Every marketing automation you add sends more email, and sending more email to a list that is not more engaged is the fastest way to lose inbox placement.
That constraint governs the whole subject and almost no automation advice mentions it. The pitch is always additive — add a welcome sequence, add cart recovery, add a re-engagement flow, add a browse abandonment trigger — as though volume were free.
It is not free. It is charged against your domain reputation, and the bill arrives as a quiet decline in delivery across everything you send, including the receipts and password resets that have nothing to do with marketing.
This article is about deciding what to automate under that constraint, rather than automating whatever a platform makes easy.
What automation is actually good at
Two jobs, and they are narrower than the category implies.
1. Doing a defined thing at a moment you cannot be present for. Somebody subscribes at 2am and gets the welcome sequence. A cart is abandoned on a Sunday and the reminder goes at the right interval. The value is timing, not intelligence — the message could have been written a year ago and often was.
2. Doing a defined thing at a volume a person cannot sustain. Tagging, segmenting, routing, deduplicating, reporting. Dull, high-frequency, rule-shaped work.
What it is bad at is anything requiring judgement about a specific person, which is most of what people now try to point AI at.
A useful test before automating anything: could you write down the rule, completely, including what happens when the input is unusual? If you cannot, you are not automating a process — you are hiding one.
Where AI changed the picture, and where it did not
AI made the generation of marketing material almost free. It did not make the distribution of that material free.
That asymmetry is the whole story of the last two years. The cost of producing an email, an ad variant, a landing page or a follow-up sequence collapsed. The cost of getting any of it seen did not — inbox placement, search visibility and attention are all still rationed, and they are rationed harder now precisely because supply went up.
What genuinely improved:
- First drafts. A usable draft in a minute rather than an hour
- Variants at volume. Twenty subject lines to test rather than three
- Summarising and classifying. Support tickets into themes, reviews into complaints, survey responses into segments
- Translating between formats. One long piece into the eight shapes it needs to take
What did not improve:
- Knowing what is worth saying. A model has no access to what your customers told you on the phone last week
- Judgement on edge cases, which is where automation reputations are made and lost
- The distribution constraint, which tightened
The constraint, stated properly
Adding automations increases send volume to your least engaged contacts, because your engaged contacts were already receiving mail.
Work through what a typical stack does:
- A welcome sequence sends to everyone who subscribes
- A re-engagement flow sends specifically to people who stopped opening
- A cart recovery flow sends to people who did not buy
- A browse abandonment flow sends to people who looked and left
- A win-back flow sends to people who left months ago
Notice the pattern. Every flow after the first targets people defined by not engaging. Those are the contacts most likely to complain, most likely to be dead addresses, and most likely to include a recycled spam trap — an abandoned address that a provider has converted into a trap precisely because nothing legitimate should still be mailing it. What spam traps do.
The consequences are measurable and they are the ones providers act on:
- Complaint rate. Google's bulk sender requirements state complaint rates should stay below 0.3%, and the working target for a healthy programme is far lower than that ceiling
- Bounce rate, which rises as automations reach older, colder addresses. What bounces signal
- Engagement rate, which falls mechanically as you send more to people who engage less
And the damage is not confined to marketing. Domain reputation is shared across everything the domain sends. A re-engagement campaign that generates complaints degrades the delivery of your order confirmations, which is why the streams should be on separate subdomains before the automation count goes up. How to separate them.
The rule that follows: before adding an automation, decide which existing send it replaces, or accept that you are spending reputation to run it. Automation budgets are real; most people never draw one up.
What to automate, and what not to
Two questions place almost anything: how often does it happen, and how much judgement does each instance need?
| Low judgement | High judgement | |
|---|---|---|
| High frequency | Automate fully. Tagging, routing, reporting, welcome sequences, receipts | Automate the preparation, not the decision. Draft and queue for review |
| Low frequency | Automate if it is cheap. Not worth much effort either way | Do not automate. Complaints, refunds, anything where being wrong is expensive |
The top-right cell is where AI actually helps and where most people misuse it. A model that drafts a reply for a person to send is doing the useful part. A model that sends the reply itself has removed the only step that catches the case it got wrong.
Three things worth keeping manual regardless of frequency:
- Anything a complaint could become. Refunds, cancellations, billing disputes
- The first message to a person who matters commercially. A prospect who can tell it was generated learns something about how you will treat them later
- Anything where being wrong is public. Social replies, review responses, anything with a screenshot risk
Automation debt
Every automation you build is a thing that can break silently, and nobody owns it.
The failure shape is specific: an automation that stopped firing looks exactly like a segment that nobody entered. There is no error. The dashboard shows zero sends, which is also what a quiet week looks like.
Four ways they break:
- An integration's API changes and the trigger stops receiving events
- A field is renamed and the personalisation token resolves to nothing
- A condition becomes unreachable because an upstream tag was retired
- A platform migration carries the sequence across but not the trigger
Two habits prevent most of it:
Keep a register. One page: every automation, what triggers it, who owns it, when it was last confirmed working. Sites accumulate flows built by people who have left, and nobody dares turn them off because nobody knows what they do.
