--- title: "My AI answered 900+ sales conversations while I slept, for $110. Every mistake it made" description: "How an AI held 911 sales text conversations in 6 weeks for about $110: how it's set up, the rules it follows and every mistake it made." url: https://talalevel.com/blog/ai-answered-900-sales-conversations/ date: 2026-11-08 updated: 2026-11-08 --- # My AI answered 900+ sales conversations while I slept, for $110. Every mistake it made An AI that answers warm leads' texts held 911 conversations in its first 6 weeks, with a median reply under 2 minutes and a total AI bill of about $110. It waits a minute or two before answering, has one goal per conversation, checks every number before a text goes out, knows when to say nothing, hands anything hot, weird or angry to a person and never pretends to be a person. I live in Madrid. My leads are in the US. So when a business owner in Texas texts back "yeah I'm interested, what do you need from me" at 8pm his time, it's 3am for me. By the time I wake up, he's probably talked to two other brokers. I know I lost deals like that. I'm still pretty new at all this, and I don't know everything. But here's what I've figured out so far. I sell business funding with a small team. In funding, speed is everything. Whoever gets the bank statements first usually gets the deal. And texting was where we were slowest. So at the end of September I built an AI that answers our leads' texts for us. Not a bot that says "Thanks! A rep will reach out shortly." It actually holds the conversation, asks for what we need, reads the bank statements people send, and hands the lead to one of us when it should. Here's what it did in the first 6 weeks: - 911 conversations - 5,326 texts and 309 emails sent - median reply time: under 2 minutes - about 1 in 5 of its replies answered a text that came in at night or on a Sunday - 48% of its texts got a reply within 48 hours - about 1% replied STOP - 49 times it read the bank statements someone sent and told them what was missing - 532 times it handed the lead to a person - 355 times it looked at a conversation and decided the right move was to say nothing - total AI bill: about $110 ![The AI texter's first 6 weeks, Sep 29 to Nov 8: 911 conversations held, 5,326 texts sent plus 309 emails, a median reply under 2 minutes, 1 in 5 replies to texts that came in at night or on a Sunday, 48% of its texts got a reply within 48 hours, about 1% said stop, 532 handed to a person, $110 total AI bill](https://talalevel.com/img/blog/ai-texter-numbers.webp) Here's exactly how it's set up, and what it got wrong. Steal whatever's useful. ## 1. It only works warm leads It doesn't blast cold lists. It picks up once someone is actually interested: they replied yes, they talked to one of us on a call, or they already sent statements. Cold outreach stays with us. ## 2. It waits before it answers People text in bursts. Almost 1 in 10 send a second message right after the first, most within a minute. If the AI answers the first bubble instantly, it answers the wrong question. So it waits a minute or two, reads everything, then replies. ![What happens when a lead texts back: the AI waits 1 to 2 minutes, reads the whole thread, the call and the file, then replies with one goal, hands the lead to a person or says nothing; with no answer, up to 3 follow-ups between 7am and 10pm in the lead's time zone](https://talalevel.com/img/blog/ai-texter-flow.webp) ## 3. It has one job per conversation For us that's the bank statements. Early on it asked a bunch of questions first and only asked for statements after like 6 texts back and forth. That's 6 chances for the lead to go cold. Now it asks as soon as it knows how much they need and what for, ideally by the third reply. ## 4. It's not allowed to make up numbers This one matters most. Before any text goes out, every number in it gets checked against what the lead actually said or sent us. If a number came from nowhere, the AI gets one shot to rewrite it. If it fails again, a person gets the conversation. It caught it once inventing loan amounts nobody asked about. ![The number check: every number in a reply is checked against what the lead said or sent; one that doesn't match gets one rewrite, and if it is still wrong a person takes the conversation](https://talalevel.com/img/blog/ai-texter-number-check.webp) ## 5. It knows when to shut up Someone says "thanks", someone's clearly busy, someone says no. The AI doesn't need the last word. Follow-ups are capped at 3 and only go out between 7am and 10pm in the lead's time zone. Replies can go any time, because they texted first. ## 6. It hands off fast Anything hot, weird or angry goes to a person with a push notification. A good AI isn't one that never needs a human. It's one that knows exactly when to get one. ## 7. It never pretends to be a person If someone asks "is this a bot?", it says yes in one honest line, tells them a person on our team reviews every conversation, and gets straight back to what the file needs. ## What it got wrong It wasn't smooth. A few that stung: - I added one line to its instructions about follow-ups ("the statements, or a quick call"). That night it followed up with a bunch of leads offering calls and asking for screenshots of all four months. One sentence changed hundreds of conversations. Lesson: replay every change on real past conversations before it goes live. - A call note said "statements from March". The AI took it literally and kept asking for March to May. It trusts whatever one of us writes down, so what we write down has to be right. - It booked a meeting and sent the calendar invite to the email on our list instead of the guy's actual email. Now invites only go to an address the lead clearly owns. - The number check was too strict at first. It kept reading a bank account name like "Chase 5230" as a dollar amount and blocking perfectly good replies. ## If you want to copy this - Warm leads only - One goal per conversation - Check every number before it's sent - Let it stay quiet - Make the hand-off instant - Replay every change on old conversations first You can build this yourself with a good model and a texting API. That's honestly how I started. It's one piece of a whole CRM I ended up building mistake by mistake (calling, pipeline, texts, all of it), so if you're building one and get stuck, write to me at hello@talalevel.com. I'm not gonna hide anything. ## Questions ### What did the AI do in its first six weeks? It held 911 conversations, sent 5,326 texts and 309 emails and replied in under 2 minutes at the median. 48% of its texts got a reply within 48 hours, about 1% replied STOP, and it handed the lead to a person 532 times. The total AI bill was about $110. ### Why does the AI wait before it answers a text? People text in bursts: almost 1 in 10 send a second message right after the first, most within a minute. Answering the first bubble instantly answers the wrong question, so it waits a minute or two, reads everything, then replies. ### How do you stop an AI texter from making up numbers? Before any text goes out, every number in it is checked against what the lead actually said or sent. If a number came from nowhere, the AI gets one shot to rewrite it; if it fails again, a person gets the conversation. ### When do the AI's follow-ups go out? Follow-ups are capped at 3 and only go out between 7am and 10pm in the lead's time zone. Replies can go any time, because the lead texted first. ### Does the AI say it is a bot? Yes. If someone asks "is this a bot?", it says yes in one honest line, tells them a person on the team reviews every conversation, and gets straight back to the conversation.