Practical AI for small-business operators: putting AI inside the software you already use, choosing tools worth paying for, and automating the right things first.
The template is the foundation Copilot learns from, not the prompt. Two documented limits that break a real workflow, and the checklist that works around both before a deck goes in front of anyone.
A better prompt does not fix content Copilot under-weighted because it sat in the middle of a long file. Three documented causes for one symptom, checked in the order that actually rules them out.
A scanned page with no text layer is read as nothing, not as an error. Here are the four documented reasons Acrobat’s AI Assistant refuses a file, and what actually fixes each one.
The setting has to be on before the edit, not after. Compare is the only tool that catches an AI edit made before Track Changes was switched on, because it reads two saved files rather than an edit history that may never have started.
A comment can be deleted if Copilot rewrites the paragraph it’s anchored to. This is documented Microsoft behaviour, not an edge case, and it’s one of six stated exclusions that decide whether you can rely on the feature at all.
Add personality is not an instruction anyone can follow. Em dashes, groups of three and uniform sentence length are. Here is the edit pass, in the order that makes each step easier than the last, and why detectors are the wrong target.
Business Starter gets nothing in Sheets and still gets Help me write in Gmail. The usage figure everyone quotes appears nowhere in Google’s documentation, and the limits that do exist are per user rather than pooled.
A rule runs while your laptop is shut and costs nothing per message. Copilot needs you awake and does not run in shared mailboxes at all. Here is where the line falls, and why the rules limit is not the number everyone quotes.
A fluent draft promises things you never asked it to promise. Here is why the button disappears on shared mailboxes and plain text, and the fifteen-second scan that catches what a generated message commits you to.
Chart type is determined by what the data is, not by taste, which makes it one of the more reliable things to ask AI. Asking whether the result looks good is not, and that is where captions get quoted back at you.