Say a customer writes at 9:40 on a Tuesday night: "Can you do coffee and pastries for 25 people on Thursday? One guest has a nut allergy."
You could answer it now, tired, from memory. You could leave it for the morning and hope. Or you could paste it into an AI assistant and get back something warm, fluent, and quietly wrong.
That last one is the part nobody puts on the brochure. AI is very good at sounding sure. It is not good at knowing your prices, your schedule, or what your kitchen can safely promise. Leave a gap and it will fill it, cheerfully.
The fix is better facts.
The problem usually isn't the writing
Most owners I work with can write a decent reply. The trouble is that the reply depends on things that live in somebody's head: what we charge, whether Thursday is open, what the hold policy actually is this month.
When those facts aren't written down, every reply starts from zero. Different people give different answers. And an AI draft just speeds up the guessing.
The routine I teach is simple and a little boring on purpose:
- Approved facts first. What's true, what isn't decided, and what needs a person to check.
- An AI draft second. Built only from those facts.
- A human review before sending. Every time.
Boring is good here. Boring means your customer gets the same honest answer on a Tuesday night as on a Saturday morning.
Start with an approved facts sheet
An approved facts sheet is not a manual. It's a short list, the kind you could tape inside a cupboard door. It holds the things you can say plainly and, just as important, the things you can't promise without checking.
That second half is where most bad replies come from. "Yes, we can do that!" feels friendly right up until it isn't.
A café: the catering question
Back to that Tuesday night message. Here's what a café's approved facts might look like (illustrative, not anyone's real menu):
- Open 7 a.m. to 3 p.m., Tuesday to Sunday.
- Coffee service and pastry trays are available for groups.
- Pickup or local delivery; delivery has a fee.
- A manager confirms pricing, capacity and dietary requirements for every event order.
- We don't promise a nut-free kitchen.
Before (fast, friendly, risky):
"Absolutely! We'd love to cater your event. Our pastries are all nut-free, and 25 people is no problem. See you Thursday!"
Lovely tone. Three promises nobody approved, and one of them is about an allergy.
After (drafted from the facts, then reviewed):
"Thanks for thinking of us. What time is the event, and would you prefer pickup or delivery? I'll pass the allergy note to our manager so they can confirm what we can safely offer, along with pricing for 25."
Less exciting. Much better. It's warm, it asks for what's needed, and a real person owns the next step.
A retail shop: "Can you hold it?"
Say someone messages a gift shop: "Is the blue lamp in stock, and can you hold it until Saturday?"
The honest answer depends on two things AI can't see: what's on the shelf right now, and your hold policy. If neither is in your facts, the draft shouldn't pretend.
Before:
"Yes, we have it! I'll put it aside for you until Saturday."
After:
"Thanks for asking about the blue lamp. Let me check current stock and our hold terms before I confirm. Could you send the product name or a photo so we can check the right one?"
If you do have a hold policy, write it into your facts ("holds last up to three days," or whatever is true for you). Then the draft can say it, and say it the same way every time, no matter who's on the till.
A service business: "What would it cost?"
A plumbing company gets: "Can someone come look at a leaking water heater Thursday? What would it cost?"
The approved facts are short. We repair and replace water heaters. The service call fee and available slots need a team check.
Before:
"Sure! Thursday works, and a service call usually runs about $150."
Where did $150 come from? Nowhere. Where did Thursday come from? Also nowhere. That's the problem.
After:
"Thanks for letting us know about the water heater. We do repair and replace water heaters. Is the water leaking from the tank itself or from a pipe or valve near it? Once we have that, our team can check Thursday's schedule and confirm the service call fee."
No invented price. No slot promised. One useful question. The quote comes later, from a person who knows the job. (And once that quote goes out, the follow-up is its own small art. I wrote about it in "Just Checking In" Is Killing Your Quotes.)
The human check before you hit send
The review step is the one people want to skip, because the draft looks finished. Look anyway. Three questions cover most of it:
- Did it invent anything? A price, a time, a discount, a guarantee.
- Did it imply something is available that nobody has checked?
- Does it ask one useful question, or name who handles the next step?
If the first two answers are no and the third is yes, you're in good shape. If not, fix the facts sheet, not just the reply. Next time the draft starts from a better place.
And name the person who does the check. "Someone will look at it" is how nobody looks at it.
Try it on one real message this week
You don't need a new system. Pick one question you answer over and over. Write five or six approved facts for it, including what you can't promise. Draft from those facts with whatever AI assistant you're allowed to use for that information. Check it. Then send it like you wrote it, because you did the part that matters.
If you'd like to try it on a worked example first, there's a short free lesson from my FieldKit AI Bootcamp that walks through exactly this routine: fictional practice input, a prompt to copy, and a reviewed reply to check yours against.
Business consulting. Practical progress.
Related: "Just Checking In" Is Killing Your Quotes. FieldKit Field Notes for owner-led businesses.