By KYN AI Advisory Team — AI implementation specialists, Singapore
Outbound with an AI agent breaks in four places: the list it works from, the mailbox it sends through, the way replies are read, and the handoff to a person. Volume is rarely the cause. A system that fails at 50 emails a day fails at 500, only faster. Fix the four points in order, because each one caps the next.
Where does outbound break first?
At the list. An agent writes confidently from whatever it is given. If the CRM holds a contact who left the company 14 months ago, the agent writes a polished first line to nobody.
Three list faults do most of the damage:
- Stale contacts. The person moved, the company closed or the domain changed.
- Duplicates. The same contact sits under two company names, so two reps (or two sequences) write to one person.
- Missing source. The record says "decision maker" but not where that came from, so no one can check it.
Fix. Put a gate before any draft is written:
- Check the email domain resolves and the address passes a verification step.
- Match the contact against the CRM on email and on normalised company name. A match blocks a new sequence.
- Require one checkable source per account: a registry record, a company page or a news item with a URL. No source, no draft.
The third rule is the one that matters. A first line the reader can verify ("Your site lists a new Surabaya warehouse") reads as research. A first line nobody can verify reads as a guess.
How do you stop sending from breaking deliverability?
Deliverability breaks when a sender looks like a bulk sender before it has earned that standing. An agent can send in an hour what a person sends in a week, so the guardrails have to be code, not habit.
Google's email sender guidelines set the bar for Gmail recipients. All senders need SPF or DKIM. Senders above 5,000 messages a day to Gmail need SPF, DKIM and DMARC, plus one-click unsubscribe on marketing mail. Google also asks you to keep the spam rate shown in Postmaster Tools below 0.10% and never reach 0.30%. Most small teams send far below 5,000 a day, yet the spam-rate guidance still applies to them.
Fix. Build the limits into the sending layer:
- a daily cap per inbox, raised in steps over several weeks for a new domain;
- a separate sending domain from your main company domain, with SPF, DKIM and DMARC set up before the first send;
- a suppression list checked before every send, covering unsubscribes, bounces and complaints;
- an automatic pause on an inbox when bounces or complaints cross a threshold you set.
The pause rule is the one teams skip. Without it, a bad list keeps sending until the domain is damaged.
Why do agents misread replies?
Replies break because language is messy. "Not now, try after Q2" is not "not interested". "Please remove me" inside a longer paragraph is an unsubscribe. An out-of-office that names a colleague is a lead for the colleague.
A weak system sorts into two bins: positive and negative. A usable one sorts into at least six:
| Label | Next step | |---|---| | Interested | Alert a named person, same hour | | Not now (with a date) | Schedule a task for that date | | Wrong person | Ask for the right contact, once | | Unsubscribe | Stop all channels, log it | | Out of office | Pause, resume on return date | | Question or complaint | Stop automation, send to a person |
Fix. Measure the sorter before you trust it. Take 100 real replies, have a person label them, and compare. Look only at the two costly errors: an interested reply labelled "not now", and a missed unsubscribe. Set a rule that low-confidence replies go to a person rather than to a guess. KYN's reply desk costs about S$0.01 per reply sorted, so sending the uncertain ones to a person costs little.
What goes wrong at the handoff to a person?
The handoff breaks when the agent finds a warm reply and nobody acts on it. This is the quietest failure, because the dashboard looks healthy: sent, opened, replied. The reply sits in a shared inbox for three days.
Three causes recur:
- no named owner for each reply type;
- no deadline, so "interested" has the same urgency as "newsletter";
- no context, so the person opens a bare reply and has to dig through the CRM.
Fix. Define the handoff as a record with five fields: who the contact is, what the agent sent, what they replied, the label, and the named owner with a time limit. Send it to WhatsApp or the CRM task list the owner already uses. If the owner has not acted within the limit, escalate to a second named person.
Which fix pays back first?
Start with the list gate. It costs the least and prevents the failures downstream. A rough order:
| Order | Fix | Why first | |---|---|---| | 1 | List gate and source rule | Bad input ruins every later step | | 2 | Sending limits and suppression | Domain damage is slow to repair | | 3 | Reply sorter with a test set | Misread replies cost deals | | 4 | Handoff with owner and deadline | Warm replies decay by the day |
If you only have time for one thing this week, run the 100-reply test. It tells you more about your system than any dashboard.
For the wider picture of what an agent can and cannot do, read what an AI sales agent is, and what it cannot do. For where a person should step in, see when an AI sales agent should escalate.
Where KYN fits
KYN builds outbound as part of a larger system in the client's own cloud account. For an insurance brokerage, we built outreach with one checkable first line, at about S$0.02 per email, and a reply desk that sorts each reply at about S$0.01. Every action is logged with the actor. Outbound messages wait for a named person. The agents run on their own restricted accounts, and the database uses row-level security.
The first 14 days are a review of your stack and interviews with your team. A full build ships in under 90 days. See sales systems and security.
Frequently asked questions
What is the most common reason AI outbound fails?
Bad input data. The agent writes fluent copy from stale or duplicated records, so the failure looks like a copy problem when it is a list problem. A list gate with a required source fixes most of it.
How many emails a day is safe to send?
It depends on the domain's age and history, so there is no single number. Start low on a new sending domain, raise the cap in steps, and watch bounce and complaint rates. Google asks senders to keep spam rates below 0.10%.
Should an agent reply to interested leads itself?
No. A warm reply carries your name and often a price or a date. The agent should alert a named person with the full context and let that person answer.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
