AI Sales Automation – Lead Generation Bot & Customer Acquisition

How an AI sales bot works – from the first message to a booked meeting

A sales bot takes over the first contact with an inbound lead: it answers questions, collects missing details, checks fit against your customer profile and steers the conversation to one of a few outcomes. A qualified lead books a meeting, a lead that isn't ready goes into a follow-up sequence, and a poor fit gets a polite answer and something useful to read.

1. Entry. A chat message, a submitted form, an email to the sales address or a messenger thread. The bot sees the source (campaign, landing page, UTM parameters), because it changes the context of the conversation.
2. Enrichment. Before asking anything, it checks what's already known: is the contact in the CRM, does the company have an open deal, what company does the email domain belong to. It doesn't ask what it can find out itself.
3. Qualifying conversation. It answers the lead's questions and picks up the missing criteria along the way: the problem, scale, timing and the person's role in the decision.
4. Scoring and routing. The criteria turn into a score and a path: a meeting with a rep, a sequence, a handoff to a person or the end of the conversation.
5. Action. It books a slot in the right rep's calendar, saves the contact and a summary to the CRM and starts the sequence.
6. Handoff. Before the meeting the rep gets a note: who it is, which company, what problem, what was already said and what to probe.

Speed is the main argument here: the bot replies while the lead is still on your site. I describe the whole acquisition funnel, including finding companies and enriching data, in B2B lead generation automation. This service is mainly about inbound leads, people who contacted you on their own.

Lead qualification criteria – turning your ICP into bot questions

I set qualification criteria from your CRM data and from talking to your reps, not from a generic framework. I look at what separates won deals from lost ones and from leads that never reached a proposal. Those differences become the bot's questions.

BANT or MEDDIC – what works in a bot conversation

BANT (budget, authority, need, timeline) is a simple filter you can check in a short conversation. MEDDIC (metrics, economic buyer, decision criteria, decision process, identified pain, champion) is for running complex deals and needs information a lead on a chat widget rarely gives. So the bot collects a BANT-style baseline and the rep fills in the MEDDIC parts in the meeting.

Disqualifying signals

• A service or market you don't serve (another country, consumers when you sell B2B).
• Scale below the break-even point your reps identified from past deals.
• Inquiries from competitors, students writing a thesis or vendors who want to sell to you. The bot points them to the right address.
• Job applicants. The bot shares the careers page instead of creating a deal.

Conversation design – qualifying leads without interrogating them

A good qualifying conversation starts by answering the lead's question, not by asking the bot's. I collect the criteria along the way, with two or three questions at most, and only about what can't be found out another way.

Buyers increasingly want to do their own research before talking to a rep. In a Gartner survey published in March 2026 (646 B2B buyers, fielded in August and September 2025), 67% of respondents preferred a rep-free buying experience. A bot that answers a pricing question with "it depends, let's book a call" works against that preference.

Rules I apply to every bot

• Value first. The bot answers the question concretely from the offer knowledge base, then asks for context.
• Straight answers on price. If you publish prices or ranges, the bot shares them. If you don't, it explains what drives the price and offers a quote.
• One question at a time. Closed questions come with answer buttons, which shortens the conversation on a phone.
• No promises. The bot never promises discounts, delivery dates or features that aren't in the knowledge base. Negotiation stays with the rep.
• Objection answers from the field. The usual concerns (price, implementation time, integrations, security) get answers written from what your best reps say.
• A way out at all times. The lead can ask for a person at any moment, or leave a number and come back later.

Example conversation

1. Lead: "How much does it cost for a company with three warehouses?"
2. The bot gives the range from the price list, explains what affects the price and asks which warehouse system they use.
3. The lead names the system. The bot confirms the integration exists and asks about the planned start date, with answer buttons.
4. The timing fits and the company behind the domain matches the profile. The bot offers three slots with the rep who covers that segment.
5. After booking, the lead gets a confirmation and the rep gets a CRM note with the whole conversation.

Handoff rules are the same as in the customer support AI agent: a request for a human is a hard rule, not something the model gets to interpret.

Lead scoring and routing – who gets a meeting and who gets a sequence

I calculate the outcome with rules applied to the data the bot collected, not with the model's gut feeling. The model extracts structured fields from the conversation, and thresholds agreed with the sales team and stored in configuration decide the path.

Who the lead goes to

• Existing customer or open deal: always to the account owner in the CRM, with no new booking by the bot.
• Segment or region: to the rep who owns it, round-robin within the team, holidays taken into account.
• Large accounts: human review before booking, even with a high score.
• Borderline score: a manual review queue instead of an automatic decision.

Technically, I extract fields from the conversation with structured outputs and run calendar and CRM actions through tool calling with limited permissions.

