All guides
What an AI Agent Should Do When Nobody Answers
When an AI agent asks a person and gets no reply, it needs a fallback ladder: a time limit, the next person, then a safe default. How to design each rung, and what never to do.
By Fruxon Team
November 7, 2026
7 min read
The short answer
When an AI agent asks a person and nobody answers, it should move down a fallback ladder: give the first person a set amount of time, then ask the next person who can decide, then fall back to a safe default that you chose in advance. While that happens, it keeps whoever is waiting informed.
What it should never do is treat silence as a yes. An agent that goes ahead with an irreversible action because nobody objected has turned one unanswered message into a mistake you'll have to clean up.
Why this happens more than you'd expect
Making an agent ask a person is the first half of the job. The second half is everything that happens when that person doesn't reply.
They will often not reply. They're in a meeting, on a flight, off sick, in another time zone or just buried in messages. Today, when an email asking for a decision goes unanswered, the case sits in someone's inbox until a person notices and chases it. Sometimes that's an hour. Sometimes it's a week, and the customer chases first.
An agent can do better than that, but only if you decide ahead of time what "better" means. If you don't design the fall-through, you get one of two failures: the work waits forever, or the agent decides on its own, without any guidance from you.
The fallback ladder
A fallback ladder is an ordered list of what to try, each step with a time limit. A typical one has three or four rungs:
| Rung | What happens | Example |
|---|---|---|
| 1. The owner | Ask the person who owns the decision. Wait a set time. | Dana, who approves large refunds, gets 15 minutes. |
| 2. The next person | Ask the next person who can make the same call. | Dana's manager, or the other person who covers refunds. |
| 3. A safe default | Take the action you agreed in advance is acceptable if nobody decides. | Hold the refund and tell the customer they'll hear back by tomorrow. |
| 4. Park and flag | Leave the case open, clearly marked, for a person to pick up. | The case appears in the team's morning review with the full history. |
Not every question needs every rung. A question about a $50 credit might go straight from the owner to the default. A question about shutting off a customer's account might skip the default entirely and park until a person decides.
How long to wait at each rung
Set the time limit by what the delay costs, not by a standard number.
- Someone is waiting live, such as a customer in a chat: minutes. Every minute is visible to them.
- Someone is waiting, but not live, such as a customer email or an internal request: an hour or two is usually fine.
- Nobody is waiting, such as an invoice exception or a record that failed to sync: the end of the business day, or longer.
Keep the first rung short enough that a busy owner doesn't hold everything up, and long enough that they have a real chance to answer. Then adjust it based on what you see. If owners usually answer in 20 minutes and your limit is 15, most questions are moving on just before the answer would have come.
Watch out for working hours. A limit that makes sense at 11am is wrong at 11pm for someone in another time zone. Either pause the clock outside working hours or start with a person who is working.
What makes a default safe
A safe default is the action you'd be comfortable with if nobody ever answered. Good defaults tend to look like this:
- Hold, don't act. Put the record on hold, flag the order, keep the money where it is.
- Take the reversible option. Issue a store credit that can be converted later rather than a refund that can't be taken back.
- Answer the part you can. Tell the customer what you can confirm now and when they'll hear about the rest.
- Decline politely, and say who will follow up. For requests the rules clearly don't allow.
And it should never look like this:
- Silence means yes. Going ahead with a payment, a deletion or an external message because nobody said no.
- Silence means no, said bluntly. Rejecting a customer's request outright because a colleague was at lunch.
- The agent decides on its own. If the agent could have made the call safely, it shouldn't have asked in the first place.
Agree the default with the person who owns the decision. It's their call, and they're far more likely to trust the agent once they've chosen what happens when they're unavailable.
When the answer arrives late
Sometimes the owner replies after the question has already fallen through. Decide in advance what happens then:
- If the default was a hold, use the late answer. That's the point of holding.
- If the default already did something reversible, use the late answer to correct it if needed, such as turning a store credit into the refund the owner approved.
- If something irreversible already happened, don't redo it. Record the late answer, flag the difference and let a person decide.
In every case, keep the late answer with the case. It tells you whether your time limits are too short.
Keep the person who asked informed
The customer or colleague who started all this doesn't care about your ladder. They care whether anything is happening. When the question moves to the next person or to a default, tell them something useful: that it's being looked at, and when they'll hear back. "We're checking this with our team and will reply by 3pm" is far better than silence followed by a surprise.
What to measure
Two numbers tell you whether the ladder is working:
- How often each rung is reached. If most questions reach the default, the problem is earlier: the wrong people are being asked, or the time limits are too short.
- How long owners take to answer. This tells you where to set the first time limit, and which people are overloaded.
If one person is the first rung for a lot of questions and often misses them, that's a staffing signal, not an agent problem. Add a second person to that area.
How Fruxon does it
In Fruxon, every question an agent asks has a fallback ladder: the next person, then a safe default. A typical setup gives the first person fifteen minutes before the question moves on, so nothing depends on one person's lunch break or time zone. The run stays open the whole time, so a late answer still finds the work where it was left, and every step on the ladder (who was asked, how long they took, what happened next) shows up in the run's trace.
To set up the asking itself, start with how to make an AI agent ask a human. For the broader pattern, see human-in-the-loop AI.
All guides
