Article
What I Don't Automate (And Why)
Jun 9, 2026 · 5 min read
Knowing what to automate is easy. Knowing what to leave alone is the harder skill, and the one that prevents real damage.
Automation is not a virtue. It is a tool.
Everyone talking about AI automation focuses on what to automate. The more interesting question is what not to. Getting that boundary wrong is where real damage happens — not catastrophic failure, but the slower kind: brand erosion, relationship damage, decisions made on bad inputs, trust quietly broken.
I automate the boring parts aggressively, and I gate the risky parts deliberately. That sentence sounds obvious. Operating it consistently is harder than it sounds.
What I do automate
The tasks I automate without reservation share a few properties: they are high-volume, the inputs are structured, the outputs are verifiable, and the cost of a mistake is low and recoverable.
Lead enrichment, dedupe, scoring, and prioritisation. These are mechanical tasks that scale well. The output is a ranked list with supporting data. If the enrichment is wrong, I catch it during review before it affects anything downstream. The worst case is fixable.
Drafting outreach variants and follow-ups. I automate the drafting. I do not automate the send decision for anything that matters. The automation generates candidates, I approve the batch. The leverage is in not writing from scratch, not in removing my judgment from the final call.
Summarising inbound emails and extracting next actions. This is one of the highest-value automations I run. The volume of inbound is high, the structure of what I need from each message is consistent — sender, context, ask, urgency — and the output feeds a triage queue that I work through. Wrong summaries are obvious. Caught quickly. No downstream harm.
Ops anomaly flags, dashboards, and reminders. If inventory drops below a threshold, I want a flag. If a pipeline has not run in twelve hours, I want an alert. These are informational signals, not decisions. The human still decides what to do. The automation just makes sure the signal gets through.
Content and creator discovery, ranking signals. Scraping signals, ranking by relevance, surfacing candidates — all of this can run automatically. The outputs are inputs to a decision, not decisions themselves. I review the top candidates; I do not automatically act on the ranked list.
What I do not automate
The tasks I refuse to automate fully have a different profile: high judgment, high stakes, hard to reverse, or brand-sensitive.
Partner negotiation messaging. The first message to a meaningful partner is not a conversion play. It is a relationship opening. Tone, framing, specificity, and timing all matter in ways that are difficult to constrain reliably with a prompt. The cost of a generic or off-tone message is not just a lost reply — it is a closed door that may not reopen. This does not mean AI has no role. It drafts, I edit significantly, and I send. The automation does the scaffolding; I do the relationship work.
Anything legal or compliance-adjacent. Terms of service, guarantees, claims, contractual commitments. AI will generate plausible-sounding text here. Plausible-sounding is not the same as accurate or defensible. Nothing in this category goes out without human review, and usually legal review.
Pricing decisions without data review. Pricing has downstream consequences across positioning, margin, and customer expectations. An automation can surface the data and model scenarios, but the decision to change a price is mine. The judgment required to weigh all the implications is not something I trust to an automated pipeline.
Brand voice content that could damage trust. There is a class of content — thought leadership, founder communications, public-facing narrative — where the voice is load-bearing. Getting it wrong does not just produce a bad piece of content; it produces a piece of content that represents me or the business in a way that is hard to walk back. I use AI to accelerate drafts here, but the final version is heavily edited. The automation assists; the author is still me.
High-stakes customer communications. Refund disputes, escalations, complaints from important customers. These require empathy, context, and the judgment to know when a policy exception is the right call. Automating a response here is not a time-saving measure — it is a risk that the customer interaction goes badly in a way that is far more expensive than the minutes saved.
The underlying principle
The line I draw consistently is between preparation and decision.
I automate preparation: enrichment, triage, drafting, discovery, summarisation, scoring. These are the inputs to human judgment, not replacements for it. They let me make better decisions faster, not skip decisions entirely.
I gate decisions: final sends, pricing moves, negotiation tone, brand-sensitive content, anything with legal weight. The automation brings me to the decision point faster and better-informed. The decision itself stays mine.
This is not excessive caution. It is a deliberate allocation of where automation creates leverage versus where it creates liability. The more I trust the preparation layer, the more confidently I can act at the decision layer. The automation and the judgment reinforce each other — as long as I keep them in their respective roles.
When I see automations that remove human judgment from high-stakes decisions entirely, I assume it is because the builder has not yet experienced a bad outcome from it. I have. The lesson is cheaper to learn from someone else's experience.
Keep the risky parts gated. The leverage is still enormous, even with those gates in place.