AI lead qualification: score every incoming lead so your salespeople call the right ones first

What is AI lead qualification?

AI lead qualification is the automatic sorting of new enquiries by how likely they are to buy, done by a language model that reads each lead, asks what is missing and gives it a score with a reason. It replaces the first five minutes a salesperson spends deciding whether an enquiry deserves a call.

The older version of this idea is lead scoring with points: plus ten for a company email, plus five for choosing “within 30 days”, minus twenty for a student. Points still work when every lead fills the same structured form. They break when leads arrive as a WhatsApp voice note, a portal message saying “rate?”, or a form where the useful detail sits in the comments box. A large language model can read that free text, pull out budget, location, product and urgency, and judge it against a rubric you write in plain English or Hindi.

So the working definition we use with clients is simple. AI lead qualification has four jobs: collect the lead from every source, complete it by asking two or three questions, score it against your ideal customer, and route it to the right person or the right automated reply. Anything that does fewer than those four jobs is a chatbot or a report, not qualification.

• Collect: forms, WhatsApp, web chat, IndiaMART, property portals, Meta lead ads
• Complete: short qualifying questions by chat, stored on the lead
• Score: hot, warm or cold, with a one-line reason a manager can audit
• Route: salesperson alert, nurture sequence or polite decline

When does a business actually need AI lead qualification?

You need AI lead qualification when salespeople spend more time sorting leads than talking to buyers, or when good leads wait hours because they sit in a pile with junk. If your team can answer every enquiry within ten minutes already, you probably do not need it yet.

The signs are usually visible in a week of CRM data. Salespeople open leads and close them as “not reachable” or “not interested” at a high rate. The best closer spends mornings calling students, job seekers and vendors who filled the enquiry form by mistake. Portal leads arrive in bursts that nobody can cover. Managers argue about which source is “good” with no data. Or leads come in Hindi, English and a mix of both, and your rules only understand one.

Industries in India where we see the pattern most often: real-estate projects buying portal and Meta leads, coaching institutes and colleges during admission season, B2B manufacturers on IndiaMART, solar installers, interior designers, clinics running paid campaigns, and insurance or loan advisers. Each gets volume that looks healthy on a dashboard but converts poorly because the call order is random.

How does AI lead qualification work, step by step?

A lead enters, gets cleaned, gets questioned if details are missing, gets scored by a model with your rubric, and gets routed, all within a minute or two. The salesperson only sees the result: a ranked list with reasons.

Under the hood, each source sends the lead to one intake endpoint through a webhook or API. The intake step normalises fields, so “Mob No”, “phone” and “WhatsApp” all become one phone field in the same format. A filter step checks for bots, invalid numbers, repeat submissions and junk text. If key facts such as budget or timeline are missing, the WhatsApp bot asks them. Then the scorer, usually a hosted large language model called through its API, receives the lead plus your rubric and returns structured JSON: score, band, extracted fields and reason. The router reads that JSON and acts.

• Intake: webhook from the form, WhatsApp Business Platform, portal or ad platform
• Normalise: one phone format, one city spelling, one source label
• Filter: bot score, phone and email validity, duplicates, junk text
• Complete: ask missing questions by chat and wait a set time for replies
• Score: model returns band, score, extracted details and a reason
• Route: CRM owner, alert on WhatsApp or email, or automated reply
• Log: every decision stored so you can audit and retrain the rubric

The intake and routing half of this pipeline is the same work described on website CRM integration; qualification adds the scoring brain on top.

What qualifying questions should an AI lead qualification bot ask?

Ask the fewest questions that change what the salesperson does next, usually two to four. Each extra question loses some people, so every question must earn its place.

A good test: if the answer would not change who calls, how fast, or what they say, cut the question. For a real-estate project, budget band, preferred configuration and purchase timeline change everything; the buyer’s profession rarely does. For a B2B manufacturer, quantity, delivery location and whether the person is the end buyer or a trader matter most. For a coaching institute, the course, the exam year and the city decide the counsellor.

How the bot asks matters as much as what it asks. On WhatsApp we prefer reply buttons or short lists over open typing, because a tap is easier on a budget phone than a typed answer. The first message thanks the person by name, says who is writing, and asks one question. The next question only appears after a reply. If the person types something unexpected, such as “call me”, the model recognises the intent and marks the lead hot instead of repeating the script.

How does LLM scoring against your ideal customer work?

LLM scoring means giving a language model a written description of your ideal customer, your disqualifiers and a few scored examples, then asking it to judge each lead the same way. The quality of the rubric decides the quality of the scores.

