Marketing automation does not usually break with an error message. It breaks by continuing to report success.
Your platform shows the email as delivered. Google has routed it to spam. Your system logs a successful API call. The message is sitting in a queue and will arrive three hours after the webinar ended. Your CRM confirms the record was created. It is the fourth duplicate of the same person. Every dashboard is green.
That gap — between a success response and a successful outcome — is where automation money goes, and it is what we design around when we build these systems as part of our performance work. Below are the specific thresholds, error codes and signals involved, because this is one of the few areas of marketing where the numbers are published and exact.
Why do marketing automations break at scale?
Three causes, and a property they share.
Bad data entering the system. Duplicates, unnormalised phone numbers, invalid addresses, form spam.
Unhandled edge cases in branching logic. A contact in two sequences at once. A contact re-entering a sequence it finished. An automation triggering itself. A retry sending the same message twice.
Undefined handover points. A lead routed to nobody, or to someone who left. An automated sequence still running after a salesperson started a conversation.
The property they share is that the expensive version of each is silent. A loud failure gives you an error code you can alert on. A silent failure gives you a success response and a wrong outcome, and you find out from a customer. Design for the silent ones first.
The Silent Success Problem
Here is the pattern, stated once so the rest of the article makes sense.
Every layer in your stack reports on what it did, not on what happened. Your email platform reports that it handed the message to an SMTP relay. Your messaging API reports that it accepted your request. Your CRM reports that it wrote a record. None of them report whether the message reached a human, arrived in time, or landed on the right record. Each layer is telling the truth about its own scope and nothing beyond it.
So the question for every automation you run is not "is it working." It is: what would I see if this were silently failing right now? If the honest answer is "the same thing I see now," you have no monitoring, regardless of how many dashboards you have.
What are the email deliverability thresholds you cannot cross?
This is the best-documented part of the whole topic, so there is no excuse for guessing.
Google's sender guidelines define a bulk sender as anyone sending 5,000 or more messages a day to personal Gmail accounts, counted across your primary domain and all its subdomains. Once you cross it, the designation sticks — you do not drop back out by sending less next week.
The requirements for bulk senders:
- SPF and DKIM authentication on all mail.
- A published DMARC record. One correction worth making, because the internet is confidently wrong about it:
p=nonesatisfies the requirement. You do not needp=quarantineorp=rejectto be compliant. Stricter policies protect you against spoofing and are worth moving toward, but having no DMARC record is the failure, not having a permissive one. - Spam complaint rate below 0.3%, with under 0.1% as the real target.
- One-click unsubscribe per RFC 8058, honoured within two days. This applies to marketing mail; transactional mail is exempt. The standard explicitly prohibits making the user land on a preference centre, log in, or confirm anything.
And the silent part. If your complaint rate goes above 0.3%, Google throttles you or routes your mail to spam. Your email platform will still report those messages as delivered, because the relay accepted them. The only place you see the truth is Google Postmaster Tools, which is a separate thing you have to go and look at. If nobody on your team opens Postmaster Tools, you do not know your deliverability — you know your send volume.
Microsoft has moved in the same direction for Outlook addresses. Treat the Google requirements as the floor everywhere rather than tracking each provider separately.
Where do messaging automations fail silently?
The clearest documented example of a success response meaning nothing comes from messaging throughput, and it catches everyone exactly once.
Messaging platforms enforce two different limits. Exceed the concurrency limit and you get a loud failure — Twilio returns HTTP 429 with error 20429, which you can catch and back off from. Good. That is a failure doing its job.
But exceed the messages-per-second rate at which the platform hands messages to carriers, and the API accepts your request, returns success, and quietly queues the message — for up to four hours, and longer on some campaign types. Only when the queue overflows do you get a loud error.
Now think about what you put through that pipe. A one-time passcode. A webinar-starting-now reminder. An appointment confirmation for a slot in twenty minutes. All of them reported as sent, all of them useless on arrival. The system did not fail. It did precisely what it documented, and your send looks perfect in every report you have.
The fix is architectural rather than clever: know your provisioned rate, pace your sends beneath it rather than firing a batch at the API, and for anything time-critical, treat delivery latency as a thing you measure rather than assume.
How do retries create duplicates?
This is the edge case that produces the most embarrassing failures — the customer who got the same message four times, or got charged twice.
Webhooks are built on an assumption people miss: the sender must hear a success response quickly, or it assumes failure and tries again. Stripe, for example, expects your endpoint to return HTTP 200 within seconds, and if it does not, it retries with exponential backoff for up to 72 hours.
So if your webhook endpoint does the actual work synchronously — writes to the CRM, enrols the contact, fires the email, all before responding — and that work takes too long, the sender concludes nothing happened and sends it again. Your slow-but-successful operation now runs a second time. And a third.
Two things prevent this, and both are boring:
Idempotency. Record the unique event identifier the sender provides, and check it before acting. If you have seen that event before, acknowledge and do nothing. Without this, every retry is a fresh duplicate.
Respond fast, work later. Return 200 immediately, put the payload on a queue, and process it asynchronously. Failed payloads go to a holding queue for inspection rather than triggering a retry storm.
