What your AI agent answers at 11pm, and what it leaves to a person
The short answer. Picture your site at 11pm: nobody at the desk, chat box still open. An AI agent covering that gap should act like a night clerk who knows which questions aren't hers. You find out which by tallying, for two nights, what people send while you're closed, then sorting it into three piles. Pile 1 gets answered from text you already have. Pile 2 waits for a person, with a time attached. Pile 3 the agent never talks through: a fixed message, then a person. We haven't run this tally, in our own business or a client's. It's the method we're proposing you run in yours.
An agent with no handoff rule is the night clerk who answers everything with "I'm sure the manager wouldn't mind." The manager would mind.
1. What shows up after dinner
Imagine three messages landing on your page this evening. One asks what time you open tomorrow. One wants a price for three services. One says something went wrong with what you did for them, and the writer is upset.
Same inbox, same hour. The first has an answer on your website. The second needs someone to decide. The third needs a human being, soon.
An agent that treats all three alike can answer the first, make up the second and send the third a polite reply. The third can be the costly one: that person was already annoyed when they started typing.
Sorting is the night job nobody wrote down
By day, someone on your team does this sorting without noticing. At night that person is asleep, and the agent inherits the job with no instructions.
2. Two nights, one sheet, three piles
Don't guess what people send after hours. Count it.
For two nights, log every message that arrives while you're closed, wherever it lands: web chat, texts, any inbox your customers use. Write it down in the customer's own words, with the language and the time. If you can, let one of the two nights be a weekend. Next morning, note whether someone replied with the same text as always. Then sort each entry by asking three questions.
- Would a copy-paste answer it? Pile 1. Hours, address, parking, booking, payments: the agent answers from your text as written, adds nothing and "improves" nothing, and your tool has to be able to stick to that (section 5). The rule underneath, that what's published gets answered and what belongs to a person goes to that person, is in Your next customer won't call. They'll send their assistant.
- Does it take a decision from you? Pile 2. "Can I get a discount?" "Can you fit me in today?" "What would mine cost?" No text answers these, because the answer depends on who's asking and on what you decide. The agent shouldn't touch the substance. It should say who will see it and when, and log the case.
- Could something be damaged, could someone get hurt, or is the writer furious? Pile 3. A serious complaint, a message that sounds like a lawyer is next, a scared or very angry person. The agent doesn't talk it through: it gives a fixed message and hands off to a person who really picks up, or to an emergency number where that applies. The rule is in section 4.
A rough tally still points to the first text to write
A half-kept tally only hints at which pile is heaviest. That's enough to choose which text to write first, and you can finish counting later.
3. Imagine a pet boarding shop at 11pm, sheet blank
Imagine a small pet boarding and grooming shop, an owner and a short-handed crew. At 11pm a customer wonders whether the dog can be picked up early tomorrow. That's an imaginary case: no real business and no client data behind it. The fields are blank by design. Where you see [N] or [time], put your own.
The sheet the owner would fill in after two nights:
| What they wrote (their words) | Language | Time | Pile (1, 2 or 3) | Do I have text that answers it? | Who sees it, and when? |
|---|---|---|---|---|---|
| "Can I pick up early on [day]?" | [EN or ES] | [time] | [1, 2 or 3] | [yes or no] | [name and time] |
| "How much for [service] for my [pet]?" | [EN or ES] | [time] | [1, 2 or 3] | [yes or no] | [name and time] |
| "My [pet] is [symptom] since pickup" | [EN or ES] | [time] | [1, 2 or 3] | [yes or no] | [name and time] |
When the two nights are up, count each pile: [N] in pile 1, [N] in pile 2, [N] in pile 3. Those numbers belong to you.
Why these columns are here:
- Their words. Rewrite them in yours and you lose how people really ask. Those phrases are what the agent has to recognize later.
- Language. In a bilingual business, messages can come in English, Spanish and both at once. If your pile 1 text exists in one language only, that's gap number one.
- Who sees it, and when? An empty box in pile 2 or 3 means there's no handoff, only a promise with nobody holding it. The sheet shows you that before anything gets switched on.
If pile 3 has rows and nobody picks up at night, the sheet just told you that handoff doesn't exist today. An agent won't fix that. Someone on call, a number to ring or a more honest set of hours will.
