AI Chatbot for Lead Qualification

Qualification is sorting, not gatekeeping

The goal of a qualification bot is not to decide who deserves attention. It decides what kind of attention fits. A visitor who will not buy this quarter still deserves a useful answer and an easy way back when the timing changes.

Many teams treat qualification as a wall: fail the questions and receive nothing. That burns goodwill and quietly discards buyers who would have been ready in six months. We design around four outcomes instead of a pass-or-fail verdict:

• Sales-ready. A real need you serve, a reason to act soon and someone involved in the decision. They get a call with the right rep.
• Promising but early. A good fit whose timing is vague. They get genuinely relevant material and a clear path to talk later.
• Not a fit. Something you do not offer, or somewhere you do not work. They get a straight answer quickly instead of a sales call that wastes both sides’ time.
• Unclear. The bot could not tell. A person reviews the conversation instead of the software guessing.

Criteria built from your closed deals, not from a textbook

Frameworks such as BANT (budget, authority, need, timeline) provide a useful vocabulary, but applied mechanically they produce poor questions. Asking an anonymous visitor for their budget in the first minute is a reliable way to lose them. So we begin with your pipeline history rather than a framework.

Together with whoever leads your sales, we review a sample of recent won and lost opportunities from your CRM and ask two questions. What did the deals you won have in common at first contact? What did the early conversation look like in the ones that went nowhere? The signals that surface are usually specific to your business, for example:

• the kind of problem the prospect describes in their own words;
• organization type or size, where it genuinely affects fit;
• a trigger event, such as a move, a failed system, a new regulation or a contract ending;
• the person’s role in the decision;
• whether the work falls within the services and geography you cover;
• a stated timeframe.

Each signal becomes a written criterion with a definition, a question or an inference drawn from what the visitor already said, and a weight. Criteria that sound important but did not separate won from lost deals get dropped. Every question costs a little of the visitor’s patience, so it has to earn its place.

A points system your reps can read and argue with

We start with an explicit, rule-based score instead of letting a language model produce a number. The model’s job is interpretation: turning free-text answers into defined categories. The scoring itself is ordinary arithmetic your team can audit.

The weights above are examples only; yours come from your own history and change as evidence arrives. Every lead record carries its breakdown, for instance “Scored 7: core service match +3, owner +2, this quarter +2”, so a rep can see why a lead scored the way it did and push back when the reasoning looks wrong.

Hard rules sit alongside the points. Some answers decide the route regardless of total: an existing customer always goes to account management, and out-of-area work always takes the redirect path. Keeping the model away from the arithmetic matters because language models can give different outputs for the same input and cannot reliably explain themselves. Fixed rules mean the same answers always produce the same outcome.

Qualifying questions that do not feel like an interrogation

The difference between a qualification bot that visitors finish and one they abandon is mostly conversation design. The principles we apply:

• Give before taking. Answer the visitor’s own question first. People share more once they have received something useful.
• Infer, then confirm. If someone mentions “our three offices,” the bot does not ask how many locations they have. It confirms and moves on.
• Say why. “So I can connect you with the right specialist, roughly when are you hoping to have this in place?” lands better than a bare question.
• Handle money carefully. Many businesses qualify better on problem and timing and leave budget to the first call. If you do need it, ask late, offer ranges and allow “not sure yet.”
• Stop once the route is clear. When enough signals are in, the bot acts. Extra questions after that point only add friction.
• Let people skip. Every question can be skipped, and a skip is recorded as information rather than treated as a failure.

After the score: what each route actually does

Rep assignment mirrors the rules your CRM already uses, whether that is territory, product line, deal size or simple rotation. During business hours a high-scoring lead can be offered a live handover if a rep is free; outside them, the bot books time instead of leaving the visitor waiting. Early-stage contacts should never be dumped into a generic newsletter. The nurture message relates to what they actually asked about, or it is not sent. Booking and follow-up mechanics beyond that first conversation are covered in our AI chatbot for sales funnels write-up.

The qualification note, not a transcript dump

Nobody on a sales team reads a twenty-message transcript before a call. What reps get is a short structured note attached to the lead, with the full transcript one click away if they want it. A fictional example:

The same facts are written into CRM fields, not just the note text, so you can filter, report and build views on them. The “not discussed” line matters as much as the rest: it tells the rep what to cover first without making them discover the gap mid-call. Connecting those fields properly is covered on our AI chatbot with CRM integration page.

Lines the bot must never cross when judging a lead

• No protected characteristics. The bot does not score on race, religion, sex, age, disability, national origin or similar traits, and it is not allowed to infer them from names, photos or writing style. The model extracts only the signals defined in your criteria document.
• No manufactured urgency. No invented deadlines, fake scarcity or pressure lines, whatever the score.
• No commitments. Prices, availability and outcomes are for a person to discuss. The bot records the question and routes it.
• Honest identity. Visitors are told they are chatting with an automated assistant, and they can ask for a person at any point.
• Minimal data. If an answer would not change the route, the question is not asked.
• Human override. Any rep can re-route a lead, and every override is logged so it feeds the next calibration.

If you sell in a regulated area such as lending, insurance, housing or employment services, rules about how you screen people may apply. We flag that early and ask that your counsel review the criteria before launch; we build software and do not give legal advice.

Checking scores against what actually closed

A scoring model is a hypothesis about your buyers, and it should be tested like one. After launch we compare scores with outcomes recorded in your CRM:

• conversion rate by score band, which should rise with the score if the model is doing its job;
• which individual criteria actually track with closed deals, and which are just noise;
• override rate and override reasons, which reveal where reps disagree with the model;
• completion rate and the question where visitors most often drop out;
• time from a sales-ready conversation to first human contact;
• the share of conversations landing in the unclear route.

Reviews run monthly at first and settle to quarterly once the model is stable. Weight changes go into a change log with the reason, so nobody wonders later why a lead that would once have scored 9 now scores 6. Once enough outcomes accumulate, a statistical model can be layered over the rules, but only with genuinely sufficient history to learn from. Many small businesses never reach that point and do not need to.

Platform lead scoring or criteria built from your pipeline?

Many marketing platforms and chat tools include lead scoring, and they are a reasonable starting point if your entire stack already lives in one of them. Their limits tend to be the same: scoring on generic behavior such as page views and email opens, fixed question templates, little ability to interpret free-text answers, and in some cases a score nobody can explain. A custom build uses your own deal history, your customers’ vocabulary and your routing rules, and it runs on your site with the data flowing into the CRM you already use.

How EVOTECH delivers one, step by step:

1. Pipeline review. We examine won and lost deals with your sales lead to find the signals that matter.
2. Criteria document. Definitions, questions, weights and hard rules, written in plain language for your sign-off.
3. Conversation and routes. The script, the four routes and the rep-assignment logic.
4. Connections. Your CRM, calendar and email tool, via their official APIs.
5. Replay test. Anonymized past inquiries are run through the scoring to see how they would have been routed, and you judge whether that looks right.
6. Supervised launch. Early leads are reviewed by a person before the routes run unattended.
7. Calibration. The regular review cycle described above.

EVOTECH IT LLC is a Houston-area company with more than two decades in business, a 5.0 Google rating and a US-based team serving clients remotely across the country. The first step is a free phone or video consultation; the second is a fixed-scope written quote. If you are still deciding where AI fits in your sales process at all, our AI consulting work is the better place to begin.

Related services

Frequently asked questions