What is the TRAI Third Amendment to TCCCPR in 2026?

The TRAI Third Amendment to TCCCPR was released in draft form on 13 March 2026, building on the 27 February 2026 direction institutionalising AI/ML-based UCC detection at access service providers. Together, the two artefacts authorise telco-side machine-learning models to score the audio content of outbound calls (not just the dialler metadata), require inter-operator intelligence sharing on flagged headers, and signal mandatory AI voice disclosure at the start of every artificial-voice call. The framework moves spam detection from a metadata layer to a content layer, with operational consequences for every Indian AI calling vendor and customer.

What are the AI/ML detection signals under the Third Amendment?

Six signals form the candidate detection space: per-number call velocity (calls per minute from a dialler number), answer-seizure ratio (fraction of calls answered), sub-1-second disconnect rate (fraction of answered calls hung up immediately), complaints per million calls per DLT header, content classification (AI-voice vs human, transcript scoring vs spam vocabulary), and DLT header–content template alignment. The exact thresholds per telco are not public, but observed flag patterns in production deployments suggest soft caps around 40–80 calls/min per number, 15–25% ASR, 25% sub-1-second disconnect rate, and 3 complaints per million calls per header.

Is AI voice disclosure mandatory under TRAI in 2026?

The Third Amendment treats AI-generated voice as "artificial" and signals mandatory disclosure in the first 3–5 seconds of every call. Formal enforcement is being drafted at the ASP layer. The defensive posture for any Indian AI calling deployment is to ship disclosure now — "Hi, I'm an AI assistant from \[brand\] calling about your \[purpose\]" in the borrower's preferred language — with the disclosure scripted as constrained generation (not LLM creative output) so the audio matches the script. Most deployments adding disclosure observe a 2–4 percentage point completion-rate hit that recovers to 0–2 within two weeks of script tuning.

How does inter-operator flag propagation work under TRAI 2026?

When one access service provider (Jio, Airtel, Vi, BSNL) flags a DLT header for AI/ML-detected spam patterns, the flag propagates to other operators through a shared intelligence layer. The defensive posture of "switch operators if flagged" no longer works — a header rate-limited on one telco will not be unrestricted on another. Vendors must implement an inter-operator flag handler that, upon detection of a flag on any one telco, automatically pauses traffic on the flagged header across all telcos within 60 seconds.

What should an Indian buyer ask their AI calling vendor about TRAI Third Amendment readiness?

Seven diligence questions filter the serious vendors: per-number call-velocity dashboard with real-time customer visibility; adaptive pacing logic tied to ASR feedback; AI voice disclosure script in Hindi and customer's regional languages; DLT header–content alignment audit on every campaign; documented runbook for flag-response (pause on all telcos, root-cause, re-enable) with SLA timing; complaint-rate dashboard scoped to customer's DLT headers; audit trail format for the disclosure event including timestamp, language, agent ID and scripted content. Vendors that have direct answers and demo screens for all seven are ready; vendors that deflect or promise "roadmap" are not.

What is the penalty for a TRAI Third Amendment violation in India?

The amendment does not specify a new penalty schedule beyond the existing TCCCPR framework. The operational penalty is operational: a flagged DLT header is rate-limited or suspended across operators, which directly costs the lender or D2C brand the outbound throughput of that header. For a lender running 1 million calls a month through a flagged header, the suspension cost is 100% loss of that channel until re-enabled, plus the compounding cost of borrower-side complaints. The regulator-side penalty is the existing TCCCPR fine schedule plus supervisory action; the operational penalty is the larger of the two for most enterprises.

How long does it take to make an AI calling vendor Third Amendment-ready?

For a vendor starting from a 2024-era pacing baseline, the rebuild fits in eight weeks: weeks 1–2 instrument the six detection signals; week 3 ship the per-number call-velocity governor; week 4 wire adaptive pacing on ASR feedback; week 5 ship constrained-generation disclosure scripts in all production languages; week 6 build the DLT alignment auditor and complaint-rate dashboard; week 7 implement the inter-operator flag handler; week 8 deliver customer-facing readiness pack including dashboards and contractual SLAs. Vendors finishing this before the Q4 2026 likely enforcement window enter the new regime detection-ready.

