By KYN AI Advisory Team — AI implementation specialists, Singapore
An AI sales agent should escalate when a reply touches price, terms, a complaint, a legal or regulatory matter, or when it has already tried twice and the objection remains. It should also escalate when it is not confident what the reply means. Build these as measurable triggers with a retry cap, not as a model's judgement call. A wrong "resolution" costs more than a handoff.
Why is a wrong resolution worse than an escalation?
An agent that answers an objection wrongly does two kinds of damage. The prospect hears an answer that may be false, and your record shows the objection as handled when it was not. Nobody follows up because the log says it is closed.
An escalation costs a person a few minutes. A false resolution can cost the deal and the next three conversations. So the default for doubt is a handoff.
What triggers should cause escalation?
Write triggers as rules the agent checks on every reply.
| Trigger | Example | Action | |---|---|---| | Price or discount | "Can you do better on the fee?" | Stop. Hand to the account owner | | Terms | Contract length, liability, payment terms | Stop. Hand to the account owner | | Complaint | "This was not what you promised." | Stop. Hand to a named person today | | Legal or regulator | Dispute, notice, data request | Stop. Hand to a named person | | Repeat objection | Same objection after two agent replies | Stop. Hand over | | Low confidence | Classifier confidence below your threshold | Stop. Flag | | Unusual sentiment | Anger, sarcasm, distress | Stop. Hand over |
Set the confidence threshold by testing, not by feel. Take 100 real replies, have a person label them, and measure where the agent is wrong. Raise the threshold until the errors that matter most, such as a complaint labelled "not now", drop to an acceptable level.
How do severity levels work?
Not every handoff has the same urgency.
- Level 1, same day: complaints, legal or regulator mentions, anything involving money owed.
- Level 2, within one working day: price, terms and repeat objections.
- Level 3, next review batch: low-confidence replies and unclear intent.
Route Level 1 to a named person by WhatsApp. Put Levels 2 and 3 in a queue with the draft, the reply and the account history in one view.
Why does escalation need a retry cap?
An agent that keeps replying to an objection can loop. It restates the same answer, the prospect repeats the objection, and the thread grows without progress. Cap it at two attempts per objection. On the third, stop and hand over.
Add a second cap on total agent messages per thread. A long thread with no human involvement is itself a trigger.
What does a good handoff look like?
The handoff is a work package, not a failure notice. It should contain:
- the prospect's last message, quoted;
- the objection in one phrase;
- what the agent already said;
- the account stage and owner;
- a drafted reply for the person to edit or approve.
The reviewer then confirms, edits or replaces. That is minutes, not an investigation. A person who sees the draft and the source clears a batch quickly.
How do you improve the triggers over time?
Log every escalation with the actor, time, trigger and outcome. Review monthly:
- Which triggers fired most, and were they right?
- Which escalations did the person resolve with no edit? Those may be safe to automate later.
- Which replies should have escalated and did not? Add a trigger.
This turns the log into the tuning tool.
What does an escalation decision look like in a real thread?
Illustrative thread. The company and numbers are made up.
| Turn | Prospect says | Agent check | Action | |---|---|---|---| | 1 | "Interested. What does it cost?" | Price mention, first time | Stop. Hand to account owner with the question quoted | | 2 | (Owner replies with a quote.) "Is there a lower tier?" | Price, repeat | Owner handles; agent stays silent | | 3 | "We are comparing two other suppliers." | Competitor mention, no price ask | Agent drafts a factual comparison page link; owner approves | | 4 | "Not now, maybe next quarter." | Timing objection, first attempt | Agent drafts a short reply that offers a date to re-contact; person approves | | 5 | "Not now." (again, after the reply) | Same objection, second time | Stop. Hand over; do not send a third reply | | 6 | "Your last email broke our data rules." | Complaint and possible regulatory content | Level 1. Named person, same day |
Three things stand out. Price goes to a human on the first mention. A timing objection gets two tries and no more. A complaint jumps the queue whatever the thread history says.
How do tolerance thresholds and second-pass checks work?
A classifier assigns each reply a label and a confidence. Use that confidence as a gate.
- First pass. The agent labels the reply: price, terms, complaint, legal, timing, competitor, interest, unclear.
- Threshold. If confidence is below the level you set for that label, flag the reply. Use stricter thresholds for costly labels. A complaint labelled "timing" is a worse error than the reverse.
- Second pass. For any reply the agent is about to treat as resolved, run a second check from a different angle. Does the thread contain a number, a date, a clause reference or a negative word that the first label did not account for? If yes, flag.
- Corroboration. Do not loosen a threshold to get an answer. If the first pass is unsure, require a second, independent signal before acting, such as a keyword rule that agrees with the classifier.