We write the rubric with your sales head in a one-hour call. It covers who buys (budget range, location served, size of business, product fit), who never buys (students, job seekers, vendors, areas you do not serve), and what signals urgency (a date, a festival, a funding event, “this week”). Then we pull 50 to 100 past leads with known outcomes from your CRM and check whether the model’s bands match what actually happened. Where it disagrees, we either fix the rubric or learn that your team’s instinct was wrong.

Three safeguards keep AI lead qualification honest. The model must return structured output, not prose, so the router cannot misread it. Every score carries a reason in one sentence, so a manager can spot nonsense. And the model never sees data it does not need; a lead’s full chat history, for example, is trimmed to the relevant lines. Scores are advice to your team, not a verdict: a salesperson can override any band, and those overrides become training examples for the next rubric revision.

Rules, machine learning or LLM: which lead scoring method fits your data?

Use rules when every lead comes through one clean form, machine learning when you have thousands of past leads with recorded outcomes, and an LLM when leads are messy, multilingual or few. Many good systems combine all three.

Rules are cheap, transparent and fast. They fail on free text and drift silently as your market changes. A classic machine learning model, trained on your CRM history, can be very accurate, but it needs a large, honest dataset where every lead is marked won or lost, and many small Indian businesses simply do not have that. An LLM needs almost no history to start, reads any language your team reads, and explains itself, but each call costs money and it can be confidently wrong if the rubric is vague.

Our usual design for AI lead qualification: hard rules first (spam, out-of-area, duplicates), then the LLM for judgement, then, after six months of outcomes, an optional statistical check that compares model bands with real wins. Another of us, who handles the AI and data side of our team, sets up that comparison so the business can see whether scores predict revenue or just look clever.

How does AI lead qualification filter spam and fake leads?

Filter in layers, cheapest first: stop bots at the form, reject impossible data, merge duplicates, and only then let the model judge what is left. Paying an AI model to read obvious junk wastes money.

At the form, Google’s reCAPTCHA v3 returns a score between 0.0 and 1.0 for each interaction, where 1.0 is very likely a good interaction and 0.0 very likely a bot; Google’s documentation suggests starting with a threshold of 0.5 and adjusting from your own traffic. A hidden honeypot field catches simple scripts. Next come data checks: a mobile number with the wrong length or an obvious pattern, a disposable email domain, a name field containing a URL. Then duplicates: the same phone in the last few days is merged into the existing lead instead of creating a new one, so two salespeople do not call the same person.

The model handles the subtle cases. Competitors filling your form to waste your time, job applicants using the enquiry form, vendors pitching SEO services, and bored visitors typing “test” all have recognisable language. The model labels these “not a buyer” with a reason, and they skip the sales queue entirely. For paid campaigns, we also log the source and campaign on every rejected lead, because a campaign producing mostly junk is a spending decision, not a sales problem.

• Bot layer: reCAPTCHA v3 score, honeypot field, rate limits per IP
• Data layer: phone format, email domain, blank or nonsense fields
• Duplicate layer: same phone or email within a set window
• Language layer: model flags vendors, job seekers, competitors and tests
• Report layer: junk share by source and campaign, reviewed weekly

Qualifying IndiaMART, property portal and Meta lead ads leads

Portal and ad leads need AI lead qualification the most, because they arrive with the least context and the highest volume. The fix is to pull them into the same scorer as your website leads, within minutes, instead of downloading spreadsheets.

Meta’s lead ads documentation describes real-time retrieval: your app subscribes to webhooks, receives a leadgen_id when a lead is created, and fetches the lead’s answers through the Graph API. That means a Facebook or Instagram lead can get a WhatsApp question within a minute of submitting the ad form. IndiaMART offers a lead API for paid sellers, and several property portals share leads by email or API depending on your plan; where only email exists, we parse the email reliably and send it on.

Portal leads bring their own quirks. Buyers often enquire with many sellers at once, so speed matters more than anywhere else. Some portal messages are just a product name. Ad leads frequently carry autofilled details the person never checked. The rubric should therefore weigh replies to the WhatsApp question more heavily than the ad form data. Our pages on Facebook lead ads integration and IndiaMART CRM integration cover the plumbing for each source in more depth.

How should qualified leads be routed in the CRM?

Route by band first, then by skill or territory, and always alert a human for hot leads. A score that sits in a CRM field nobody looks at changes nothing.

A typical routing table looks like this. Hot leads get assigned to the right salesperson by city, product or language, with an instant WhatsApp alert that includes the reason line and a tap-to-call link. Warm leads go into a nurture flow: a helpful message, a brochure or price guide, and a reminder task for the salesperson in two days. Cold leads receive a polite automated reply and stay in the CRM, because some cold leads warm up later. Rejected spam is stored separately for reporting and never assigned.