Test quarterly by entering your own automations. Subscribe with a spare address, abandon a cart, trigger the win-back condition. Twenty minutes, and it is the only thing that catches a silent failure before a customer does. The same discipline applied to sequences.
Measuring it honestly
The number every platform shows you is the one that means least.
Revenue attributed to an automation is not revenue caused by it. A cart recovery email that goes to someone who was returning anyway gets credited with the sale. The attribution window does the work, and a longer window flatters the automation.
Three questions that produce a truer answer:
- What is the incremental effect? A holdout group — even 5% who receive nothing — turns an attribution number into a measurement. Almost nobody runs one, and it is the single highest-value change available
- What did it cost in reputation? Complaint and unsubscribe rate for that flow specifically, not the account average
- What did it cost in attention? A contact has a finite tolerance. A flow that earns £200 and burns the audience for a launch has a cost that does not appear in its own report
Open rate is not a measure of anything since privacy protection began pre-fetching images. It is directionally useful for comparing two sends to the same list on the same day, and it is not a performance metric. What to use instead.
A realistic starting stack
For a small business, in the order worth building.
- A welcome sequence. The highest-engagement mail you will ever send, going to people who just chose you. How to build one
- Transactional email that works. Receipts, resets, confirmations, authenticated properly. Boring, and it fails more often than marketing does because nobody monitors it
- One recovery flow, whichever fits your model — cart, booking, or trial
- Tagging and segmentation rules, so later flows can target engagement rather than everyone
- A reporting digest, so the numbers reach you without you fetching them
Then stop and measure for a quarter before adding anything.
What to resist early: re-engagement flows, which target your least engaged contacts and are the most likely to damage reputation on a young domain; and anything that sends to a list you did not build yourself. Why bought lists are not a shortcut.
Frequently asked questions
What should a small business automate first?
A welcome sequence, because it reaches people at their most engaged, and transactional email, because it fails more often than anyone checks. After that, one recovery flow matched to your business model. Add nothing else until you have measured a quarter.
Does marketing automation hurt deliverability?
It can, and the mechanism is specific: most flows after the welcome sequence target contacts defined by not engaging — lapsed, abandoned, dormant. That raises complaint and bounce rates and increases the chance of hitting a recycled spam trap, and the reputation damage applies to everything the domain sends.
Can AI write my marketing emails?
It writes usable first drafts and generates variants to test, which is genuine time saved. What it cannot do is know what your customers told you on the phone, or judge an unusual case. Use it for the draft and keep a person on the send.
How do I know if an automation has stopped working?
You mostly do not, which is the problem — a broken automation looks identical to a quiet segment, with no error and zero sends. Keep a register of every flow and what triggers it, and test quarterly by entering your own automations with a spare address.
Is attributed revenue a reliable measure of automation performance?
No. Attribution credits the automation for sales that would have happened anyway, and a longer attribution window flatters it further. A holdout group of even 5% turns the number into a measurement, and almost nobody runs one.
Should transactional and marketing automation share a sending domain?
Better not to. Domain reputation is shared, so complaints generated by a re-engagement campaign degrade the delivery of order confirmations and password resets. Separate subdomains keep one from taking the other down.
How many automations is too many?
The count matters less than what they send and to whom. The useful discipline is deciding, before adding one, which existing send it replaces — or accepting that you are spending reputation to run it.
Free: The 60-Minute Email Authentication Fix
A no-fluff checklist to set up SPF, DKIM & DMARC correctly and pass Gmail & Yahoo's sender requirements.
Muhammad Basim has worked in digital marketing since 2013, focused on email deliverability and AI-assisted content production. He is the author of the Email Deliverability Playbook and the Email Copywriting Playbook, and has run 100+ email campaigns for ecommerce brands, coaches, and B2B senders. He writes about email, SEO, WordPress, and AI — with a bias toward what can be tested over what sounds good.
Related Articles
What to Give AI Before You Ask
Generic output is almost always a context problem, not a prompt problem. A model with nothing specific to work from produces the average of what has been written on the subject. That is not a flaw to be worked around with better phrasing — it is the correct behaviour given no information. The fix is […]
Connecting Your Tools: Automation That Does Not Break
Every integration you add is a thing that can fail without telling you. That is the property that makes connected systems different from the tools they connect. A broken tool announces itself — you open it and something is wrong. A broken integration produces silence, which is indistinguishable from a quiet week. And the failure […]
AI Email Personalisation: What Is Worth Doing
Personalisation built from what a customer told you works. Personalisation built from what a model inferred about them produces complaints. That line is the entire subject, and it does not move as the technology does. The risk is not that the inference is bad. It is that it is confident, specific and occasionally wrong — […]