PortuguêsEspañol

AI Sales Agent in 2027: The SDR That Texts Your Ad Leads in Seconds, Qualifies Them and Fills the CRM

OBy Onne Company · Published
Quick answer: An AI sales agent replies to new leads by text or chat in under a minute, asks the qualifying questions, books a call or visit and writes structured fields into the CRM. Built with n8n, an AI model and a registered SMS number, it costs roughly US$ 50 to US$ 120 a month for 150 leads, and it must never quote prices or promise savings.
AI Sales Agent in 2027: The SDR That Texts Your Ad Leads in Seconds, Qualifies Them and Fills the CRM
In this article
  1. What this agent does, and what it does not
  2. The flow, step by step
  3. The agent's instructions (copy and adapt)
  4. The output that goes into the CRM
  5. A real sample conversation
  6. Follow-up for leads who go quiet mid-conversation
  7. n8n, Make, Zapier or a custom GPT?
  8. How much it costs per month
  9. The rules that apply to this agent
  10. When I would NOT build this agent
  11. Common mistakes
  12. Want to skip this part?
  13. Step by step
  14. Frequently asked questions

A residential solar installer in the Phoenix area spends about US$ 9,000 a month on Facebook and Google ads and gets roughly 150 leads. The owner pulled the numbers for October: leads that got a reply within five minutes booked a site visit almost three times as often as leads that waited until the next morning. And 38% of the leads came in after 6 p.m. or on weekends, when nobody was at a desk. That gap is the job for an AI sales agent: reply in seconds, ask the qualifying questions, book the visit and write everything into the CRM, so the human rep starts the day with a list of people ready to talk.

This post shows the whole build: what the agent does and refuses to do, the flow step by step, the system prompt, the prompt that fills the CRM (with real output), a real sample conversation, follow-up rules, tool choice and the monthly cost. For the bigger picture of where AI helps a sales team, start with the AI for sales guide.

What this agent does, and what it does not

DoesDoes not
Replies to a new lead by text or web chat in under a minute, 24/7Quote a final price or a monthly payment
Asks 4 to 6 qualifying questions (homeowner, roof, power bill, timeline)Promise savings, tax credits or "free solar"
Answers questions from an approved knowledge baseNegotiate or offer discounts
Books a site visit or a call on the rep's calendarPretend to be a person
Fills the CRM deal with structured fields and a scoreKeep texting someone who said stop
Hands the conversation to a human when it shouldReply to complaints or legal questions

The right column matters as much as the left. The agent is a fast junior SDR, not a closer. In solar especially, a bot that says "you'll save 60%" is making a claim your company has to back up, and the FTC has gone after solar sellers for exactly that kind of promise.

Advertisement

The flow, step by step

I will describe it in n8n, which is what I would use. In Make or Zapier the logic is the same; only the block names change.

  1. Lead arrives. A Facebook Lead Ads form, a Google lead form or the website form sends a webhook with name, phone, ZIP code and the consent checkbox.
  2. Consent check. If the form did not capture written consent to receive texts, the flow stops texting and creates a task for a rep to call during business hours. No consent, no automated text.
  3. First text in under 60 seconds. Sent through Twilio (or a similar provider) from a number registered for A2P 10DLC.
  4. Conversation. Every reply goes to the AI model with the system prompt below, the knowledge base and the conversation history.
  5. Booking. When the lead qualifies, the agent offers three open slots from the rep's Google Calendar or Calendly.
  6. CRM. After each conversation, a second AI call reads the history and returns the fields. n8n writes them into HubSpot, Pipedrive or Close.
  7. Handoff. Keywords (STOP, lawyer, complaint, "talk to a person") or the model's own HANDOFF tag alert the rep in Slack or by text.

The agent's instructions (copy and adapt)

This goes into the system prompt field. Rules 3 and 4 are the ones that keep you out of trouble: a savings figure or a price that comes from your number is your offer.

You are the automated assistant for [COMPANY], a residential solar installer serving [AREA, e.g. the Phoenix metro area]. You text with people who asked for information through our ads or website.
Goal: qualify the lead and book a free site visit with a consultant.
Rules:
1. In your first message, say you are an automated assistant. Never claim to be a person.
2. Ask one question per message. Keep messages under 300 characters. Plain, friendly English.
3. Never quote a final price, monthly payment, savings percentage or tax credit amount. If asked, say the consultant calculates it from the actual roof and bill, and offer the visit.
4. Only state facts from the KNOWLEDGE BASE below. If the answer is not there, say you will check with the team and add [HANDOFF].
5. Qualifying questions, in this order: Do you own the home? What is your average monthly electric bill? Roughly how old is the roof? When would you want to start?
6. If the person does not own the home, explain politely that we only install for homeowners and end the conversation.
7. If the person says STOP, wants a human, mentions a complaint, a lawyer or a previous installer problem, reply briefly and add [HANDOFF].
8. Offer exactly three time slots from AVAILABLE SLOTS. Confirm the address only after they pick one.
9. Do not text between 8 p.m. and 8 a.m. Arizona time. If a message arrives then, reply once and say the team follows up in the morning.
KNOWLEDGE BASE: [products, service area by ZIP, financing partners, warranty, how the site visit works]
AVAILABLE SLOTS: [inserted by the automation]

Rule 9 is stricter than the federal quiet hours on purpose. It keeps you inside the TCPA window everywhere and avoids the 11 p.m. text that annoys people into reporting spam.

