authority to change a rule with the responsible business role.
Recruitment and decisions about people
Have HR and privacy advisers assess any use involving applicant or employee information. Review necessity, legal authority, fairness and the ability to challenge a result.
Don’t treat a generated ranking as an objective assessment of suitability. Keep a person accountable for each hiring decision, and give candidates a way to ask for a human review.
A safer first trial is drafting a job description from requirements the hiring manager has signed off. Leave candidate screening out of it.
Document the operating boundary
For each use, record allowed inputs, permitted output, reviewer and actions requiring separate approval. Test those boundaries, including a request the system should refuse or escalate.
Keep the decisions consistent with the staff AI policy and recheck them when integrations or features change.


