Ten things we learned building an AI that talks to prospects

After a year of running AI agents on our own LinkedIn outreach and on customer accounts, ten lessons stand out: the list matters more than the wording, observations beat questions, follow ups add replies and cost goodwill, speed is table stakes, a pleasantry is not interest, booked is not held, facts need rules, handoffs rot, machine mail is not a reply, and opt outs must be mechanical. Each lesson below points to the numbers that taught it.

We make software that writes to strangers, which means we get to be wrong in public a lot. These are the lessons we now build around. They are not a boast. Some of them cost us weeks, and a few are the kind that sound obvious once someone says them.

1. The list matters more than the wording

We spent months tuning messages. Then we compared lists. Intent signal lists replied at 54.6% and booked 4.54 meetings per 100 people messaged. Untagged lists replied at 37.2% and booked 1.16. Sales Navigator searches replied at 21.7% and booked 0.71. The source of the signal mattered too: people who engaged with AI SDR vendors' and automation creators' posts booked at 2.6% to 3.7%, those who engaged with enterprise sales thought leaders under 0.4%. The details are in the Sellinger case study. The lesson: before you touch the wording, ask who is on the list and why now.

2. Observations beat questions

A message that states something true about the person's work out-replied a message that asks a question by 10 to 15 points within the same month: 51.9% against 37.3% in June 2026, for example. We expected the opposite, because questions invite answers. In practice a question reads like a script, and an observation reads like attention. Observation openers beat questions has the examples.

3. Follow ups add replies and cost goodwill

The first message alone got 26.5% of people to reply. With follow ups the total reached 38.4%, 45% more. But the share of first replies that were negative rose from 0.9% on the intro to 14.7% by the third follow up. Both numbers are true at once. So we cap follow ups, make each add something new, and end with a clean exit line. The Shelly installer programme case study gives the same advice in its own words: plan for the follow up, and cap it where partners start to feel spammed, which there was two follow ups.

4. Speed is table stakes, not the lever

Our agents' reply time fell from about 19 minutes to about 2 over the year. Meetings per conversation did not rise. Speed keeps you in the conversation, and the offer and the list decide whether it goes anywhere. For inbound, speed does more: the Hotel Pirin agent's median first reply to a guest was 41 seconds. The post on speed and the handoff draws the line.

5. A pleasantry is not interest

LinkedIn offers quick replies such as "thanks for connecting". They count as replies and inflate any reply rate. In the referral network case, 1,751 first replies were pleasantries and 2,025 said something real. The agent answered the pleasantries with a question about the prospect's own clients, and 1,035 became real conversations. We now report two numbers: replied, and replied with substance.

6. Booked is not held

Our agents booked 351 showcase meetings through the booking link, and 33 are confirmed held, by a stored recording or a team mark. The calendar marks a meeting completed when its slot ends, whether or not anyone joined. We stopped counting that. The ladder in count meetings honestly came out of this.

7. Facts need rules, not hope

A model knows your business only through what you tell it. At the hotel, the agent answered fast but only half of its replies carried a price. After we added a rule to quote as soon as it had dates and party size, 92% did. The rule changed the behaviour, and no model change did. Put prices, terms and exceptions in a knowledge base and write the rules about when to use them. See what a good AI reply looks like.

8. Handoffs rot

An agent can answer in seconds. A person may not. At the same hotel, after a guest was handed to staff the wait ran from three hours to 14 days, and staff answered group requests in a median of almost four hours. A handoff without an alert, an owner and a clock is a place where leads go to die. We now treat the handoff as part of the product, not an afterthought.

9. Machine mail is not a reply

Out of office messages, helpdesk acknowledgements, satisfaction surveys and delivery failures look like replies. If they count, your reply rate is wrong and your agent may answer a robot. Sellinger now marks these as machine mail: they never count as a human reply, the agent does not answer them, and the lead goes to a person with a named reason. The Shelly programme's case study excludes out of office and delivery notices from its 360 genuine replies for the same reason. The long version is keeping AI agents out of machine mail.

10. Opt outs must be mechanical

When a contact writes STOP or unsubscribe, in English, Bulgarian or German among other forms, the contact goes on the team's suppression list, the agent stops and the lead is closed out. A weak or unclear phrase never writes the permanent record. The agent stops and a person decides. And before every send on email, LinkedIn and WhatsApp, the platform checks the suppression list and refuses the send when the contact is on it or the check cannot be completed. The rule behind it is plain: a system with good intentions and one leak is worse than a boring one with none, so we test the stop paths as carefully as the send paths.

What do the ten have in common?

Most of them are about not fooling ourselves. Reply rate with pleasantries, meetings that were booked and never held, a fast agent in front of a slow human, a model asked to guess facts. In every case the first number looked good, and a second look found the problem.

If you take one thing from the list, take the habit: for every figure you report, write down what it counts and what it leaves out.

What would we tell a team starting today?

Start small and read everything. Run the first fifty conversations in approval mode, with one narrow audience and one clear offer. Read every draft, and read every reply, including the rude ones, because they show you where the message is wrong. Count carefully, then change one thing a week. The agent will do the repetitive work from the first day. What you add is the judgement about who deserves a message and what is true enough to say.

A checklist

• [ ] I know where my list came from and why those people, now.
• [ ] Openers state an observation that is true.
• [ ] Follow ups are capped and each adds something.
• [ ] I report replies with substance, not only replies.
• [ ] I report meetings held, not only booked.
• [ ] Prices and terms live in a knowledge base with rules.
• [ ] Every handoff has an owner and a clock.
• [ ] Machine mail and opt outs are handled by code.