The output that goes into the CRM

The second AI call does not talk to anyone. It reads the history and returns fields. This prompt is why the CRM gets filled without a rep typing anything:

Read this text conversation between our assistant and a solar lead. Return JSON with:
homeowner (yes/no/unknown), monthly_bill_usd (number or null), roof_age_years (number or null), shading_mentioned (yes/no), timeline (now / 1-3 months / 3+ months / unknown), financing_interest (cash / loan / lease / unknown), objections (list, in the lead's words), visit_booked (date and time or null), score (A = owner, bill over US$ 150, starting within 3 months; B = owner, missing one criterion; C = everything else), next_step (one sentence for the rep).
Use only what the lead said. Do not guess.
Conversation: [HISTORY]

A real sample conversation

I ran the flow with the instructions above and a test knowledge base: residential systems, financing through two partner lenders, service in Maricopa County. The lead filled the Facebook form at 6:12 p.m. on a Saturday, with the texting consent box checked. Real output, lightly trimmed:

Assistant (6:12 p.m.): Hi Kevin, this is the automated assistant for Sunridge Solar. Thanks for your request! Quick question so we send the right consultant: do you own the home where you'd install the panels?

Kevin: yes we own it. Gilbert

Assistant: Great, Gilbert is in our service area. What's your average monthly electric bill, roughly? Summer and winter are fine as separate numbers.

Kevin: like 380 in summer, maybe 140 in winter. how much would it cost

Assistant: Those summer bills are common out here. Honest answer on price: it depends on roof size, shading and your usage, so the consultant calculates it at the free visit and gives you cash and financing options side by side. How old is your roof, more or less?

Kevin: 8 years. my neighbor said he pays nothing now is that true

Assistant: Some homeowners get their bill very low, but it varies a lot by system and usage, so I don't want to promise a number. The consultant can show you a projection based on your actual bills. When would you like to get started?

Kevin: asap before next summer

Assistant: Makes sense. I have these times for a free visit (about 45 minutes): Mon 11/16 at 9:00 a.m., Mon 11/16 at 4:30 p.m., or Tue 11/17 at 10:00 a.m. Which works?

Kevin: monday 4:30. wife will be there

Assistant: Booked for Monday 11/16 at 4:30 p.m. Can you confirm the street address in Gilbert? Having your wife there is great, since the consultant goes over the numbers with both of you.

And the CRM extraction from that conversation:

{"homeowner": "yes", "monthly_bill_usd": 380, "roof_age_years": 8, "shading_mentioned": "no", "timeline": "1-3 months", "financing_interest": "unknown", "objections": ["how much would it cost", "my neighbor said he pays nothing"], "visit_booked": "2026-11-16 16:30", "score": "A", "next_step": "Bring a projection using summer and winter bills (US$ 380 / US$ 140); both spouses will attend; ask about financing preference early."}

Note what the assistant did with "my neighbor pays nothing." It did not confirm, did not deny, and moved toward the visit. That behavior comes straight from rule 3. Without it, models tend to say "yes, many customers eliminate their bill," which is the sentence you do not want in a screenshot.

Advertisement

Follow-up for leads who go quiet mid-conversation

About a third of leads stop replying after the second or third question. The rules I would use:

  • Nudge 1, after 4 hours (inside allowed hours): one short text that repeats the open question. "No rush, Kevin. Whenever you have a sec: roughly how old is the roof?"
  • Nudge 2, the next day: a different angle, still a question. "Would a quick phone call be easier than texting? I can have a consultant call you."
  • Nudge 3, after 3 days: the breakup. "I'll close this out for now. If you want a free visit later, just reply here."
  • Then stop. Three nudges, then the lead goes to a monthly nurture email list, if they gave email consent.

Every text includes a way to opt out ("Reply STOP to opt out" in the first message is the norm), and a STOP reply ends all automated messages immediately. Mark the contact in the CRM too, so no other tool texts them.

n8n, Make, Zapier or a custom GPT?

