What a good AI reply looks like

A good AI reply answers the question in one or two sentences with a real fact, repeats two or three specifics the prospect gave, and ends with exactly one next step. It admits a mistake in one honest sentence. The worst replies acknowledge with no next step, polish every sentence until it sounds like a brochure, or ask question after question.

We read a great many agent conversations, our own and customers', because the reply is where an AI SDR either keeps a person or loses them. This is what separates the good ones. It is qualitative on purpose: scoring individual accounts is internal, but the patterns repeat.

Trait 1: it answers the question that was asked

It sounds obvious. A prospect writes "what do you actually offer for outbound?" and the reply begins with "Great question! Thanks so much for reaching out!". The question is still unanswered three sentences later.

A good reply starts with the answer. In one of our own threads a product leader asked what we offer for outbound and mentioned his target buyers. The reply, three minutes later: "Good question. It's a system where AI agents manage the full outreach process, engaging leads like the OTEs and trip planners you're targeting." One sentence of answer, tied to his words.

Trait 2: it uses a real fact

The answer needs something true and specific behind it. "We help companies grow" is not a fact. "Agents answer in about two minutes, including at night" is one, and it can be checked.

The corollary matters more: when the agent does not have the fact, it must not produce one. A language model will happily fill a gap with something plausible, and that is the most expensive failure in AI replies, because the prospect may act on it. The fix is not a better tone. It is to give the agent exact capability guardrails on day one, as in the Lamothermic case study, where the owner added lines such as which alloys the foundry does not cast, and to let the agent say "I will check and come back" and hand over. See should an AI agent give a price quote or stall.

Trait 3: it repeats what the prospect said

People want to feel heard. The cheapest way is to repeat two or three specifics they gave: the segment they target, the tool they built, the problem they mentioned. It also proves the message was read.

In another thread, a consultant who had built his own outreach system with workflow tools asked for data. The reply credited what he had built before moving on. A different consultant described how he optimises processes and the agent answered with that detail: "That's a really comprehensive system you've built, especially the self-improving prompts idea." Then it connected it to the offer and asked for 15 minutes.

Repeat the specifics, but do not flatter. Mirror, do not gush.

Trait 4: it ends with one next step

One. Not "would you like a demo, or a trial, or to see a case study, or I could send a link?" A reply with four options asks the prospect to do the work of choosing. Pick the step that fits the message.

• Factual question: answer, then offer to show it.
• Interest: a calendar link or two times.
• Objection: acknowledge, give one fact, ask one question.
• Not now: a clean exit and an open door.

The first-reply pattern in our data: the booking link typically went out after three prospect messages, and 56% of bookings confirmed in the chat needed at least one more nudge after the link. The next step has to be small and repeatable.

Trait 5: it owns a mistake in one sentence

Our host once joined a 15 minute call 15 minutes late. The prospect wrote that he was not interested. The agent replied within the hour: "you're absolutely right. My apologies for being late and wasting your time. I respect your decision." Twenty six minutes later he wrote back that honesty is important too, and rebooked.

One sentence, no excuse, no reframing, no offer wrapped around it. Long apologies make the problem bigger.

What do bad replies look like?

• Acknowledgement with no next step. "Thanks so much for your response, that is really helpful!" and nothing else. The thread stalls and the next move is the prospect's.
• Over polish. Every sentence balanced, every reply the same length, and a closing line repeated word for word. Prospects notice. Vary length with the message.
• Interrogation loops. Three questions in a row. After a prospect has answered two, the agent should offer something, not ask a third.
• Reminders that drop the question. A follow up that does not carry the original ask forces the person to scroll.
• Claims about what the product does. A feature mentioned that is not live, or a number that does not exist.

What does a weak reply look like when rewritten?

The prospect message and both replies below are made up for illustration, not taken from a customer thread.

Prospect: "We target dental clinics in two cities. How would you handle follow ups if someone goes quiet?"

Weak reply: "Thanks so much for your message, that is a great question! We offer a comprehensive solution with many powerful features. Would you like a demo, a trial or a case study? Also, how many clinics do you have, and what tools do you use today?"

It does not answer, offers three next steps, asks two new questions and names no fact.

Better reply: "If someone goes quiet, the agent sends follow ups on a schedule you set, usually a few days apart, and stops the moment they reply or opt out. For dental clinics in two cities I would keep it to two or three touches. Want me to show you the follow up settings on a 15 minute call?"

It answers first, uses a fact about how follow ups work, repeats the clinics and the cities, and ends with one step.

How long should a reply be?

Match the message you received. A one line question gets two or three sentences. A long message that lists the prospect's setup gets a reply that picks up the two or three details that matter and leaves the rest. Replies that are always the same length are one of the over polish signs listed above. A practical rule for a rule file: answer in the first sentence, keep the whole reply under about 60 words unless the question needs more, and end with the one step.

What should the agent do when it does not know?

Say so and hand over. "I will check and come back to you today" is a good reply when the answer is a price, a date, a legal point or a feature the agent has no record of. The platform stops the agent on its own when it cannot answer, for example a price it has no answer for, and a person can take over any conversation. A short honest delay costs far less than a confident wrong answer, which the prospect may act on.

Does speed matter more than wording?

Both matter, in different ways. The median agent reply took 2.3 minutes on our own account since May 2026, and speed alone does not save a reply that stalls the thread. A fast reply with no fact and no next step is just a fast acknowledgement. Treat speed as the entry fee, and wording as what earns the next message. The handoff to a person is where threads are most often lost, which the speed and handoff post covers.

How do you check replies on a weekly basis?

Sample twenty threads from last week and read them in order, not in a dashboard. For each reply ask five questions:

1. Did it answer what was asked?
2. Did it rest on a fact the agent actually has?
3. Did it repeat something the prospect said?
4. Did it end with exactly one next step?
5. Did it avoid promising anything the business cannot deliver?

Count the failures by type. If one type dominates, fix the rule that governs it, not the single message. Where a conversation matters, approval mode lets a person read the reply before it sends; human in the loop covers when to use it.

What is the point?

Reply quality is the part of an AI SDR you can see and judge. It is also only one of the levers. On our own account the audience and the opener moved results more than reply polish did, as in what 27,737 messages taught us. But a reply that is wrong, vague or invented can undo the work that got the person to write back. For what follows a first reply, read what an AI SDR does after the first reply.