TRAI Third Amendment 2026 — AI/ML Spam Detection for AI Calling

A founder of an Indian conversational AI startup got the alert at 2:14am IST on a Tuesday in April 2026 — three DLT headers belonging to one of his largest NBFC customers had been "ASP-flagged" by Vodafone-Idea's spam detection system overnight, and 18 percent of the night's outbound queue had been rate-limited. He had no public-facing documentation from any operator describing the flag criteria, no error code that mapped to a specific behaviour, and no SLA from his telephony partner that committed to a resolution time. By 9am, his CTO and his customer's collections head were on a call trying to reverse-engineer what the AI/ML detector had reacted to: answer-seizure ratio, sub-1-second disconnect rate, call-velocity-per-number, complaint rate, content classification, header reuse pattern — six candidate variables, no ground truth, an angry customer, and a regulator that had announced this exact direction six weeks earlier.

This post is for that founder, that CTO, and the heads of operations at every Indian voice AI vendor and lending customer who will hit the same alert in the second half of 2026. The TRAI Third Amendment to TCCCPR introduced on 13 March 2026, plus the February 27 direction institutionalising AI/ML-based UCC detection at access service providers, are the two regulatory artefacts that change how every outbound voice call is judged in India. We will walk through what the Third Amendment actually says at the operational level, what the detection signals look like as engineering parameters, what the AI voice disclosure mandate requires in the first five seconds of every call, how to design call pacing that stays under the flag thresholds, and what to do when your DLT header gets ASP-flagged. By the end, you will have a vendor-side implementation checklist and a customer-side diligence list.

What the Third Amendment changes operationally

Strip out the regulatory language and three things actually change for an AI calling deployment in India.

First, content-level analysis becomes lawful at the ASP layer. The Third Amendment empowers access service providers (the telcos — Jio, Airtel, Vi, BSNL) to run AI/ML models against the audio and content of outbound calls, not just against the dialler metadata. Historically the spam-detection layer was metadata-only — registered header, registered content template, call frequency. Under the new framework the ASP can score the actual audio: is the voice synthetic, does the conversation pattern match a known scam template, does the inferred call purpose match the registered DLT content template, does the post-pickup behaviour (sub-1-second disconnect, immediate hang-up by the customer) suggest annoyance. This is content-level scoring, not metadata-only scoring.

Second, inter-operator intelligence sharing becomes structured. When Vodafone-Idea's detector flags a header, the flag propagates to the other telcos through a shared intelligence layer. A header rate-limited on Vi will not be unlimited on Jio. The defensive posture of "switch operators if flagged" no longer works. For a vendor running outbound at scale, the implication is that the bar must clear across all four operators, not just the most lenient.

Third, AI voice classification triggers a disclosure obligation. The Third Amendment treats AI-generated voice as "artificial" and signals that disclosure at the start of the call is required. Most Indian AI calling deployments today do not disclose. Some open with "Hi, I'm calling from X bank, this is about your EMI" — which omits the AI nature of the voice. The Third Amendment direction is that the first 3–5 seconds must communicate the artificial nature of the voice clearly. Vendors that ship disclosure now will not see a regulatory penalty even if enforcement is delayed; vendors that do not will face a sudden cliff when enforcement starts.

These three operational changes — content-level scoring, inter-operator flag propagation, mandatory AI voice disclosure — are the heart of the Third Amendment from a vendor perspective. The rest of the amendment text is process-level (consultation, drafting authority, definitions) and matters less to engineering than to policy.

The detection signals — what ASPs actually score

The exact AI/ML model used by each telco is not public, but the candidate detection signals are inferable from the regulator's direction text, from telco statements at industry events, and from observed flag patterns in production AI calling deployments through late 2025 and early 2026. Six signals are the high-confidence list.

Call velocity per number. Calls per minute from a single dialler-side number. Anomalously high velocity (above the natural rate for legitimate enterprise outbound) triggers a flag. The threshold is not published but observed flag patterns suggest a soft cap around 40–60 calls per minute per number for legitimate enterprise telephony and a hard cap above 100 calls per minute.

Answer-seizure ratio. Fraction of dialled calls that get answered (lift the receiver) versus those that don't. Legitimate enterprise outbound runs 25–40% ASR. Spam patterns frequently run lower (5–15%) because spam callers churn through pre-filtered databases that are stale. Anomalously low ASR triggers a flag.