An empty pile 3 after two nights doesn't mean it can't happen. Write its rule anyway.
4. The handoff rule
Once the sheet is full, write the rule. One page is enough.
Rule: whatever the agent has no text for, it doesn't answer. It should say a person will see the message, say when, and log the case. Your tool has to be able to follow that, and section 5 asks how to check. The handoff is where the rule works or fails: it holds only if someone with a name and a time is on the other end.
Copy this and fill the blanks:
Pile 1. If the question is about [topic], the agent answers with this text, as written: [text]. If the text doesn't cover the question, it moves to pile 2, or to pile 3 if it sounds urgent or risky.
Pile 2. The agent doesn't answer the substance. It says: "A person on the team will look at this. [Name or team] will reply [day] before [time]." It logs: name, how to reach them, the question in the customer's own words, and the language they wrote in.
Pile 3. The agent doesn't continue the conversation. It says: "An assistant can't handle this. Call [number that gets answered; for a pet shop, an emergency vet clinic] now. If a person is at risk, call emergency services." It alerts [on-call person], if that on-call exists. If not, the message only says who to call and the case is logged for the morning.
In every case. The agent should say in its first message that it's an automated assistant for [your business].
Three details in that rule.
Pick a time you can keep. "We'll reply before [time]" only works if someone opens the log at that time. Unsure? Pick a later time and keep it.
Log it where you'll read it. Not in a chat nobody opens. First thing in the morning, before email.
Let it say what it is. An FTC business guidance blog post says that "certainly, people should know if they're communicating with a real person or a machine." The post is about AI that can sway consumers, not after-hours support; we use it to back our own recommendation, not as a rule. It's a sentence from a guidance blog, not a summary of what the law requires. If you're unsure what you need to disclose, check with whoever advises you.
5. Questions only you can answer
These are open questions, not findings. They depend on your business and your tool.
Can your tool be held to a text? Some can be set to answer only from what you give them, and some improvise. Ask which kind yours is, then ask it something your text doesn't cover and watch what it does.
Who decides which pile a new message goes to, the tool or a person? Without an answer, the rule has no mechanism. Find out before you switch anything on.
Who opens the log in the morning? If the answer is "me, when I can," the time you promise in pile 2 is a wish. Pick who and when, and make it a named person.
What about messages that fit no pile? Some will land on the border. When unsure between pile 1 and 2, it goes to 2. When unsure between 2 and 3, it goes to 3: an extra alert beats an urgent case with a reply for tomorrow.
This piece picks up where Your first AI agent should carry your messages, not decide left off. That one was about messages inside your team; this one is about customers. If your repeated message is one you pass to your own team, that's the other job, and it's there.
Underneath all of it is your website. Pile 1 comes from what it says, and an agent can repeat a stale promise at 11pm without blinking. We wrote about that in Your website still promises things that stopped being true, and AI repeats them. And if you want to see how customers find you through search, AI chat and voice, it's in SEO vs. GEO vs. AEO: Which One Does Your Business Need in 2026?.
This week: two nights
A note and a sheet.
- Tonight, open a note with the section 3 columns and log whatever arrives while you're closed, in the sender's words.
- After the second night, sort every row and fill in the rule from section 4. If "who sees it, and when" is empty in pile 2 or 3, that's your first job, ahead of any agent.
Frequently asked questions
To close
An agent earns the night shift by knowing what to leave alone.
You fill in the sheet, the agent doesn't: what repeats, what needs a person's call, what nobody touches. Two nights and a note turn into a one-page rule.
If you want to go over your sheet with us, book a free 30-minute session.
Keep reading
- Your website still promises things that stopped being true, and AI repeats them - what your agent repeats if your site is out of date
- Your next customer won't call. They'll send their assistant - the front desk when an AI assistant calls
- SEO vs. GEO vs. AEO: Which One Does Your Business Need in 2026? - how customers find you through search, AI chat and voice
- FTC business guidance blog post on AI and consumer trust - the source of the line about telling people they're talking to a machine
Sources
- Federal Trade Commission, business guidance blog post on AI and consumer trust (May 2023), accessed November 3, 2026. Includes the sentence "people should know if they're communicating with a real person or a machine." https://www.ftc.gov/business-guidance/blog/2023/05/luring-test-ai-engineering-consumer-trust
