If you have been pitched an AI qualification layer in the last year, you have probably been shown a version of this number: it removes 40 to 60% of unqualified conversations.

We went looking for where that figure comes from. It comes from conversational software vendors, and nowhere else. There is no independent research behind it, and more importantly it measures the wrong thing — it counts conversations a human did not have, which is a cost metric, not revenue. Our performance work now starts by refusing that number, because the framing behind it is what breaks these systems.

Here is what the evidence actually supports, what the two Indian regulatory regimes require, and why the goal is not filtering.

What is the Silent Reject?

Every filter makes two kinds of mistake. It lets through leads that were never going to buy, and it turns away leads that would have.

The first mistake is visible. Your telecaller wastes twenty minutes and tells you about it.

The second is invisible, and that is the entire problem. A lead your system deprioritised or dismissed leaves no record anywhere. It does not appear in your CRM as a lost opportunity, because it never became an opportunity. Nobody calls it back to find out what would have happened. Your filter's accuracy is therefore unmeasurable in practice, and every vendor dashboard reports only the half that flatters the system.

Call it the Silent Reject, and do the arithmetic on it. In a high-ticket category — real estate, a premium programme, enterprise software, a capital purchase — one wrongly rejected buyer can cost more than a year of saved telecaller time. There is no published research quantifying false negatives in automated lead qualification, which means nobody selling you a filter can tell you its error rate against actual closed business.

That is not an argument against automation. It is an argument against measuring it as a filter.

Is faster response actually worth it? What the research really says

The speed claim is the most-quoted and least-examined number in this category, so it is worth getting right.

The source is a 2007 Lead Response Management study by Dave Elkington, then chief executive of InsideSales.com, with James Oldroyd of MIT Sloan. It analysed three years of data from six companies — more than 15,000 web leads and over 100,000 call attempts. The study is still available, and its findings were:

• The odds of contacting a lead drop by 100 times between a five-minute and a thirty-minute response
• The odds of qualifying a lead drop by 21 times across the same gap
• Waiting an hour cuts contact odds more than tenfold

Three qualifications that almost never travel with the statistic. The study explicitly states it did not address close ratios — it measured contact and qualification, not revenue. It is from 2007, which predates smartphones, WhatsApp and instant calendar booking. And it was co-authored by the chief executive of a company selling lead-response software.

So treat it as strong evidence that speed drives contact, which is genuinely useful, and not as proof that a five-minute response multiplies your revenue by a hundred. The honest version of the claim is narrower and still worth acting on.

Does automating the conversation help or hurt?

Here the evidence is sharper, and it cuts against interrogation.

Luo, Tong, Fang and Qu published a field experiment in Marketing Science in 2019 (citation here) covering roughly 6,000 customers of a financial services company, randomly assigned to either a human agent or a chatbot. INFORMS summarised the findings. Two results matter:

Undisclosed chatbots were about four times more effective at selling than inexperienced human staff. Note the comparison group — inexperienced staff, not your best closer.

Disclosing that the agent was a chatbot reduced purchase rates by more than 79.7%. The researchers attribute this to customers perceiving the bot as less knowledgeable and less empathetic, becoming curt, and disengaging.

Some commentary reads that second finding as an argument for concealing automation. It is not, and we would not advise it. In India, advertising standards and data protection notice requirements both push toward disclosure, and a business caught hiding it has a worse problem than a conversion dip. The study itself varied disclosure across four timings — none, at the start, after the conversation, and after purchase — which points at the real lesson: the penalty is about establishing value before identity, not about deception being profitable.

Read with the false-negative problem, it gives you a clear design rule. An automated first touch that opens with a questionnaire is the worst of both worlds: it discloses nothing useful, demands something, and drives exactly the disengagement the study measured.

So what should the system actually do? Route, don't interrogate

Three jobs, in order of value.

1. Answer instantly, usefully. The enquiry gets a real response in seconds — the thing they asked about, the price band, the next available slot, the document they need. Not "thanks, our team will reach out." You are buying the contact-odds effect the 2007 study measured, without asking the buyer to earn it.

2. Route the strongest enquiries to a human fast. Use the signals you already have and did not have to ask for: which page they came from, what they typed, which campaign, which product, time of day, whether they returned. These are observed rather than interrogated. The highest-signal enquiries go to a person within minutes, by whatever channel the buyer used.

3. Sequence everything else. Not reject — sequence. A useful follow-up, a document, a callback window, a slot booking. Leads you cannot serve now get something rather than nothing, and they stay measurable, which is precisely what a filter destroys.

Where qualification questions do belong: after value has been delivered, framed as service rather than screening. "Which of these three programmes are you comparing?" is a useful question. "What is your budget?" at first contact, from a bot, at 11pm, is how you lose an enquiry in a market where the relationship is the sale.