Related timing bugs in the same family, worth checking in your own setup: a contact enrolled in two sequences simultaneously, a contact re-entering a sequence it already completed, an automation triggering itself because it updates a field that is also a trigger, scheduled sends that shift by an hour when a time zone or daylight rule changes, and a record whose category changes mid-sequence while the sequence continues as though it had not.
What breaks first in Indian CRM data?
The phone number, and it is specific enough to fix this afternoon.
In most Indian businesses the mobile number is the real identity key. If your forms do not normalise input to a single international format, your CRM will cheerfully create separate records for 09876543210, 919876543210, +91 98765 43210 and 98765 43210. They are the same person. Your system now has four relationships with them, each getting its own sequence.
The same defect propagates outward with no error anywhere. Meta's Conversions API requires server-side and browser-side events to share an identical event ID and event name to be deduplicated, within a 48-hour processing window, and contact data must be normalised to a consistent format before hashing. Get the normalisation wrong and the match quality degrades — Meta's deduplication documentation sets out the mechanics. Nothing in your stack raises an error. Your targeting just gets worse, and you attribute it to the algorithm.
Contact data also decays continuously as people change jobs, numbers and companies. We have seen decay rates quoted confidently in the twenty-to-thirty percent range annually, but every version we could find traces to a data-cleaning vendor rather than independent research, so treat the direction as real and the number as marketing. Your own bounce trend is better evidence than anyone's benchmark, and it is a figure you already have — which is the general argument in measuring marketing honestly.
What about the "cost of bad data" statistics?
Worth a short detour, because these numbers are the main sales evidence for data tooling and we went looking for them.
The $3.1 trillion figure is not research. It is traceable — via a third-party audit of these statistics — to a 2011 trade-magazine article in which the author extrapolated from a single healthcare estimate and wrote, in the piece itself, that the number stood "no matter how far off my estimate is." An IBM infographic repeated it in 2013 with no methodology attached, and a 2016 Harvard Business Review article made it famous as IBM's. No national estimate with a published method appears to exist.
The $12.9 million per organisation figure is real, with a catch. It does come from Gartner, in its 2020 analysis of data quality tooling, and it is based on 154 reference customers of 16 data quality vendors. So the number is companies that were already buying data-quality software estimating their own costs. That is not nothing. It is also not a measurement of the general business population.
We mention this because the sequence matters: if someone is selling you data infrastructure on the strength of either number, they have not checked, and you should wonder what else they have not checked. The case for clean data does not need them. Your duplicate count makes the case on its own.
Which Indian rules apply to which channel?
Businesses get sold compliance for the wrong regime in both directions, so this needs stating plainly.
SMS and voice are governed by TRAI's commercial communication framework. Under the TCCCPR regime, you register as a principal entity, use registered telemarketers, get headers and content templates pre-approved, scrub against the preference register before sending, and use the designated numbering series — the 140 series for promotional and the 1600 series for transactional and service traffic.
The enforcement is not a fine, which is why it matters operationally. Under the February 2025 amendments, a first violation by a registered sender means all of that sender's telecom resources are barred for 15 days — SIMs, PRI and SIP trunks included. Repeat violations mean blacklisting for up to a year. Your SMS compliance failure takes your phone lines with it.
That framework does not cover WhatsApp, email or in-app messaging. If a vendor is selling you preference-register scrubbing or TRAI template registration for WhatsApp, you are paying for compliance with a regime that does not reach the channel. Those channels are governed by the platform's own rules and by data protection law.
Consent withdrawal is an engineering requirement, not a policy page. Under the Digital Personal Data Protection Act, 2023, consent must be free, specific, informed, unconditional and unambiguous, withdrawal must be as easy as giving it, and the burden of proving valid consent sits on you. For a running automation that means a withdrawal arriving mid-sequence has to propagate across every channel before the next step fires — CRM, email, WhatsApp, ads audiences — and you must be able to produce evidence of what the person consented to and when.
The Act's Schedule sets the exposure: up to ₹250 crore for failing to maintain reasonable security safeguards, ₹200 crore for failing to notify a breach or for children's-data breaches, ₹150 crore for significant data fiduciary obligations, and up to ₹50 crore under the residual provision, which is where a consent or withdrawal failure lands. The substantive obligations commence 13 May 2027 — far enough away that doing nothing now means rebuilding your consent records later, because they cannot be reconstructed retroactively.
Where does the handover actually fail?
The boundary between automation and a human is the least engineered part of most systems and the most expensive.
The canonical evidence here is older and narrower than people claim. The 2007 lead response study by James Oldroyd with InsideSales — roughly 15,000 web leads across six companies over three years — found that moving first contact from five minutes to thirty minutes cut the odds of reaching the lead by about 100 times and the odds of qualifying them by about 21 times. Two caveats we would insist on: it predates smartphones and WhatsApp entirely, it was co-authored with a vendor whose product sold speed, and it is frequently confused with a separate 2011 study that measured one hour against twenty-four and found a roughly sevenfold effect. The direction is well supported. The multipliers are from a specific context.