Sub-1-second disconnect rate. Fraction of answered calls that the customer disconnects within one second. Legitimate calls retain attention; spam calls get hung up immediately. Above a threshold (observed flag pattern suggests 30–40%) this triggers detection.

Complaint ratio. Direct customer complaints submitted via TRAI's UCC complaint mechanism per thousand calls fired. The threshold is a moving target — 2024 enforcement reportedly used 5 per million calls as a soft signal — but the trend is tightening.

Content classification. AI/ML scoring on the audio: synthetic vs human voice classification, language model classification of the spoken transcript against known spam vocabulary, prosody analysis. This is the new content-level layer the Third Amendment formally authorises.

Header / template reuse pattern. A DLT header registered for a specific purpose used for calls whose content classification falls outside that purpose. A header registered as "loan reminder" fielding calls whose content is "policy renewal" will be flagged at the inter-template level.

These six signals are the candidate detection space. A vendor running enterprise outbound at scale must instrument and manage each of them as production SLOs. The good news is that legitimate enterprise outbound naturally sits well inside the threshold envelope on every signal — the bad news is that one misconfigured campaign can drag the aggregate into the flagging zone within an hour.

What an AI calling vendor must implement — the eight-item checklist

These eight items are what the engineering and operations teams at an Indian AI calling vendor must ship before the Third Amendment enforcement window opens in the second half of 2026.

1. Per-number call-velocity governor. A rate-limiter at the dialler that caps outbound calls per minute per number to a configurable threshold (default 40 calls/min, tunable downward when flag pressure rises). Calls above the cap queue, not fail.

2. Adaptive pacing on ASR feedback. When the answer-seizure ratio on a campaign drops below the campaign's healthy baseline (typically 25–40%), the dialler slows the call-fire rate by 20–40% within 60 seconds. Static pacing is the old game; adaptive pacing is the new game.

3. Sub-1-second disconnect tracking and feedback. Every disconnect within one second of answer is logged with the borrower's number, the campaign, and the time-of-day. Aggregates roll up per campaign and per DLT header in real time. When the rate exceeds 25%, the campaign is paused and re-evaluated.

4. AI voice disclosure script in first 3–5 seconds. Every call opens with "Hi, I'm an AI assistant from [brand] calling about your [purpose]" or the language-appropriate equivalent. The disclosure runs as a constrained-generation block (not LLM-generated freely) so the audio matches the script. Disclosure timestamp logged into the audit trail.

5. DLT header–content template alignment audit. Every campaign cross-checks its content template against the actual conversation content the AI agent will deliver. If the campaign content has drifted from the registered template (common when a script is re-tuned and the DLT template is not re-registered), the campaign blocks until the template is re-aligned.

6. Complaint-rate dashboard at the header level. Track TRAI complaints per million calls per DLT header in near-real-time. When a header crosses the complaint-rate threshold, route new traffic to a different header and investigate the campaign.

7. Inter-operator flag-propagation handler. When one operator flags a header, automatically pause traffic on the same header across all other operators within 60 seconds — assume the flag will propagate, don't wait for it to.

8. Pacing-and-routing rebalancer. When flag pressure rises on one telephony partner (Plivo, Exotel, Knowlarity, Ozonetel, Tata Tele, Twilio), the dialler rebalances new traffic across partners with healthier flag posture, without breaking campaign-to-DLT-header bindings.

These eight items are the vendor-side implementation list. The customer side has a parallel diligence list — which we cover in the next section — but the engineering burden falls on the AI calling vendor.

What a customer must demand at procurement — the seven-item diligence list

The procurement teams at top-30 Indian NBFCs and top-50 D2C brands started adding Third Amendment readiness to RFPs in February 2026. The seven diligence items below are what serious buyers ask.

1. Show me your per-number call-velocity dashboard. Real-time per-number pacing should be visible to the customer's ops team, not just the vendor's.

2. Show me your adaptive pacing logic. Specifically, what happens when ASR on my campaign drops 30%? Vendors that cannot answer this lose serious procurements.

3. What is your AI voice disclosure script, in Hindi and my regional languages? The vendor must show actual scripted disclosure language for every language the deployment will use. Vendors that don't have this default to no disclosure.

