Why e-commerce brands still need human-led email marketing
The button
There's a button in your Klaviyo dashboard that didn't exist eighteen months ago. You click it, type a sentence describing your ask, and sixty seconds later, you're looking at a fully formed email sequence with subject lines, copy, segmentation, and send times. It's reasonably good. And if you're paying an agency to manage your email, you've probably hovered over that button and wondered whether to just press it and cancel the retainer.
Flow Butler is a Klaviyo partner, and we use AI tools every single day. We've watched the platform evolve from a sending tool into something that, at first glance, seems like an existential threat to our industry. So we sat down and asked ourselves the uncomfortable version of the question: not whether AI will replace agencies, but whether it should.
This paper discusses what we found. We tried to be honest about it.
What the AI can do right now
Klaviyo's Composer takes a single sentence — something like "build me a spring reactivation campaign targeting lapsed customers across email and SMS" — and produces audience segments, messaging, channel selection, and timing. Their Marketing Agent goes further, autonomously planning and launching campaigns, creating on-brand content, personalising per recipient, and learning from the results. Their Customer Agent handles support via email, SMS, and WhatsApp, including order tracking, return processing, and product recommendations, all in your brand voice. Personalised Send Time uses reinforcement learning to determine the exact hour each person on your list is most likely to engage — not "Tuesday mornings work for retail," but the specific window for the specific individual.
These aren't future promises. They're shipping features that work, and the quality is often surprisingly close to what a competent human would produce if given the same brief.
So the question becomes: what's the gap? Where does "surprisingly close" fall short of "right for your business at this specific moment"?
Winbacks that don’t win
Anyone who's worked with AI coding tools has noticed something: the AI builds a feature that runs perfectly, seems to pass acceptance criteria, and outputs a close facsimile of the solution you expected. On closer inspection, you notice that there are components that don't seem quite right. It didn't know about the conversation you had with your customers last week or that you're pivoting the product next quarter. Constraints that no longer apply were ignored, because the feature is unaware of what your business needs right now.
Email marketing has the same gap. Klaviyo Composer builds a winback flow drawing on fourteen years of platform data and billions of consumer interactions. It knows what a good winback flow looks like across thousands of brands. And like all systems based on Large Language Models (LLMs), its stochastic nature operates within a narrow band shaped by everything it's already seen.
Consider a premium Australian skincare brand with two customer segments that both lapse around 90 days. Composer builds a single winback sequence — a gentle reminder, then a discount offer a week later, then a final "last chance" email. It is best practice to increase urgency across the three and have a "we miss you" sentiment throughout.
A strategist who knows this brand may recognise that for this brand, the standard playbook is precisely the wrong approach. They may discern that the two segments require different approaches. One may be seasonal gift buyers, purchasing for loved ones predictably during the Christmas period and Mother's Day. They are swayed by discounts and plan to return only in the next buying window. The second segment may be repeat buyers of consumable products, who will see a "we miss you" email as an awkward attempt at emotional connection when they simply forgot to replenish their hand wash, and a simple email focused on utility and expediency would suffice.
Simply put, the AI builds for the category, but the strategist builds for the business. That gap won't last forever, and we'll get to why, but right now the distance between "what a good winback flow looks like in general" and "what your winback flow should actually do" is still a space filled by people who know your business deeply enough to make those calls.
The homogenisation problem
The risk of AI-generated email isn't that it produces bad work. It's that it produces the same work for everyone. Every brand using platform AI draws from the same pool of training data, and the output is shaped by the same aggregate signals about what "works." The result is the platonic ideal of a marketing email — statistically optimal, structurally sound, and increasingly identical in tone, cadence, and approach across every brand in your customer's inbox.
Your customers don't just receive your emails. They receive everyone's. And when the underlying intelligence generating those emails converges on the same patterns — the same subject line structures, the same urgency triggers, the same three-email sequence with the discount in email two — the brain classifies repetitive patterns as "already processed" and stops registering them. Neuroscientists call this habituation, and research into creative fatigue shows it kicks in within three to four weeks for top-of-funnel content. When every brand follows the same AI-optimised template, that fatigue compounds across the entire inbox, not just your sends.
The more brands adopt AI-generated email at default settings, the more inboxes fill with structurally similar messages. Engagement drops as people stop noticing them. And because the platform's recommendations are shaped by aggregate performance data, the definition of "what works" gradually drifts into homogenous mediocrity. This is regression to the mean operating at industry scale.
Why “better prompts” aren’t enough
The obvious counter-argument is that if generic output is the problem, you should just ask the AI for something unconventional. Push it toward creativity. We thought about this for a while, and there are two reasons it's harder than it sounds.