OptionGood forLimit
n8nFull control, lower cost at volume, self-hosting optionSteeper learning curve
MakeVisual building, good HubSpot and Twilio modulesGets expensive with many operations
ZapierFastest to set up, huge app catalogMost expensive per task; long conversations get clunky
Custom GPTTesting the instructions and the toneCan't text leads or write to your CRM by itself

My firm opinion: prototype the instructions in a custom GPT or Claude project for a day, then build the real thing in n8n. And text from a properly registered business number. Personal-phone workarounds and unregistered numbers get filtered by the carriers, and a blocked number in the middle of the season costs more than the whole setup.

How much it costs per month

For the installer's 150 leads a month, with n8n in the cloud. Check each vendor's current pricing.

ItemApproximate monthly cost
n8n Cloud (Starter)About US$ 20 to US$ 25
AI model via API (small model, 150 conversations plus extraction)US$ 3 to US$ 10
Twilio SMS (about 2,000 segments) plus numberAbout US$ 20 to US$ 25, plus carrier fees
A2P 10DLC brand and campaign registrationOne-time fee plus a small monthly campaign fee
CRM (HubSpot free or Pipedrive)US$ 0 to about US$ 50 per rep
TotalRoughly US$ 50 to US$ 120

Compared with US$ 9,000 in ad spend, the agent costs about 1%. If it lifts the booking rate on after-hours leads by even a few visits a month, it pays for itself many times over. Plan two or three weekends for the setup.

The rules that apply to this agent

  • TCPA consent. Automated marketing texts need prior express written consent. Put a clear consent checkbox on the form, with your company name, and keep the record (time stamp, form version).
  • Do Not Call. If a rep calls leads who did not give consent, scrub against the National Do Not Call Registry and your internal list.
  • Disclosure. California's bot disclosure law requires it when selling to Californians; I would disclose to everyone. Rule 1 handles it.
  • Claims. Savings, tax credits and "free" are the claims regulators watch in solar. Keep them out of the bot.

When I would NOT build this agent

  • Fewer than 50 leads a month. A rep with phone notifications can answer that volume. The automation does not pay back the setup.
  • Enterprise B2B with long cycles. A VP of operations who filled a demo form does not want to text with a bot. Fast human email works better.
  • No written process yet. If the team cannot say what makes a lead qualified, the agent will qualify at random.
  • Forms without consent. Fix the form first. Texting without consent is the expensive mistake here.

Common mistakes

  1. Letting the bot "help" with price. One "around US$ 18,000" in a text becomes the number the customer quotes back to your consultant.
  2. Asking all the questions in one message. Reply rates drop; one question per text works.
  3. No owner for HANDOFF alerts after hours. Decide who gets them before going live.
  4. Forgetting the CRM extraction. Without it, you have a fast bot and the same empty CRM.
  5. Never reading the conversations. Read 20 a week for the first month and fix the instructions.

The same design, adapted to listings and showing requests, appears in the AI guide for real estate agents. And if your leads arrive through Instagram DMs instead of forms, the AI for social media guide covers keyword automation in the inbox.

Ready-made AI agents for sales reps

Want to skip this part?

Ready-made AI agents for sales: nothing to build, no prompts to write, they already work for you by text, chat and in the CRM.

  • Replies to every ad and website lead in seconds, including Saturday night
  • Asks the qualifying questions, books the meeting and fills the deal in your CRM
  • Follows up with leads who went quiet and hands you only the ones ready to buy
I want the ready-made agents
Advertisement

Step by step

  1. Add a clear texting consent checkbox to your lead forms
  2. Register your business number for A2P 10DLC
  3. Write the knowledge base and the system prompt with the price and savings rules
  4. Build the flow in n8n: webhook, consent check, AI reply, calendar, CRM
  5. Add the CRM extraction prompt and HANDOFF alerts
  6. Read 20 conversations a week and adjust the instructions

Frequently asked questions

What is an AI sales agent?

It is an automation that talks to new leads by text or chat, asks qualifying questions, books meetings and fills the CRM. It works like a fast junior SDR and hands off to a human for pricing, negotiation and complaints.

Does an AI sales agent replace the SDR?

It replaces the first reply and the data entry, which is where most of the delay happens. A person still calls the qualified leads, handles objections and closes.

Can the agent text leads from a regular phone number?

Use a business texting provider with A2P 10DLC registration. Unregistered or personal numbers get filtered by carriers, and the TCPA requires written consent for automated marketing texts either way.

How much does an AI sales agent cost?

For about 150 leads a month, roughly US$ 50 to US$ 120 in automation, AI, SMS and CRM costs, plus setup time. Ready-made agents skip the setup for a monthly fee.

Which CRM works with an AI sales agent?

Any CRM with an API: HubSpot (free plan), Pipedrive and Close all work with n8n, Make and Zapier.

Do I have to tell leads they are talking to AI?

California requires disclosure when a bot is used to sell, and honesty helps everywhere. A first message that says "automated assistant" solves it.

Sources

O

Onne Company · the Núcleo Aprendizado team

We test, compare and explain what is worth learning, with a focus on what gets results at work. How we work.

Advertisement

Read next