What stops AI outreach from sending the wrong message?
By Matthew Eagle
Founder, Eagle AI Agency. Eight years as a staffing account executive.
Published . Updated .
This week a prospect received our AI's private note about them as the opening message of a sequence. Here is how it got through, how we made sure it was the only one, and what checks the words now.
AI that writes your outreach can put the wrong words in a message and nothing downstream will notice, because a filled-in field looks like a finished job. What stops it is a check on the words themselves before anything sends, a rule that a "not a fit" verdict takes the person out of the sequence, and a person reading any suggested reply before it goes back. We learned this the hard way this week, when a prospect received our AI's private note about them as a LinkedIn message.
What happened when our AI sent its private notes?
One of our outreach sequences uses an AI step for two jobs. First it judges whether a new lead fits the kind of company we are going after. Then it writes a personal opening message and two follow-ups for the people who fit.
For one lead it decided "not a fit", and then wrote that decision into all three message fields. The opener that went out on LinkedIn read, word for word:
"The company's Ideal Customer Profile status does not match the requirements for this motion."
The person was still enrolled in the sequence, so it sent. The next morning they replied with a single question mark.
How did we catch it?
We caught it by reading the conversation itself. A reply came in, the question on our side was simply whether they had replied, and opening the thread showed what we had sent.
No report flagged it. The field had text in it, so as far as every system was concerned the AI step had worked.
How did we make sure it was only one person?
- We paused that one person's sequence, then read it back to confirm the pause held.
- We swept every live enrolment across all four accounts we run this way, more than 3,000 people.
- We read every LinkedIn conversation on the affected account, 1,273 threads.
It was one person and one message. The sweep is the step people skip. One bad message is an apology. The same bug across a whole list is a reputation problem, and you only find out which one you have by checking all of it.
Why didn't anything stop it before it sent?
Three parts each looked fine on their own:
- The AI step finished and filled its fields, so it reported success.
- The fit verdict was stored in one place and the enrolment in another, so a "no" never took anyone out.
- The send delivers whatever is in the field. It has no idea what the words mean.
Every part did what it was built to do, and nothing checked the handoff between them. We wrote about that pattern before in what breaks when you automate a process nobody is watching.
Where does AI-written outreach go wrong, and what catches it?
| What goes wrong | Why nobody sees it | The check that catches it |
|---|---|---|
| The AI writes its own notes or verdict into the message | The field is full, so the step counts as done | Read the words before sending and pause anything that reads like an internal note |
| A lead judged "not a fit" stays in the sequence | The verdict and the enrolment live in different places | A "no" removes the person, it does not just label them |
| The AI decides not to write, but leaves text behind anyway | Something in the field still looks like a message | When the AI passes on a lead, the fields stay empty and an empty field stops the send |
| A suggested reply makes the mistake worse | It is written by software reading the thread | A person reads every reply before it goes back |
| The safety check only runs on a schedule | Anything between runs goes out unchecked | Check at the moment the message is written, as well as on the schedule |
What did we change?
We added a guard that runs every hour across every account. It reads the opener and both follow-ups for every live enrolment and pauses anyone whose message reads like an internal verdict instead of a note to a person.
We tested it both ways. It fired on the message that went out, and it stayed quiet on normal messages.
It has limits, and we would rather say so. It runs hourly, so a bad message could still slip out inside that hour. It also treats the symptom. The real fixes sit further up the chain: a lead marked "not a fit" should never be enrolled, and the AI step should leave the fields empty when it decides not to write. Those are next.
How should you reply when your automation sends something wrong?
The tool that drafts our reply suggestions read the thread and proposed telling the person the message was not from us, then pitching them in the same reply. It was from us. We did not send it.
A reply to a mistake like this should come from a person, own it in a line, and leave the pitch out. This is the same standard we hold the outreach we run in our sales work to: the automation does the volume, and a person owns anything that goes wrong.
Questions people ask
Is it safe to let AI write outreach messages?
Yes, if something checks the words before they go out. Read a sample of real messages before launch, then keep an automatic check running after it.
How do I check AI-written messages when there are hundreds of them?
Look for text that never belongs in a note to a person: verdicts, scores, field names, the language of an internal report. Anything that matches gets paused instead of sent.
What should happen to a lead the AI says is not a fit?
It should leave the sequence. A label that says "no" while the person stays enrolled is exactly how this message got out.
Should I reply when my automation sends something wrong?
Yes, from a person and briefly. Own it, apologize once, and do not pitch in the same message.
If you want AI writing your outreach with checks that catch this before a prospect does, book a call.
More notes
- Why do verified emails still bounce?This week a list that passed two separate email checks still bounced hard enough to stop its own campaign. Here is what the checks missed, and what we now run before every send.
- How do you move to a new CRM without losing your history?We moved a small firm's years of CRM history into a new system and checked every record. The next day they told us the notes had not come over. They had.
- Why did the link in your email open a warning page?We approved an email with a clean payment link. Our customer clicked it and landed on a warning page instead.
- Can you use AI on work that is under an NDA?A confidentiality agreement that never says the word AI can still rule out the tools you were about to use.
- Is your outreach broken, or is the benchmark wrong?The fastest way to wreck a campaign that is working is to measure it against a number that came from somewhere else.
- Do more pages actually help your website rank?The advice to publish more assumes pages are the unit of measurement. They are not, and a site full of near identical pages can be worth less than a site with a handful of good ones.
- How do you check if your cold email is going to spam?We spent days proving our email was landing in spam. Every test was pointed at a kind of mailbox almost nobody on the list used.
- Should a staffing firm hire a recruiter or rent a desk?Most small staffing firms are not short of a recruiter. They are short of a recruiter they can afford to carry while one ramps.
- Why does AI make things up, and how do you stop it?A model handed something unusable does not stop. It produces something plausible, and it will produce a different plausible thing the next time you ask.
- What breaks when you automate a process nobody is watching?Automation rarely fails loudly. It fails quietly, and keeps reporting that it is fine.
- Why did your cold outreach run out of leads?Usually it is not the message. You have simply run out of people to send it to.
- Do you need to hire a salesperson, or a system?The honest answer depends on one thing, and it is not budget.
- What do you do about sales when you are already at full capacity?The trap is not that you are too busy. It is that being busy feels like being safe.
Want the version of this for your business?
Twenty minutes, your actual situation, a straight answer about what we would do first.
20 minutes. You leave with a straight answer.