A reply that passes both checks still goes to a draft queue if the action is outbound. The check reduces how often a person is needed. It does not remove the person.
How do you test the triggers before go-live?
Use real replies. A synthetic set hides how people write.
- Collect 200 replies from past campaigns, across clients and channels.
- Have two people label each as escalate or not, and settle disagreements.
- Run the agent in draft mode and compare.
- Count the two error types separately:
- Missed escalations: the agent would have handled a reply that needed a person. Treat each as a defect.
- Unnecessary escalations: the agent handed over something it could have handled. These cost time, not deals.
- Tune thresholds until missed escalations reach a level you accept, and write that level down.
- Re-test after each rule change.
Add hostile and odd cases: a reply with two objections, a joke that hides a complaint, a message in mixed English and Chinese or Bahasa, a forwarded email chain and an out-of-office reply. Each should have an expected action before the test begins.
What are the edge cases?
- Two objections in one message. Handle the highest-severity one first. If any trigger fires, the whole message escalates.
- Silence after a handoff. If the person has not replied within the stated window, remind them and tell the prospect. Do not let the thread sit.
- Out of hours. The agent may acknowledge and state when a person will reply, but it must not answer price or terms.
- Language switch. If the prospect changes language, confidence usually drops. Flag until the classifier has been tested in that language.
- Forwarded or quoted text. A quoted earlier message can trigger a rule falsely. Strip quotes before classification.
- Multiple contacts at one account. An escalation by one person pauses automated replies to colleagues at the same account until the owner decides.
- Unsubscribe. Treat as an immediate stop on all channels, with no persuasion attempt.
What is the economics of a human in the loop?
Illustrative arithmetic. Substitute your own figures.
Suppose an agent sorts 1,000 replies in a month. At about S$0.01 per reply sorted, as in KYN's live brokerage deployment, the model cost is about S$10. Suppose 15 percent escalate, which is 150 replies, and a person spends four minutes on each. That is 10 hours of staff time.
| Line | Amount | |---|---| | Model cost for 1,000 replies | about S$10 | | Escalations (assumed 15 percent) | 150 | | Human time at 4 minutes each | 10 hours | | Replies handled with no person | 850 |
The person's time is the main cost, not the model. That is why the useful lever is the quality of the handoff: a package with the quote, the objection, the history and a draft reply lets a person clear an escalation in minutes. A bare "please look at this" takes much longer.
If the escalation rate is 2 percent, the agent may be missing things. If it is 60 percent, the triggers are too wide or the data is poor. Investigate both extremes.
Who owns the escalation queue?
A queue without an owner becomes a backlog.
- Name a person for each level and a back-up for leave.
- State a response window for each level and display it in the queue.
- Show the age of each item and sort by severity, then age.
- Send Level 1 items to WhatsApp so that they do not wait for someone to open a dashboard.
- Review the queue in a weekly meeting for the first month.
The agent should tell the prospect who will reply and by when, in plain words. A statement such as "I have passed this to the account owner, who will reply today" sets a clear expectation.
What should the log capture?
For every escalation, record:
- The prospect's message, quoted.
- The trigger that fired and the classifier's label and confidence.
- The agent's previous replies in the thread.
- The person who received it, the time and the time they replied.
- What the person did: approved the draft, edited it or replaced it.
- The outcome at 30 days, such as meeting booked, deal lost or no change.
Use the log to improve the rules. Triggers that always end in an unchanged draft may be candidates to automate. Replies that were handled automatically and later led to a complaint show where to add a trigger.
Start with strict triggers and loosen them only on evidence from the log. It is easier to hand over too often in the first month and relax later than to recover a deal lost to a confident wrong reply.
Where KYN fits
KYN builds sales agents with rules-based logic in the client's own cloud account. Agents have their own restricted accounts, every action is logged with the actor, and outbound messages, money and deletions wait for a named person. In a live insurance brokerage system, a reply desk sorts each reply at about S$0.01, and replies that fall outside the rules go to a person.
The first 14 days are a review of your stack and team. A full build ships in under 90 days. See sales systems and security. For limits on the agent's authority, read what an AI sales agent can and cannot do.
Frequently asked questions
How many objections should an agent handle before escalating?
Two attempts on the same objection. After that, a person takes over. For price, terms and complaints, escalate on the first mention.
Should the agent tell the prospect it is handing over?
Yes, in plain words, and with a time. "I have passed this to the account owner, who will reply today." Do not leave the prospect waiting in silence.
How do I measure whether escalation rules are working?
Track missed escalations, unnecessary escalations and time to human reply. Missed ones matter most: review a sample of non-escalated replies each week.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