4. How does your DLT header–content alignment audit work? Vendors that say "the customer is responsible for DLT template registration" are technically right but operationally weak. Demand that the vendor's system audits the alignment.

5. What happens when my DLT header gets flagged on one telco? The vendor must have a documented runbook: detection, pause on other telcos, root-cause investigation, re-enable. If the answer is "we wait and see", the vendor is not Third Amendment-ready.

6. What is your complaint-rate dashboard, scoped to my DLT headers? The customer must see complaints per header in near-real-time, not in a monthly report.

7. Show me your audit trail format for the disclosure event. Disclosure timestamp, language used, agent ID, scripted content. The audit trail is the artefact a regulator will request, and the customer must be able to produce it independently of the vendor.

Customers that ask these seven questions filter the serious vendors from the not-serious vendors inside the first procurement call. The serious vendors have direct answers and demo screens. The not-serious vendors either deflect or promise "roadmap by Q4."

What goes wrong — five recurring failure modes

Five patterns repeat in production deployments through 2026.

Failure 1: a script update breaks DLT alignment. Ops re-tunes a collections script for tone, the content materially drifts from the registered DLT template, the new content fires for two weeks, complaints rise, the header gets flagged. Fix: every script change triggers a DLT-template re-alignment audit before going to production.

Failure 2: adaptive pacing is configured but not connected to ASR feedback. Vendor's UI shows an adaptive-pacing toggle, but the ASR feedback loop is not implemented — the dialler maintains static velocity even when the campaign is bleeding ASR. Fix: instrument ASR-to-pacing as a closed loop with a sub-60-second adjustment cycle.

Failure 3: AI voice disclosure script is configured but skipped under timing pressure. When the dialler is at peak load, the LLM occasionally trims the first 3 seconds, dropping the disclosure. Fix: constrained generation for the first 5 seconds — the audio for that window must match a fixed script, no LLM creativity allowed.

Failure 4: complaint-rate dashboard exists but does not page anyone. A header crosses the complaint-rate threshold at 2am, the dashboard updates, no one is paged, by 9am the header is ASP-flagged across three operators. Fix: page on-call when complaint rate breaches thresholds; do not rely on humans noticing.

Failure 5: inter-operator flag propagation is ignored. Vi flags a header at 6pm, vendor keeps firing the same header on Jio and Airtel through the night, by morning the header is flagged on all three. Fix: assume propagation, pause traffic on all operators within 60 seconds of the first flag, investigate.

What "good" looks like in the numbers

The vendor-side and customer-side both want to monitor the same six metrics.

Metric	Healthy band	Yellow zone	Red — flag-imminent
Calls/min per number	< 40	40–80	> 80
Answer-seizure ratio	25–40%	15–25%	< 15%
Sub-1-second disconnect rate	< 15%	15–25%	> 25%
Complaints per million calls per header	< 1	1–3	> 3
DLT alignment audit pass rate	100%	95–100%	< 95%
AI voice disclosure presence (first 5s)	100%	90–100%	< 90%

Hit the green band on every metric and the Third Amendment detection layer leaves you alone. Drift into the yellow on any two and you are on borrowed time. Hit red on any one and the flag arrives in days, not weeks.

The asymmetric reality: red on call-velocity gets flagged within an hour; red on complaint-rate gets flagged within a day; red on disclosure presence gets flagged when enforcement starts, which we estimate Q4 2026 to early 2027 based on TRAI's typical drafting-to-enforcement cadence.

Build vs partner — how the vendor question shapes the customer answer

The Third Amendment shifts procurement economics in favour of vendors that have invested in detection-readiness. Three patterns are visible in 2026 procurements.

The first pattern is large customers (top-10 banks, top-30 NBFCs) consolidating to vendors who can demonstrate live dashboards and runbooks for all six detection signals. The cost of vendor switching is high; the cost of being on a non-ready vendor when flag pressure rises is higher. Procurement teams trade off explicit feature lists for detection-readiness.

The second pattern is mid-market customers (mid-tier NBFCs, large D2C brands) demanding contractual SLAs from vendors on time-to-resolution when a flag hits — typical contract clauses now specify 4-hour pause-and-investigate response, 12-hour root-cause, 24-hour re-enable. Vendors that won't sign these clauses lose the deal.