We build routing in whichever CRM you already use, such as Zoho CRM, HubSpot or LeadSquared, or in a custom pipeline, and keep round-robin logic fair by tracking assignments rather than guessing. If a hot lead is not opened within a set time, the system reassigns it or alerts the manager. That escalation rule is often worth more than the scoring itself, because a hot lead left unanswered for a day is usually lost to the next seller.

Using Zoho already? Our Zoho CRM implementation page explains assignment rules and blueprint set-up that pair well with scoring.

How much does AI lead qualification cost in India?

A custom AI lead qualification build from BtechWaleTech starts at ₹40,000 (US$600), and running costs are paid directly by you to the AI model provider and to Meta for WhatsApp template messages. What pushes the build price up is the number of sources, the depth of the chat flow and the routing logic.

Running costs depend on volume and model choice. Each lead costs one or a few model calls; smaller, faster models cost far less per lead than the largest ones and are usually enough for scoring. Meta’s WhatsApp pricing page states that the platform moved to per-message pricing from 1 July 2025, that you are charged when a template message is delivered, and that non-template messages inside an open customer service window are free. Because a lead who fills your form has not messaged you yet, the first WhatsApp question is usually a template; the replies that follow happen inside the window.

Quotes from other freelancers and vendors vary widely for what sounds like the same project. The difference usually comes from whether routing, spam filtering and a lift dashboard are included, whether the vendor charges monthly platform fees on top, and who owns the code and prompts afterwards. Compare scope line by line. For wider numbers across AI projects, see AI automation cost for small businesses and AI agent development cost.

How do you measure whether AI lead qualification improved conversions?

Measure with a before-and-after baseline and, ideally, a holdout group: a random slice of leads handled the old way while the rest go through AI lead qualification. Compare response time, contact rate and deal rate between the two, not just the overall trend.

Start by recording the baseline for four weeks before launch: average time to first contact, share of leads reached, share marked junk, and deals per hundred leads by source. After launch, keep maybe one lead in ten outside the system for a month. If the scored group converts better and faster, you have evidence; if not, you have learned something cheaply. Also check the model’s calibration: do hot leads really close more often than warm ones? If bands do not separate outcomes, the rubric needs work.

Tie the loop back to marketing. GA4’s enhanced measurement can record form_start and form_submit events, which shows drop-off before a lead exists. Google Ads documentation describes offline conversion import, where you store the Google Click ID (GCLID) with the lead and later upload it when the lead converts; Google now recommends enhanced conversions for leads for new set-ups. Sending “qualified” and “won” back to the ad platform lets campaigns optimise for buyers instead of form fills. Our conversion tracking setup page covers that side.

Data privacy, consent and the DPDP Act in lead scoring

Collect only what you need to qualify, tell people why, and keep the data inside accounts you control. That is good practice and it lines up with India’s Digital Personal Data Protection Act, 2023, which PRS Legislative Research summarises as requiring a notice about the personal data collected and its purpose before consent, with the right to withdraw consent.

In an AI lead qualification build, that translates into concrete choices. The form carries a short consent line and a link to your privacy policy. The WhatsApp bot identifies your business in the first message. The model receives the lead’s answers, not their whole CRM history or unrelated documents. API keys sit in your cloud account, logs have a retention period you choose, and access to the lead database is limited by role. If a person asks to be removed, the removal covers the CRM, the logs and any nurture lists.

We build these controls; we do not give legal advice. PRS notes penalties under the Act reaching up to ₹250 crore for failing to take security safeguards, so please have your own lawyer review the consent wording and retention plan. For regulated sectors such as lending or insurance, sector rules on calling and data use also apply, and your compliance team should sign off the scripts before launch.

AI lead qualification for Hindi, Hinglish and regional-language leads

Modern language models read Hindi, Hinglish and major Indian languages well enough to extract budget, place and timing, so one AI lead qualification rubric can cover mixed-language leads. The weak point is usually the replies the bot sends, which should be written and approved by you, not improvised.

Indian leads rarely stick to one language. “Sir 2 bhk chahiye Wakad ke paas, budget 60 tak” is a perfectly clear real-estate lead to a human and to a good model. We test the scorer with real, anonymised examples from your own inbox, including spelling variations and Roman-script Hindi, before launch. For outgoing messages, we write the English and Hindi versions with your team and let the bot pick the language the lead used.