Which brings us to the framework problem. BANT — budget, authority, need, timeline — originated as an IBM sales convention, not as validated research, and there is no body of evidence showing that having a machine interrogate a buyer against it at first contact increases closed revenue. It was designed for American enterprise selling with a procurement process at the other end. Importing it wholesale into an Indian high-ticket sale is a cultural mismatch dressed as methodology.

What does Indian law require? Two separate regimes

This is where most implementations are exposed, because vendors treat it as one problem and it is two.

Track one: data protection

The Digital Personal Data Protection Act, 2023 governs the capture. Its Rules were notified on 14 November 2025, with phased compliance over an eighteen-month window — the Government's own explainer sets out the notice and consent requirements.

Two consequences for a qualification system:

Consent is purpose-limited. Consent to be contacted about a specific enquiry is not consent to be entered into an unrelated marketing sequence, and arguably not consent to have that data train a scoring model. If your system does both, that needs separate, specific consent rather than a single tick box.

Withdrawal has to actually work. Someone who opts out must be removable — from your CRM and from whatever downstream system holds their data. Build that before you need it.

Track two: commercial communication

Entirely separate, and this is the one with teeth. TRAI's TCCCPR framework governs calls and messages, and the amendments of 12 February 2025 sharpened it considerably. Storyboard18's summary sets out what changed, and the regulation itself is published:

• First violation by a sender: outgoing services barred across all its telecom resources for 15 days. Repeat violations: disconnection of all telecom resources including PRI and SIP trunks, across all access providers, for a year, plus blacklisting.
• The trigger is five complaints in ten days — tightened from ten complaints in seven days.
• Customers can now report spam from unregistered senders without having registered a preference first, and have seven days to do so rather than three.
• Header suffixes distinguish promotional (-P), service (-S), transactional (-T) and government (-G) messages.
• Ordinary ten-digit numbers are barred for telemarketing. Promotional calls use the 140 series; the 1600 series is for transactional and service calls.
• Promotional messages must carry an opt-out, and after someone opts out you cannot request consent again for 90 days.

Read the five-complaints-in-ten-days trigger against an automated follow-up sequence running at volume. That threshold is easy to hit with a mis-targeted list, and the consequence is your outbound capacity switched off for a fortnight. TRAI's own advice to senders is worth reading before you turn anything on.

The design consequence: an automated system cannot freely generate and send novel message text. Promotional communication runs on registered headers and pre-approved templates with variable fields. A language model writing bespoke follow-ups and pushing them through a messaging channel is not a clever growth tactic; it is the thing the penalty regime exists to stop.

What breaks these systems in practice

Four failure points, all organisational rather than technical.

Handover latency. The system qualifies a lead in forty seconds and then it sits in a queue for six hours. You have automated the fast part of a slow process and changed nothing. Measure time from enquiry to human contact, not time to qualification.

Sales not trusting the score. If your team believes the model is wrong, they will call everyone anyway, and you have added cost without removing work. This is the most common quiet failure and it is a trust problem, not a model problem.

No closed loop. A scoring system learns from outcomes. If your CRM does not reliably record what happened to each enquiry — won, lost, why, at what value — the model has nothing to learn from and will not improve past the rules you wrote on day one. There is no published data on CRM hygiene in Indian mid-market businesses, so audit your own: take last quarter's enquiries and see how many have a recorded outcome. If it is under half, fix that before buying anything.

Language. Indian business conversations code-mix constantly, and vendor engineering guidance — including from Indian providers building for this specifically — is candid that speech models trained on clean monolingual audio degrade badly on Hinglish and regional accents over ordinary telephony, and that latency above a few hundred milliseconds breaks natural turn-taking. Those sources sell Indian voice AI, so weigh the interest, but the direction is consistent and matches what anyone who has tested a global voice agent on an Indian call has heard.

When not to buy this

Four situations where we would tell you to spend elsewhere.

Your lead quality problem is upstream. If most enquiries are junk, the fault is in targeting, channel mix or offer, and filtering downstream treats the symptom while you pay a subscription. Fix the source. This is the mechanism behind most lead generation programmes that fail.

Your enquiry volume is low and your ticket is high. If you get forty enquiries a month and each sale is worth lakhs, the expected cost of one Silent Reject swamps any efficiency gain. Call all forty. Hire a good person to do it well.

Your real constraint is response coverage. If enquiries arrive at 9pm and nobody is on until 10am, the fix may be shift coverage or an additional inside-sales hire rather than a qualification layer. Price both honestly — at Indian inside-sales salary levels that comparison is genuinely live, and sometimes the hire wins.

