The short version: automate the repetitive, low-stakes mechanics of marketing (scheduling, first-draft reporting, basic admin) but keep your hands on anything that involves a real relationship, a real apology, or a real decision about what your business stands for. The tasks that cost you the most when they go wrong are exactly the ones people try to automate first, because they’re the ones that eat the most time.
Next step on this topic: Is an AI Assistant Better Than a Chatbot for Business Tasks?.
Why this question matters more in 2026 than it did two years ago
Every tool now has an “AI assistant” bolted onto it. Your CRM will draft the follow-up email. Your scheduler will write the caption. Your chatbot will handle the “quick question” from a prospect at 11pm. None of that is the problem. The problem is that automation has got so good at sounding human that we’ve stopped asking whether a task should be automated, and started only asking whether it can be.
Those are different questions. I’ve spent five years rebuilding my own business in public after a period where I leaned too hard on tools and outsourcing, and the lesson I keep relearning is boring but true: the tasks that feel most tedious are usually the ones doing the actual work of building trust. Automate those and you don’t save time, you borrow it, and it comes back with interest.
The comment I never should have automated
Back when my LinkedIn following was around 300,000, I tried a commenting tool for three weeks. It scanned posts from people in my network and dropped a relevant-sounding comment automatically, three or four times a day, so it looked like I was engaged everywhere at once. Engagement on my own posts went up slightly in that window. But two things happened that I didn’t expect.
- A woman I’d known for years, whose father had just died, posted about it. The tool commented “Great insight, thanks for sharing!” under her post. She messaged me directly and asked if I was okay.
- A prospective client later told me, almost as a joke, that he’d noticed my comments “didn’t sound like me anymore” and it made him hesitate before booking a call.
Neither of those cost me a client outright. But they cost me something harder to get back: the sense, in two specific people’s minds, that the person behind the profile pays attention. That’s the whole product when you’re selling trust-based services. I switched it off after week three and never turned it back on, and I still tell this story to clients who ask me to set up the same thing for them.
Never automate: replies where someone is talking to you
There’s a difference between a comment that says “nice post” and a comment that says “we tried this last quarter and it flopped, any idea why.” The first is noise and you can safely batch a reaction to it. The second is a person handing you an opening, and an automated reply (even a good one) reads as a slight when someone can tell it wasn’t read.
The test I use: if the comment or message took the other person more than about ten seconds to type, it deserves a human reply. That’s most of your DMs, most of your comments on personal posts, and every single reply to a complaint. Batch-schedule your content all you like, but keep a human on the replies where a human clearly wrote the question.
Never automate: apologies, complaints, and anything with a name attached to real money
I worked with an agency client last year whose support team used a canned AI response for refund requests. It was polite, on brand, grammatically perfect, and it made three separate customers angrier because it didn’t acknowledge what they’d said, it just pattern-matched to “refund” and fired off the standard line. One of those three left a public review naming the exact phrase the bot used, word for word, because it was so obviously copy-pasted.
Money and apologies are the two places where “close enough” isn’t close enough. If a customer is unhappy, they need to feel heard before they need to feel processed. A human doesn’t have to be slower here, they just have to read the message first.
Never automate: outreach to people who already know you
Automated outreach sequences are fine for cold prospects who’ve never heard of your business, that’s arguably what they’re for. But I still see agencies put existing clients, old colleagues, and warm referral partners into the same drip sequence they use for cold leads. It shows. People who know you can tell within one email that they’ve been dropped into a funnel, and it reads as a demotion in the relationship, even if nothing else about the message is wrong.
If you’re doing partnership outreach, joint webinar invites, or anything to a name you could recognise without looking them up, write it yourself or dictate it and clean it up. It takes five minutes longer and it’s the five minutes that keeps the relationship a relationship.
Never automate: the strategy underneath the content
AI is useful for drafting captions, restructuring a blog outline, or summarising a competitor’s last ten posts. It is not useful for deciding what your business should stand for this quarter, which audience you’re trying to reach, or which of three campaign directions is worth the budget. I use tools constantly for the mechanics of this (I’ve written before about the AI prompts for marketing that produce usable drafts instead of generic copy-paste output), but the decision about direction has to sit with a person who knows the business, the market, and what’s happened in the last three months that no model has been trained on.