The first is that while AI has lifted everyone's floor of creativity, it has compressed the ceiling. When you ask an AI to write something surprising, it reaches for its training data's version of surprising, i.e. the expected unexpected. Everyone's "unconventional" converges towards the same centre, surprising in the way a Marvel plot twist is surprising: technically unexpected, structurally familiar to anyone who's seen the previous twelve films. A 2026 theoretical analysis suggests this isn't a bug but a structural constraint: the system trades off between reliable output and surprising output, and the mathematics suggest it can't maximise both simultaneously. A human strategist can intentionally hold novelty and effectiveness in tension, knowing which rules to break and which to keep, because they understand the purpose behind the decision.
The second is that the real work isn't writing the email - it's making decisions around it. To generate something genuinely unconventional, the AI needs context that doesn't exist in one place: your audience's fatigue patterns, what competitors sent this week, which cultural moments are worth referencing, and which segment needs disruption versus consistency. Assembling that context and deciding what matters is the strategic work. And even with perfect context, there's the question of timing - you don't pattern-interrupt every send. The judgment of when to be unconventional is temporal, and harder to automate than the creative execution itself. It requires reading signals that exist outside the platform's data: social sentiment, competitive positioning, and founder decisions that haven't hit the marketing calendar yet.
The person deciding what to ask the AI to do, and when, and why — that's the job.
Three to five years out
We don't want to make the mistake of arguing against AI's current limitations as though they'll stay fixed. They won't. Agents with persistent memory are coming — systems that retain context about your brand across months of interaction, accumulating understanding of what works and why. Direct access to Shopify data, CRM records, and customer service transcripts alongside email performance. Multi-step reasoning over entire customer journeys rather than individual touchpoints. The ability to observe outcomes, form hypotheses, and adjust strategy without human input.
In three to five years, an AI agent might genuinely know that your gift-buyer segment is seasonal and your consumables segment is logistical. It might learn the difference by watching outcomes over many months, without a human telling it. So if agents get that good, does the agency disappear?
We think the answer depends on who owns the agent and who operates it. If the agent lives on the platform side — built into Klaviyo or similar — then you manage it. You configure it, evaluate its output, course-correct when it drifts. This works for large brands with dedicated marketing teams that treat AI management as a core part of someone's role. But for a founder running a five or ten-million-dollar DTC brand who hired an agency precisely because they don't want to think about email marketing every day? They've traded one job for another. Instead of managing campaigns, they're managing an AI that manages campaigns. The cognitive load hasn't gone away — it's just changed shape, like pouring water from one cup into another.
If the agent lives on the agency side — operated by people who do this work across multiple brands every day — then the agency becomes the agent operator. They handle the configuration, the prompt engineering, the quality control, and the "is this actually good?" evaluation that requires comparing output against deep brand knowledge. The client gets outcomes without managing the system. That's a stronger value proposition than writing emails manually, not a weaker one.
There's a practical reality underneath all of this. Getting an AI agent to consistently do the right thing requires someone who understands the problem deeply enough to build the scaffolding around it. Guardrails, evaluation criteria, feedback loops, exception handling — what happens when the agent sends something odd at 2 am on a Saturday before your biggest product launch? Someone needs to be across the sharp edges of the system, the way a parent needs to know which playground equipment their kids can handle. That someone is either on your team or on your agency's team.
The accountability question
Revenue from email drops twenty per cent in a quarter. What happens next? With an AI-only approach, who's accountable? The algorithm doesn't sit across from you and explain what went wrong or what it plans to do differently. You can turn it off, but then what? You can adjust settings, but which ones? The system optimised for what it was told to optimise for, and if the outcome was wrong, the debugging requires expertise in the system itself — understanding whether the goal was wrong, the data was wrong, or the model's interpretation was wrong. Most founders don't have that expertise, and acquiring it defeats the purpose of automation.
With an agency, there's a person. A Slack conversation. A meeting where someone says "here's what we think happened, here's what we're changing, and here's when you'll see the impact." If that plan doesn't work, there's another conversation. There's a human entity that can be held to account, that can reflect, that can change course based on a conversation rather than a parameter adjustment.
As our customers attribute a large portion of their online revenue to our services, they look for more than dashboards and performance attribution. They need to know that someone is watching, someone cares about the outcome, and that if things go wrong, someone will listen. AI will get better at explaining its decisions — decision logs and chain-of-thought transparency is improving rapidly. But explaining why something happened is not the same as being responsible for making it right. Under Australian Consumer Law, services must be supplied with "due care and skill." When an autonomous agent makes a poor strategic decision, who bears that obligation? The platform? The brand? It's legally untested. An agency relationship has clear contractual accountability, defined services, and a human entity responsible for delivery. That clarity has value.
