← Back to the blogBefore you automate a task with AI, write down the exceptionsEdgeIt·Nov 10, 2026 · 3 min readAn inquiry lands in your inbox: “Can you rebuild our website before Friday? We need a price today, but we haven’t decided which pages to keep.” AI could help summarize that message and prepare a reply. It shouldn’t quietly turn the requested deadline into a promise your team never made. If you’re a business owner choosing your first AI workflow, start by writing down where the task needs someone’s judgment. That gives you a useful brief before you choose a tool or connect your inbox. Give the task a finish line “Handle our leads” leaves too much open to interpretation. Try a narrower job: “Read an approved inquiry, summarize the requested service, identify missing information and prepare a reply for review.” Decide what the output should contain. For this example, it’s a short summary, a link back to the original message, a proposed next question and the name of the person who will review it. The workflow finishes when that draft reaches the reviewer. Sending it is a separate step. This matters because generative AI can produce plausible statements that are wrong. NIST describes that behavior as confabulation in its Generative AI Profile. A polished reply still needs checking against the actual inquiry and your business’s current information. Work through one awkward inquiry Here’s a fictional example for a small service business. It illustrates a possible workflow; it isn’t a client result or a claim about a particular AI product. The incoming request is the website inquiry above. A useful summary would say: website rebuild requested; Friday is the customer’s preferred deadline; page scope is undecided; price requested today. The distinction between “requested” and “confirmed” is the whole point. The output should preserve what the customer said without treating it as an agreement. A reasonable draft might be: “Thanks for getting in touch. Could you send your current website link and the pages you expect to need? We’ll review the scope and our availability before confirming a quote or delivery date.” The reviewer then checks the message, decides whether the business can help and edits or sends the response. During an initial pilot, keep every reply in that review queue, including the easy ones. Name the exceptions before connecting the tools Write a short rule for missing information: leave the field marked “not provided” and suggest a question. Don’t fill gaps with a likely budget, deadline or service package. Write another for commitments: pricing, discounts, availability and delivery promises go to a named person. Specify what that person needs to see. “Human review” isn’t very useful if the reviewer receives only a draft with none of the original context. Conflicting messages need their own route. If someone asks for a full rebuild and later says “just update the homepage,” show both requests and ask the reviewer to resolve the scope. Finally, plan for failure. If the inquiry can’t be read or the draft can’t be delivered, keep the original message available, flag the problem and tell the responsible person. A silent failure shouldn’t make a lead disappear from the normal inbox. These are suggested rules for this example. NIST’s AI Risk Management Framework supports defining and documenting human oversight; it doesn’t prescribe this particular inquiry process. Test the handoff, including the inconvenient cases Use made-up inquiries first: a straightforward request, one with missing details, one with conflicting instructions and one asking for an urgent commitment. Decide the expected outcome before each test. Check the summary and the route to the reviewer, not just whether the reply sounds good. Include a simulated delivery failure. Can the responsible person still find the inquiry? Is it clear that no reply has been sent? A successful example alone won’t answer those questions. Before using real customer information, confirm what data the chosen tool receives and how your account handles it. Share only what the task needs. Keep the original workflow available while you assess the pilot, and track the corrections your reviewer makes. Repeated corrections tell you which instructions or steps need work. Bring one repetitive task and its difficult cases EdgeiT’s AI implementation service, led by LJ, covers workflow reviews, tool setup, approval controls and team training. A useful starting point is one task your team repeats, along with a few examples of when it gets complicated. Tell us where the task gets stuck, and we can discuss a practical place to start.♡LikeCommentsSort Newest firstOldest firstPost
Las Vegas based.