What that means for architecture rather than for a motivational slide: every automated path needs a named owner and a timeout. Not a round-robin into a shared inbox — a person, and a rule for what happens when they do not act. And your sequences need to know when a human conversation has started, so the automation stops talking over your own salesperson. That failure has no error code at all; the only signal is a customer asking why they are getting emails about a product they already bought. We go further into the design of this in why lead generation fails.
The Monitoring Floor: five things nothing will tell you
If you monitor only what your platforms volunteer, you are blind to every failure above. The minimum:
- Your spam complaint rate, from Google Postmaster Tools, weekly. Your email platform will not tell you this.
- Delivery latency on time-critical sends, not just send counts — the gap between accepted and actually delivered.
- Duplicate rate on your identity key, which for Indian businesses is usually the phone number. Count it monthly.
- Failed webhook and API payloads, in a place a human looks. A holding queue nobody clears is the same as no queue.
- Leads with no owner and no action after your timeout window. This is the handover failure, and it is invisible by design.
None of that needs engineering. It needs five recurring checks and someone whose job it is to run them — the point we make about the operator in what an AI marketing stack really costs.
When you should not build this
The honest counter-argument, and it applies to more readers than you would think.
A resilient system nobody maintains fails worse than a simple one somebody watches. Dead letter queues, idempotency layers, asynchronous workers and state machines all assume someone is monitoring the parts that catch failures. If nobody clears the queue, the queue is where your leads go to die, and you will not find out for a month.
At genuinely low volume — a few dozen enquiries a day — a person with a shared inbox and a simple checklist beats a branching sequence with no owner, and complexity is itself a failure source. The right moment to engineer for scale is when volume, not ambition, has made manual handling unreliable. Before then, reduce your branching, cut the number of live sequences, and spend the effort on response speed instead.
Four signs you have outgrown simple. You have crossed 5,000 daily emails to personal inboxes. You are sending time-critical messages in batches. Your duplicate count is rising faster than your contact count. Or you have had an incident where nobody could say what the system actually did.
Want your automations audited against this? Send us what you run — platforms, live sequences, channels, rough monthly volume — and what went wrong the last time something broke. We will come back with which failures in your setup are silent, what your actual deliverability position is, where your duplicates are originating, and what your consent records would have to look like before 2027. If the answer is that you should be running less automation rather than better automation, we will tell you that. Talk to Midgrow, or message +91 9632398896 with your sector and monthly volume. Engagement structures are on the services page, pricing is covered in what agencies don't tell you, measurement setup in our analytics piece, content workflows with our social work, and ecommerce-specific flows under retail and ecommerce.
Frequently asked questions
Why does marketing automation break at scale?
Three causes: unclean data entering the system, unhandled edge cases in branching logic, and undefined handover points between automated and human steps. The expensive version of each is silent — the platform returns a success response while the outcome is wrong — so automation should be designed around the failure path first.
What is Google's spam complaint threshold for bulk senders?
Below 0.3%, with under 0.1% as the real target. Bulk sender status applies at 5,000 or more messages a day to personal Gmail accounts, counted across your primary domain and subdomains, and it does not lapse if volume drops. Breaching the rate means throttling or spam-folder routing while your email platform still reports the messages as delivered.
Is a DMARC policy of p=none enough for Gmail and Yahoo?
Yes. Publishing a DMARC record satisfies the bulk sender requirement and p=none meets it; having no record at all is the failure. Stricter policies like p=quarantine or p=reject protect your domain against spoofing and are worth working toward, but they are not required for compliance.
Why do automated messages arrive late even when they show as sent?
Because messaging platforms accept requests faster than they hand messages to carriers. Exceeding the concurrency limit returns a visible error such as HTTP 429, but exceeding the messages-per-second rate returns success and silently queues the message — for hours in some cases. Time-critical sends need pacing below your provisioned rate and latency monitoring, not just send counts.
How do I stop automation sending duplicate messages?
Use idempotency and respond fast. Record the unique event identifier each sender provides and check it before acting, so a repeated event is acknowledged and ignored. Return HTTP 200 immediately and process the work asynchronously, because senders that do not get a prompt success response retry — Stripe, for example, retries with exponential backoff for up to 72 hours.
Does TRAI's DLT framework apply to WhatsApp marketing in India?
No. TCCCPR governs commercial messaging routed through telecom operators — SMS and voice — including sender and template registration, preference-register scrubbing, and the 140 and 1600 numbering series. WhatsApp, email and in-app messaging fall outside it, and are governed instead by platform policy and data protection law.
What happens if you violate India's SMS marketing rules?
Resource barring rather than a fine. Under the February 2025 amendments, a first violation by a registered sender means all that sender's telecom resources are barred for 15 days, including SIMs and PRI or SIP trunks, with blacklisting for up to a year on repeat violations. An SMS compliance failure therefore takes your phone lines down with it.
What does DPDP require of a running marketing automation?
That withdrawal of consent is as easy as giving it, propagates across every channel before the next sequence step fires, and is evidenced — the burden of proving valid consent sits on the business. Substantive obligations commence 13 May 2027, and consent records cannot be reconstructed retroactively, so the capture work has to start before the deadline.