Meeting booking and CRM integration – exactly what the rep receives

The bot books the meeting directly in the rep's calendar and saves the contact, company, deal and a conversation summary to the CRM. A rep should never open a call by asking questions the lead has already answered.

Fields I write to the CRM after every conversation

• Contact and company details, after checking whether the record already exists (deduplication by email and domain).
• Source: channel, campaign, landing page, UTM parameters.
• Qualification criteria values and the resulting path.
• A summary of a few sentences: problem, scale, timing, questions asked and objections.
• The reason for disqualification or handoff as a picklist rather than free text, so it can be reported on.
• A link to the full transcript.

The calendar, and meetings that actually happen

• The bot shows only free slots, with buffers between meetings and a minimum notice period.
• The confirmation includes the meeting link, an agenda and one prep question, such as which system the company uses.
• A reminder before the meeting with one-click rescheduling. A moved meeting beats a missed one.
• If the lead doesn't show up, the CRM gets the status and the bot offers a new slot.

The service page lists the systems I work with most often; anything else I connect through its API. More on what AI changes in CRM work is in AI CRM and sales automation, and on generating proposals after the meeting in AI sales proposal automation.

Follow-up, cold email and the law – consent rules, GDPR and the AI Act

The bot can send follow-ups to a lead who agreed to be contacted in that channel. I don't automate mass sending to people who never got in touch, because in Poland sending commercial information by email or phone requires prior consent, including between businesses. Rules in other EU countries come from the same ePrivacy framework but differ in detail, so check the market you sell into.

What the rules say in 2026

• Poland's Electronic Communications Law (PKE). In force since 10 November 2024. Article 398 prohibits using telecommunications terminal equipment and automated calling systems to send commercial information, including direct marketing, without the prior consent of the subscriber or end user. As the Prawo.pl (Wolters Kluwer) commentary notes, the protection covers all subscribers, legal entities included, and consent has to be specific to the channel.
• GDPR. Consent under telecom rules doesn't replace a legal basis for processing personal data. You still need a privacy notice, a retention period and an easy way to object. More in GDPR and AI.
• AI Act. According to the European Commission, the Article 50 transparency obligations apply from 2 August 2026. A bot talking to leads introduces itself as an AI assistant in its first message.
• Platform terms. Meta requires opt-in before you message someone on WhatsApp, naming the business they agree to hear from. LinkedIn's User Agreement prohibits bots and other unauthorised automated methods, including for sending messages.

How I build follow-up within those rules

• Consent collected in the conversation or the form, separately for each channel, with the wording, date and source stored in the CRM.
• The sequence stops when the lead replies, books a meeting or opts out, and an opt-out in one channel reaches the CRM immediately.
• Every message refers to what the lead actually asked about instead of sending everyone the same template.
• Before launch I ask your lawyer or data protection officer to approve the consent wording and the legal basis. I describe technical practice, not legal advice.

Email deliverability

Even a compliant sequence fails if it lands in spam. According to Google's email sender guidelines, anyone sending to personal Gmail accounts must set up SPF or DKIM and keep the spam rate in Postmaster Tools below 0.3%. Senders of more than 5,000 messages a day need SPF, DKIM and DMARC, and marketing messages must support one-click unsubscribe. That's why I send follow-ups from the same properly configured domain, at low volume, and only to people who expect them.

Where a sales bot runs – website, forms, email, messaging apps

I launch the bot where most of your leads come from and adapt its behaviour to the channel. On the website it's about fast answers and page context, on forms about an instant reaction to the submission, and in messaging apps about the platform's rules.

If pre-sales questions are mostly about products in an online shop, e-commerce AI automation or a support agent that answers from the catalogue is a better starting point.

Built-in CRM chatbot or a custom sales bot – how to choose

A chatbot built into your CRM or marketing platform is enough when qualification comes down to a few questions and all of sales lives in that one system. A custom build makes sense when the bot needs data from several systems, has to hold a natural conversation backed by knowledge of your offer, or needs routing the platform can't do.

Testing and measuring a sales bot

Before launch I check that the bot qualifies leads the way your reps would. After launch I measure held meetings with well-matched leads and what happens to them further down the funnel, not the number of conversations.

Pre-launch tests

• Agreement with reps. A set of past inquiries gets scored twice, once by the bot and once by a rep. The disagreements show which criteria need work.
• Difficult leads. Someone who only asks about price, answers in fragments, writes in another language or angles for a discount.
• Abuse attempts. Instructions hidden in a message, attempts to extract the system prompt or other customers' data. More in prompt injection.
• Integrations. Booking an already taken slot, an existing CRM contact, a rep on holiday, a different time zone.

Every week I review a sample of conversations with a rep, including the disqualified ones. Methods for automated conversation scoring are in LLM evaluation, and how to calculate the return in AI automation ROI. Other services in this area are on the AI & Automation page.