Look at how brands that win long term operate. The Notion marketing strategy worked because someone made deliberate, sometimes odd, human calls about tone and community that no automated content calendar would have suggested. Same with the ASOS marketing strategy, where the personalisation only works because someone decided what “on brand” meant before any tool touched it. Automate the execution, never the judgment call about direction.
Never automate: your own opinions in thought leadership content
Here’s the uncomfortable bit that most people writing about “AI and marketing” tend to skip past: a huge amount of LinkedIn thought leadership right now is fully AI written, and readers can tell, and it is quietly costing those accounts reach even when the engagement numbers still look fine. I see it daily. Posts with the same three-sentence-paragraph rhythm, the same “here’s what nobody tells you” opener, the same tidy bullet list of lessons. They get likes from people scrolling fast. They don’t get replies, DMs, or client enquiries, because nobody feels like they’ve met a person.
I use Claude and ChatGPT constantly for research, structure, and getting a messy first draft down fast (I’ve written a full breakdown on using Claude for business tasks that’s more specific than the generic “10 prompts” lists floating around). But the opinion, the specific story, the number from my own experience, that has to come from me typing it or saying it out loud. The moment you let a model generate the opinion as well as the sentence structure, you’ve stopped doing thought leadership and started doing content filler, and your audience notices faster than you’d like.
The uncomfortable truth about “saving time”
Automation doesn’t save time on any of the tasks above. It moves the time. You spend less time now and more time later fixing a relationship, writing a correction, or explaining to a client why the campaign missed what mattered. I’ve watched agencies automate client status updates only to spend three times as long later doing damage control on a client who felt ignored. The invoice for automating badly always arrives, it’s just delayed and usually bigger than the bill you’d have paid in staff hours.
That’s not an argument against automation generally. Automating invoicing, scheduling, first-pass reporting, and routine admin is one of the smartest moves a small marketing team can make, and it does save time with almost no downside (I’ve laid out exactly how in a piece on using AI for invoicing and admin in marketing agencies). The distinction isn’t “AI good, AI bad.” It’s whether the task involves a relationship, a judgment call, or money attached to a person’s name. If yes, keep a human on it. If it’s mechanical and low stakes, automate it without guilt.
A quick way to decide what’s safe to automate
When a client asks me whether a task can be automated, I run it through four questions:
- Would the person on the other end notice, and mind, if they found out it was automated?
- Does getting it wrong cost more than the time saved by automating it?
- Does it involve a specific name, a specific complaint, or a specific amount of money?
- Does it require a judgment call about the business’s direction, not just its execution?
One “yes” and I keep it human. Two or more, and it’s not even close. The tasks that pass all four (formatting reports, scheduling posts, tagging leads, drafting first versions for a human to edit) are where automation earns its keep. If you’re not sure where your own business sits on this, that’s usually the first thing I map out with a client before we touch a single tool, and it’s worth checking the current numbers on where AI moves the needle for small businesses before you assume automating more is automatically progress; the data in our AI statistics roundup for small business owners is a decent reality check on that.
If you want someone to sit down with your specific workflows and mark, task by task, what’s safe to hand off and what isn’t, that’s exactly the sort of decision an outside pair of eyes is useful for, and it’s the core of what I do when I work as an AI implementation coach with small teams.
Frequently asked questions
Is it ever okay to automate customer replies?
Yes for simple, high-volume, low-stakes questions like “what are your opening hours” or order-tracking updates. No for anything involving a complaint, a refund, or a message that clearly took the customer real effort to write. The line is whether the message contains a specific problem with a specific name attached to it.
Can AI write my LinkedIn posts for me?
It can draft structure and tighten wording, but the opinion, the story, and the specific number or example need to come from you. Readers can tell the difference between a post with a real experience in it and one built entirely from a prompt, and it shows in replies and DMs even when the like count looks similar.
What’s the single biggest mistake businesses make when automating marketing?
Automating based on what’s technically possible rather than what’s appropriate. Just because a tool can send 500 personalised-looking messages a day doesn’t mean it should, especially to people who already have a relationship with the business. The mistake isn’t using the tool, it’s using it on the wrong list.
How do I know if automating a task will save me time?
Ask what it costs if it goes wrong. If the worst case is a typo, automate freely. If the worst case is a damaged relationship, a public complaint, or a lost client, the time you save now is almost always smaller than the time you’ll spend fixing it later.
Related reading: How to Plan a Month of Social Content Without a Big Team and Why Your LinkedIn Account Got Restricted (and How to Fix It).
I go much deeper on this in the AI marketing guide.
