I've spent two years putting these tools into real client marketing operations rather than evaluating them in the abstract, and I've arrived at a fairly firm line about what I'll automate and what I won't. It isn't where the vendors draw it, and it has less to do with capability than with what happens when something goes wrong.
What I automate without hesitation
Anything that reads rather than speaks, and anything whose output a human sees before a customer does. Summarising the week's sales calls into recurring objections. Clustering support tickets. Assembling pre-call research. Drafting the first version of a report. Turning a workshop recording into a structured brief. These remove genuinely tedious work, and a mistake costs someone ten minutes of confusion rather than a customer relationship.
What I won't automate, and why
Anything that reaches a customer without a person having read it. Not because the output quality is bad — it's frequently fine — but because the failure mode is asymmetric. A hundred good automated emails produce a modest gain. One wrong automated email to your largest account produces a problem that takes months to repair. The expected value calculation looks favourable right up until the tail event, and the tail event is the only one anyone remembers.
- Automate: reading, summarising, clustering, drafting, research, monitoring
- Never automate: anything sent to a customer without human review
- Never automate: pricing, contractual claims, or anything about a named customer
- Be careful with: anything published under a named person's byline
The quiet cost of over-automation
There's a second-order effect that doesn't show up in a time-saved calculation. When the summarising is automated, nobody reads the calls. When nobody reads the calls, the team's intuition about customers slowly degrades, and in eighteen months you have a marketing team whose feel for the buyer comes entirely from summaries. I now deliberately keep some manual contact in the loop — a rule that everyone listens to two full customer calls a month, unsummarised.
Audit what you've automated once a quarter
Automations accumulate exactly like tools do, and they degrade silently — a data source changes shape, a prompt stops matching reality, and the output is subtly wrong for months because nobody's checking. Once a quarter, go through every automation, look at the actual output, and confirm someone still owns it. Several will turn out to have been producing quiet nonsense, and nobody noticed because nobody was reading it either.
Automate everything that reads. Be very careful with anything that speaks. The asymmetry between those two isn't close.
Want to apply this to your business?
Book a free 30-minute strategy call and we'll diagnose your specific growth blockers together.
Book a Free Call