For Tamil, Telugu, Marathi, Bengali, Kannada, Malayalam, Gujarati or Punjabi replies, the approach is the same: you supply or approve the wording, and the bot sticks to it. Voice notes on WhatsApp can be transcribed first; Google Cloud Speech-to-Text, for example, lists Indian locales including Hindi, Tamil, Telugu, Bengali, Marathi, Gujarati, Kannada, Malayalam and Punjabi. For a fully Hindi conversational bot, see Hindi AI chatbot.

Risks and red flags when buying an AI lead qualification system

The biggest risk is a black box: a vendor’s tool that scores leads without showing why, on servers you do not control, with a monthly fee you cannot escape. The second is over-automation, where the bot argues with buyers instead of handing them to a person.

Watch for these warning signs when you evaluate anyone, including us. Promises of a fixed percentage increase in sales before they have seen your data. No mention of a baseline or holdout test. Scores with no stored reason. Prompts and code kept on the vendor’s side. A bot that cannot hand over to a human. WhatsApp messages sent from an unofficial tool rather than the WhatsApp Business Platform, which can get your number restricted. And any plan to buy lead lists or scrape contacts, which creates legal and reputational risk.

There are also honest technical limits. Language models make mistakes, especially on short or ambiguous messages. Rubrics drift as your offers and prices change. Seasonal spikes, such as admission season or festive property launches, change lead mix quickly. Plan for a monthly review in the first quarter, where someone from sales and someone from our team look at twenty random scored leads together and adjust the rubric.

• Ask where prompts, code and logs live, and who can export them
• Ask to see the reason text for a sample of scored leads
• Ask how a buyer reaches a human in one step
• Ask what the monthly running cost looks like at your volume

Worked example: AI lead qualification for a hypothetical Pune housing project

Say a mid-size builder in Pune runs Meta ads and two property portals for a new project, and three sales executives share every lead from a spreadsheet. This is an illustration of how we would design it, not a client story.

First, the intake: Meta lead ads through webhooks, portal leads through their email or API feed, and the website form directly. Every lead is normalised, checked with reCAPTCHA where applicable, and merged if the phone already exists. Second, the WhatsApp template asks one question with buttons: “Which home are you looking for? 1 BHK / 2 BHK / 3 BHK / Just exploring”. A second question asks timing: “This month / Within 3 months / Later”. Third, the model scores with a rubric that says: serious if configuration and timing are given and the budget in the ad form fits the project; cold if “just exploring” and no timing; not a buyer if the text mentions jobs, broker tie-ups or loan agent pitches.

Routing then sends hot Marathi-speaking leads to the executive who speaks Marathi, others round-robin, with a WhatsApp alert and a two-hour escalation to the sales manager. Warm leads get the floor-plan PDF and a site-visit booking link. After a month, the dashboard shows contact rate and site visits by band and source, and the builder can decide which portal or campaign deserves more budget. For a property-specific bot, compare chatbot for real estate.

AI lead qualification checklist before you start

Have these ready before the kickoff call and an AI lead qualification build can start the same week. Missing items do not stop the project, but they slow the tuning.

• A list of every lead source, with rough daily volume for each
• Admin access to your CRM, or a decision to start with a Google Sheet
• Your WhatsApp Business Platform number, or a plan to set one up
• A written description of your best customer and three kinds of lead you never want
• 50 to 100 past leads with known outcomes, anonymised if you prefer
• Who gets hot leads, by city, product or language, and the escalation time
• Approved wording for the first WhatsApp message in each language
• Consent line and privacy policy link for forms
• A baseline: current response time and conversion by source
• One person on your side who can answer questions within a day

Ownership is part of the checklist too. We create the AI model account, the WhatsApp app and the cloud resources under your business, share the code in your repository and write a short handover note explaining the rubric and how to edit it. If you stop working with us, the system keeps running.

Not sure your leads justify AI yet? Our comparison of AI automation vs hiring staff helps you decide.

AI lead qualification across India

We build AI lead qualification remotely for businesses in every state, over WhatsApp and video calls, so your city does not change the price or the process. What changes is the lead mix: real-estate enquiries dominate in Pune, Gurgaon and Hyderabad, admission leads in Jaipur and Patna, IndiaMART trade enquiries in Ludhiana, Rajkot and Coimbatore, and clinic or service leads in Kochi and Indore.

Language is the other big variable. A Chennai business may want Tamil replies for some leads, a Kolkata one Bengali, and a Lucknow one pure Hindi. The scoring rubric stays the same; the outgoing wording is what you approve per language. Payments are by UPI or bank transfer with a GST-ready invoice request handled in your written quote, and the whole project runs on one WhatsApp group with the three of us.