You cannot produce closed-loop data. Without it you are buying a system that cannot improve. Six weeks of CRM discipline is the cheaper purchase and it makes everything after it work better. That is the general point behind getting measurement right first and behind the analytics layer more broadly.

How we would build it

1. Instrument first. Every enquiry tagged by source and campaign, every outcome recorded. Two to six weeks, and it is non-negotiable.
2. Instant useful response on the channel the enquiry arrived on. Answer the question; do not ask one.
3. Observed-signal routing using data you did not have to request. Best enquiries to a human within minutes.
4. Sequencing, not rejection, for everyone else — with something genuinely useful in each step.
5. Compliance architecture before launch: DPDP notice and purpose-specific consent at capture, registered headers and approved templates for anything promotional, a working withdrawal path.
6. Measure human-contact latency, qualified-pipeline value and conversion — never deflection rate. If a dashboard leads with conversations avoided, it is measuring your cost, not your revenue.
7. Audit the rejects quarterly. Pull a sample of what the system deprioritised and have a human call them — a step we build into every engagement of this kind. This is the only way to see the Silent Reject at all, and nobody does it.

That last step is the one we would insist on, because it is the only honest check on whether the thing works.

Want this assessed against your own pipeline? Send us your monthly enquiry volume by source, your current response time, your close rate and what share of last quarter's enquiries have a recorded outcome in your CRM, and we will come back with whether automation is justified at your volume, what your DPDP and TRAI exposure looks like, and what we would instrument before anything else. If the answer is that you need a hire or a cleaner CRM rather than a platform, we will say so. Talk to Midgrow, or message +91 9632398896 with your sector and enquiry volume. We work on this with education and real estate businesses in particular, where enquiry volume is high and the ticket makes a wrong rejection expensive. Pricing is in what agencies do not tell you, and how we think about AI generally is in our approach to AI-powered marketing.

Frequently asked questions

Does AI lead qualification really remove 40 to 60% of unqualified leads?

That figure comes from conversational software vendors and has no independent research behind it. It is also a deflection metric — it counts conversations your team did not have, which is a cost saving, not revenue. It says nothing about leads the system wrongly rejected, because a rejected lead leaves no record in your CRM at all.

Is the five-minute lead response rule true?

Partly. The 2007 Lead Response Management study by Elkington and Oldroyd, covering over 15,000 leads and 100,000 call attempts across six companies, found the odds of contacting a lead drop 100-fold and of qualifying one 21-fold between a five-minute and thirty-minute response. But it explicitly did not measure close ratios, it predates smartphones and WhatsApp, and it was co-authored by the chief executive of a lead-response software company. Speed drives contact; it is not a proven revenue multiplier.

Should a chatbot disclose that it is AI?

Yes. A 2019 Marketing Science field experiment across roughly 6,000 customers found disclosure reduced purchase rates by more than 79.7%, which some read as a case for hiding it. That is the wrong conclusion: Indian advertising standards and data protection notice requirements both point to disclosure, and the study's design suggests the penalty is about establishing value before identity rather than about concealment paying. Disclose, lead with something useful, and offer one-click access to a human.

What does Indian law require for automated lead qualification?

Two separate regimes. The Digital Personal Data Protection Act, 2023, with Rules notified on 14 November 2025 and phased over eighteen months, governs capture — requiring notice and purpose-limited consent, so enquiry consent does not automatically cover later marketing or model training. TRAI's TCCCPR framework separately governs calls and messages, requiring registered sender headers, pre-approved templates and preference scrubbing.

What are the penalties for automated spam messaging in India?

Under TRAI's February 2025 amendments, a sender's first violation means outgoing services barred across all its telecom resources for 15 days; repeat violations mean disconnection of all telecom resources including PRI and SIP trunks, across all access providers, for a year, plus blacklisting. Action is triggered by five complaints in ten days.

Can an AI write and send its own follow-up messages?

Not for promotional communication in India. Messages must run on registered headers and pre-approved content templates with variable fields, so a model generating bespoke text and pushing it through a messaging channel is exactly what the penalty regime targets. Service and transactional messages are categorised differently, which is why the categorisation matters as much as the content.

Is BANT a good basis for an AI qualification script?

BANT originated as an IBM sales convention, not as validated research, and no credible evidence shows that having a machine interrogate a buyer against it at first contact increases closed revenue. It was built for American enterprise procurement. Asking budget and authority questions at first touch in an Indian high-ticket sale is more likely to end the conversation than to score it.

What should I measure instead of deflection rate?

Time from enquiry to human contact, qualified pipeline value, and conversion by source. Then, quarterly, pull a sample of enquiries the system deprioritised and have a person call them — that is the only way to detect false negatives, and it is the check almost nobody runs.