"I deleted half our marketing automation. Conversions went up."
A long conversation with the head of growth at a UK-based mobile app company who, over eighteen months, stripped out roughly half the automated lifecycle messaging her team had inherited. What stayed, what went, and why the numbers improved.
The company is a B2C app in the personal-finance space, founded in 2019, with — by its current published figures — somewhere around six hundred thousand monthly active users. The head of growth, who I will call H at her request, joined the company in 2023. She inherited, at that point, a marketing automation stack that had been built up over the preceding four years and that included, by her count, around two hundred and forty separate automated messages running across email, push, and in-app channels.
Over the eighteen months following her arrival, she deleted around half of those messages. The conversion rates on the messages that survived — and the overall business metrics the team was measuring — improved noticeably. The deletion exercise has, in her view, been one of the most consequential things she has done at the company. The conversation that follows is lightly edited and runs roughly in the order it was had.
The conversation
Marigold: Let's start with the obvious question. Why delete?
H: The honest answer is that the automation stack I inherited had been built one campaign at a time, by a series of different people, over four years, with very little ongoing maintenance. Nobody currently at the company had a complete picture of what it was doing. I asked, in my first month, for a complete inventory of the automated messages we were sending and how often each one fired. It took the team two weeks to produce. The inventory, when it arrived, was alarming.
M: Alarming how?
H: Several of the messages were firing thousands of times a month and had not been reviewed by anyone in over two years. Several had been built to address specific situations that no longer existed — features the product had since changed, segments we no longer served, partner integrations that had since been deprecated. Several were duplicating other messages by accident: the same user, in the same situation, was receiving the same essential prompt through two or three different channels. Several were, in candour, just bad — old copy, broken links, references to product features that had been renamed two years earlier.
M: What did you do first?
H: I turned off all the messages that had not been reviewed in the past twelve months. Not deleted — turned off, on a fourteen-day window, so we could observe what happened. There were about ninety messages in that bucket. The effect on the headline business metrics was, over those two weeks, essentially zero. The effect on the volume of complaints we were receiving from customers about email frequency — which I had not, before that point, fully understood was a significant complaint category — was substantially positive.
"I turned off ninety automated messages. The business metrics did not move. The customer complaints went down. That alone told me the automation we had was, in aggregate, doing harm rather than good."
On what was kept
M: So you turned off ninety. You said you deleted around half over eighteen months. What about the rest?
H: The rest came out more slowly, after more careful analysis. The framework we developed, after the initial turn-off, was to evaluate each automated message against three questions. Does the message do something that the user would want done on their behalf? Does the message, on incrementality testing, actually produce a behaviour change? Is the message tonally consistent with how the brand currently presents itself? A message that failed any of the three got cut.
M: What survived?
H: Roughly half of the automations we started with. The survivors are, in most cases, doing one of three jobs. The first is transactional confirmation — receipts, billing notifications, account changes — that the user genuinely wants. The second is what I would call situational support — messages that fire when the user has done something that suggests they might need help, and that offer help in a way that does not assume they are about to churn. The third is a small number of true marketing messages — feature announcements, product updates, the occasional well-crafted re-engagement prompt — that we have built carefully and that we monitor closely.
M: Anything else?
H: Yes — and this is the part that took me longest to figure out. We replaced several of the automations with a single weekly digest email that is partly templated and partly written by an actual human. The digest covers, for each user, the things that have happened in the past week that they might want to know about. It replaces, by my count, perhaps fifteen of the old automated emails that were each handling a single event. Users have, in their feedback, told us they prefer it considerably to the previous approach.
On what got worse
M: Did anything get worse as a result of the deletion?
H: Two things, on the metrics I watched most closely. The first is that our short-term reactivation rate of users who had been inactive for between fourteen and thirty days went down by about eight per cent. We had, in the previous stack, a particularly aggressive series of reactivation prompts for that window, and removing the aggression reduced the short-term reactivation. The second is that the rate of users completing their second purchase within thirty days of their first went down by about three per cent — we had, previously, been pushing the second purchase quite hard, and we stopped pushing it.
Neither of these decreases, on closer analysis, turned out to be net negative. The users who had been reactivating under the more aggressive prompts were, on a six-month follow-up, more likely to subsequently churn permanently. The users who had been pushed into a second purchase were, on the same follow-up, more likely to be net detractors in the product's NPS work. The short-term metric losses were, in lifetime-value terms, actually small net gains.
M: What about the engineering and operational overhead?
H: Significantly reduced. The marketing operations team, which had previously spent a substantial fraction of its time maintaining and debugging the automation stack, now spends much less time on maintenance. They have, accordingly, been able to spend more time on the small number of automations that remain, which is one of the reasons those automations have improved.
On the cultural change
M: You said earlier that this exercise was one of the most consequential things you have done at the company. What did you mean?
H: I meant that the exercise changed the cultural default about what marketing automation is for. The previous default was: more automation is better. The new default is: each automation has to justify its existence. The new default has, in the eighteen months since we established it, prevented a lot of bad automation from being built. We have, since I started, added perhaps fifteen new automated messages — each of which has been thoughtfully designed and tested. We have, on the old default, would have added perhaps eighty in the same period.
M: What would you say to a head of growth at another company who is wondering whether to do this exercise?
H: I would say: do the inventory first. Almost everyone I have talked to about this — and I have talked to several, because the exercise is, in candour, somewhat fashionable in the heads-of-growth community at the moment — has been surprised by what the inventory revealed. The starting point is almost always worse than the team thinks. The exercise gets easier once you can see the full picture.
I would also say: do the turn-off before the delete. We learned a lot from the fourteen-day observation window that we would not have learned if we had deleted the messages outright. The observation is essentially free. The deletion is, organisationally, much more difficult to reverse. Doing the observation first gives you the evidence you need to make the deletion confidently.
What I left thinking
I left the conversation with two main impressions. The first is that marketing automation, as it has accumulated in most mature app companies, is in substantially worse shape than the people responsible for it usually realise. The default tendency to add automations without removing them, combined with the lack of a regular maintenance practice, produces stacks that are confused, redundant, and frequently counter-productive. The simple act of auditing what is actually being sent — and turning off the messages that fail a basic justification test — is, in many companies, the highest-ROI marketing operations work that the team could be doing.
The second impression is that the cultural default H described is, in my view, the more important change. The "each automation must justify its existence" default produces a marketing operation that is smaller, more thoughtful, and more strategically valuable than the "more automation is better" default. The change in default is harder to achieve than the deletion itself; it requires sustained commitment from the head of the function. The companies that have managed it are, in our observation, in noticeably better operational health than the companies that have not.