The third pattern is small customers (mid-D2C, small NBFCs) discovering that flag pressure on shared telephony partners affects them even when they themselves are not the source. A DLT header issue at a sister-tenant of the same Plivo or Exotel account can degrade the small customer's deliverability. The small customer now asks the vendor about flag isolation: is my traffic isolated from other customers' patterns?

The strategic posture for an AI calling vendor in 2026 is: detection-readiness is a procurement question, not a technical question. Customers buy from the vendor that has the dashboards, the runbooks and the contractual SLAs. Caller Digital's productisation of the six detection signals as a customer-visible operations layer is what makes the procurement conversation about outcomes rather than features. See the voice AI India compliance stack for the broader regulator picture and the comparison vs Bolna, Exotel and 6 more for vendor positioning.

Implementation timeline — eight weeks from start to detection-ready

For an AI calling vendor starting from a 2024-era pacing baseline, the rebuild fits in eight weeks.

Week 1–2: instrument the six detection signals. Build the per-campaign and per-header telemetry for call velocity, ASR, sub-1-second disconnects, complaints, DLT alignment, and disclosure presence. Dashboards live for ops review.

Week 3: per-number call-velocity governor. Rate-limiter at the dialler. Tunable thresholds. Hard cap at 80 calls/min per number; default soft cap at 40.

Week 4: adaptive pacing on ASR feedback. Closed-loop adjustment with sub-60-second cycle.

Week 5: constrained-generation disclosure scripts. Per-language opening templates. Audio-timestamp logging. Verify across all production languages.

Week 6: DLT alignment auditor + complaint-rate dashboard. Block campaigns when alignment audit fails. Page on-call when complaint rate breaches thresholds.

Week 7: inter-operator flag handler. Detection, pause, investigate, re-enable runbook. Wire to all four telephony partners.

Week 8: customer-facing readiness pack. Dashboards visible to customer ops. Runbook document for incidents. Contractual SLA template. Ready for procurement diligence.

Vendors that finish this in 8 weeks are detection-ready before the enforcement window. Vendors that don't are surviving on inertia until the first flag.

What changes after the Third Amendment fully enforces

Three forward signals over the next 12 months.

The ASP-side AI/ML detection model will get better. The 2026 model is a first generation; the 2027 model will incorporate the inter-operator intelligence shared layer and will close some of the false-positive gaps. Vendors that survive 2026 with healthy detection metrics get the breathing room to focus on use-case quality rather than detection-defence.

The customer-side demand for proprietary instrumentation will rise. Top-10 banks will eventually demand pacing telemetry exposed via their own observability stacks (Datadog, Grafana). Vendor APIs for telemetry export become a procurement requirement by mid-2027.

The disclosure pattern will standardise. By Q2 2027 we expect a TRAI-recommended template for AI voice disclosure language — at which point the disclosure becomes a checkbox rather than a creative choice. Vendors that ship disclosure flexibility ("we support any language and tone") will continue to win procurements; vendors that hardcode a 2026 template will need to refactor.

Bottom line

The TRAI Third Amendment to TCCCPR (released 13 March 2026) and the February 27 direction on AI/ML-based UCC detection together reshape the operational posture of every Indian AI calling vendor and customer. Three operational shifts matter: content-level analysis at the ASP layer, inter-operator flag propagation, and mandatory AI voice disclosure. Six detection signals — call velocity per number, answer-seizure ratio, sub-1-second disconnect rate, complaint rate, content classification and DLT alignment — define the new operational envelope. Eight vendor-side engineering items implement the response: per-number rate-limiting, adaptive pacing, sub-1-second disconnect tracking, AI voice disclosure scripts, DLT alignment audit, complaint-rate dashboards, inter-operator flag handlers and pacing-and-routing rebalancers. Seven customer-side procurement questions filter the serious vendors from the not-serious. The eight-week implementation timeline fits between today and the Q4 2026 likely enforcement window. The vendors and customers that ship this list now will not see the 2am alert in the middle of the night; the ones that don't will.

For the broader multi-regulator picture, see the India voice AI compliance stack 2026. For the RBI-side parallel work, see RBI draft recovery norms effective 1 July 2026. For the production AI calling platform context, see the AI Caller India pillar and voice AI India 2026. For telephony partner context — Plivo, Exotel, Knowlarity, Ozonetel, Tata Tele, Twilio — see telephony integrations.