Earlier business notes
Where AI suggestions should stop and human decisions begin
How to define approval boundaries for AI-assisted work before a suggestion turns into an action that affects customers or a team.
By Rajiv Sankarlall · Updated
An AI tool can produce a recommendation that sounds ready to act on. Before connecting it to a workflow, decide what it is allowed to do with that recommendation. Drafting, recommending and taking action are different levels of responsibility.
This is a practical technology note. The coaching offered here focuses on founder burnout, values and purpose; it does not include automating your company or installing an AI employee.
Name the actual decision
Choose one proposed use rather than “AI for operations.” For example, consider drafting a reply to a routine support request. What information does the draft need? Who checks it? What should happen if the request involves a refund, sensitive data or an unhappy customer?
Write down the boundary in words a person can understand. “Suggest a reply, but do not send it” is clearer than “use judgment.” If the software can act, enforce the limit through its permissions and workflow. A reminder in a prompt is not a substitute for restricting access.
Make uncertainty visible
Decide what the tool should show alongside its recommendation: source material, missing information and the parts that need checking. Do not make the person reviewing it reconstruct the whole situation from a polished paragraph.
Try the process with ordinary requests and awkward ones. Include incomplete information, conflicting instructions and material from outside the company. For security considerations, OWASP's GenAI Security Project provides guidance about risks in AI applications.
Keep someone responsible for the outcome
Name the person who owns the workflow, not just the person who installed the tool. They need a way to inspect what happened, pause the process and correct mistakes. Agree how errors are reported and when the experiment should stop.
Start with a scope small enough that you can review actual results. If checking the output takes more work than doing the task, reconsider the use case. Expanding automation is a choice that needs evidence from your own workflow.
For a founder, the difficult part may be defining responsibility clearly enough to let another person or tool help. That question also applies without AI. Read a practical approach to shared decisions.