The founder’s day
A typical founder we work with may be running a ten million dollar e-commerce brand with a team of eight to fifteen. They're making product development decisions, managing supply chain, watching cash flow, handling team problems, negotiating brand partnerships, dealing with customer escalations. Email marketing sits in that category they know intimately: important but not urgent, revenue-generating but not where their genius lives. They know it matters — it might represent thirty to forty percent of revenue — but they don't want to think about it every day the way they think about product or team.
The AI button promises something appealing: set and forget. But most founders have been burned by at least one "set and forget" promise — Facebook ads, SEO tools, automated social posting. The initial results were fine, but then performance drifted, and it was hard to work out why. What these founders actually value isn't email execution. It's cognitive offloading. They want to be able to say "we're launching a new product line in March" and trust that someone will handle the email implications — the teaser sequences, the launch campaign, the segmentation changes, the flow adjustments — without requiring detailed instructions at every step.
An AI button doesn't give you that today. It gives you execution on demand, but you still decide what to demand, still evaluate the output, still determine when to change direction. And the platform itself is not simple. Klaviyo has over 350 integrations, five types of flow triggers, conditional and trigger splits with unlimited branching paths, predictive models for lifetime value and churn risk, layered behavioural segmentation, deliverability settings, and a feature set that ships major updates every quarter. It's closer to Salesforce than it is to Mailchimp, with AI features that sit on top of that complexity. A founder who decides to go AI-native and manage it themselves isn't just pressing a button — they're operating an enterprise platform that changes faster than they can keep up with while also running their actual business. For smaller brands with one or two million in revenue and no dedicated marketing lead, AI tools are compelling. They're cheaper than an agency and far better than nothing. But these founders often can't tell the difference between a winback flow that follows best practice and one that's been optimised for their specific business, because they've never seen the latter. They don't know what they don't know. An agency provides the strategic layer that they can't yet build internally.
The evolving agency
None of this argues for agencies to stay the same. The agencies that survive this transition are the ones that change what they do while maintaining why they exist. AI handles first drafts, variant generation, data analysis, and personalisation at scale. A campaign that took a week to ideate, write, design, and build can be created much faster by having humans refine and redirect AI models' output.
What stays is brand immersion — spending real time with the founder, understanding the business trajectory, absorbing the brand's world so deeply that judgment calls become intuitive. Strategic sequencing — deciding which flows to build, when to change approach, how to respond to business pivots that haven't hit the data yet. Creative disruption — the pattern-interrupt email that only works because someone understood what pattern needed interrupting. And accountability — a human who owns the outcome and answers for it.
What's new is the skill set emerging around AI itself: prompt engineering for brand voice, evaluating AI output for the subtle ways it can be wrong that only domain expertise reveals, orchestrating multiple AI tools across a client's ecosystem, and connecting platform metrics back to business strategy in a way that drives action rather than just reporting. The role shifts from "we write your emails" to "we run your email intelligence operation and take responsibility for the outcomes." That's a harder job, a more valuable one, and further from full automation than the old version was, because it sits at the intersection of AI capability and human judgment — exactly where the leverage is highest.
There's one more thing worth naming, and it's something we noticed while writing this paper. We used AI to help draft it, feeding it examples of our own writing to get the voice right. It got close with many iterations and hundreds of prompts back and forth to clarify points, refine the structure, and adjust the tone. (Interestingly, even though we provided explicit instructions to remove the common tropes found in AI-generated written work, such as hedging and em-dashes, we still had to handcraft the final output.) But the moment we accepted that output as "our voice," we realised we'd frozen ourselves in time. Even with the perfect prompt that captures the essence of our writing style, AI cannot know how we are evolving, what we are currently reading that is changing our viewpoints, or what experiences have shifted the way we communicate. A writer's voice changes the way a person changes — gradually, in response to life. If you hand an AI your brand guidelines and your last fifty emails and say, "This is how we sound," it may replicate that voice faithfully, but lure you into a trap. How will your brand voice mature as your audience matures? If you're not continually writing and forging your own identity, how will you shift tone when a cultural moment demands it? A strategist embedded in your business notices when it's time to become who you're becoming. That instinct — not to be trapped by your own past — is something a brand needs and something an AI trained on historical data cannot provide unprompted.
Where this leaves us
AI is making email marketing more powerful, not less complex. The tools compress timelines, unlock personalisation at scale, and make sophisticated tactics accessible to brands that couldn't previously afford them. That's good for everyone. But power without direction is just activity. The question isn't whether the machine runs — it's whether it's running toward the right destination for your specific business right now, given everything that's true about your brand and your customers that no platform dataset can fully capture.
The agency is evolving from email production house to an essential backstop that serves your business in ways an algorithm alone cannot guarantee. The question isn't whether AI can write your emails. It can. The question is whether anyone is making sure they're the right ones, for the right people, at the right time, for the right reasons. That's a job for people who know your business well enough to care about the answer.