By manual customer segmentation, I mean the practice of manually dividing a customer base into distinct groups based on shared characteristics, such as demographics, psychographics, geography, and behaviour. In small and resource-constrained organisations, where I have worked, this almost always boils down to segmenting customers using transactional data—recency, frequency, and monetary value (RFM) of transactions—because that is the only (more or less) reliable data they have on hand.
Imagine a sales manager poring over spreadsheets, sorting customers into buckets—“purchased within the last 12 months”, “1-2 years ago”, “more than 2 years ago”—then further dividing by purchase count and total spend. It is a process rooted in experience, intuition, and Excel formulas—a respectable process even in the age of AI. But it is, at its heart, manual. The boundaries, such as the 12-month cutoff for “high-value” customers, are chosen for their convenience, not for their analytical rigour.
Think about it. What makes a customer who bought 11 months ago distinctly different from one who bought 13 months ago? Probably not much. Yet, these arbitrary divisions guide marketing campaigns, promotions, and product launches.
The Case Against Manual Segmentation
The internet is teeming with comments on why manual customer segmentation is ineffective. If you type “Why is manual customer segmentation ineffective” in the Google search bar, the AI will respond with a comprehensive list of reasons why. While I wouldn’t repeat each and every one of them, here are what I see as the most important ones to consider:
Time-consuming and labour-intensive
Data teams, who are often charged with this job, frequently find themselves trapped in the Sisyphean task of neverendingly tweaking spreadsheets to make stakeholders happy. This, which does feel like a punishment from time to time, is not just a drain on time. It is a misallocation of talent that could be spent uncovering meaningful insights.
Biased and subjective
Segmentation boundaries are drawn from subjective experiences, guesstimates, or my favourite (I’m being sarcastic)—inherited templates. Hardly evidence-based, they reinforce rather than challenge assumptions. Biased segmentation overlooks or stereotypes customers, leading to inaccurate targeting, irrelevant messaging, and missed opportunities.
Limited human insights
I’m not one to claim machines are always superior to humans. Humans possess unique, wonderful qualities, such as context awareness, that machines do not (at least yet). Machines do excel in certain areas, however. Customer segmentation is one of them. No matter how meticulous we humans try to be, we struggle to find intricate relationships in data. We fail to notice complex behaviour and hidden motivations, which machines could help us uncover.
Unscalable and static
Manual segmentation relies on criteria that do not update dynamically. As organisations grow, manual segments buckle under the weight of new data and evolving customer behaviour. Customers who are in the “purchased within the last 12 months” bucket now might bear little to no resemblance to those who were in the same bucket a year ago. The sales manager, nevertheless, continues to give these customers the same treatment, sending messages that are irrelevant to them.
Manual customer segmentation, while familiar and seemingly straightforward, is riddled with limitations that are incompatible with deeper personalisation—something customers take for granted in today’s increasingly competitive market.
Clustering: A Better Alternative
Machine learning—specifically, clustering—offers a more dynamic, evidence-based approach. Algorithms like k-means or hierarchical clustering do not force customers into predefined buckets. They reveal natural groupings based on patterns that might otherwise go unnoticed, adapting as new data arrives. With clustering, data teams can move from reactive slicing to proactive discovery, and segments derived more likely correspond to meaningful business differences in terms of loyalty, churn risk, and lifetime value. Running k-means clustering on RFM data has become increasingly common and can reveal who are “champions”, “potential loyalists”, “at risk”, and so on—labels grounded in data patterns, not guesswork.
Clustering is not merely a technical upgrade. It is a strategic unlock leading to campaigns that resonate, services that adapt, and decisions that reflect the true diversity of customers they serve.
Why Manual Segmentation Persists
If there are better alternatives out there, why does manual segmentation persist?
One obvious answer has to do with capability and capacity. Small organisations rarely have dedicated in-house data science teams. Even when data and insights specialists are present, their skills may not extend to advanced analytics or machine learning. The default is manual segmentation—not necessarily for lack of imagination but for lack of capability and capacity.
With some imagination, this obstacle can be overcome. Organisations can engage external agencies or consultants with experience in machine learning. While outsourcing might sound expensive, it shouldn’t have to. Many agencies or consultants would be happy enough to tailor offerings for smaller budgets. Alternatively or additionally, investing in staff training and upskilling—through online courses, workshops, or mentoring—can boost analytics capability.
Ultimately, the barrier is less about capability and capacity. The truth is more uncomfortable—much more so.
When Clustering Fell Flat
A few years ago, I built a clustering-based customer segmentation model I was genuinely proud of. It was methodologically sound and full of potential for strategic insight. I presented it with quiet confidence, expecting stakeholders to be intrigued—maybe even excited.
They weren’t.
The model was politely acknowledged and then silently shelved. Instead of adopting my model, the team stuck with manual segmentation methods. I just shrugged and chalked it up to the usual suspects: status quo bias, resistance to change, and lack of data literacy. They just didn’t get it, I told myself.
In hindsight, it wasn’t their fault. It was mine.
I hadn’t explained the model clearly. I hadn’t walked the team through the logic, the assumptions, and the implications. I hadn’t translated the clusters into human terms or connected them to strategic decisions. I had delivered insight without scaffolding—and expected buy-in without trust.
That experience taught me something humbling: Technical excellence means nothing if it’s not accompanied by narrative stewardship. If we want our work to be understood, adopted, and used, we have to do more than build. We have to translate.
The Comfort of Explainability
Manual segmentation persists not because it is more accurate but because it is more explainable. Stakeholders can trace the logic, justify the groupings, and defend the decisions in a way that feels safe and familiar. Machine learning, by contrast, is a black box. Its internal decision-making processes are not readily interpretable or transparent. It asks stakeholders to take the plunge with algorithms that operate beyond their intuitive grasp.
This isn’t just status quo bias or resistance to change. It’s a rational response to risk. When we fail to explain clustering clearly, we inadvertently reinforce the perception that it’s “too complex”, “too academic”, or “not ready for prime time”.
To shift this mindset, we must make clustering explainable. We must:
• Translate the math into meaning, illuminating what each cluster represents in human language and linking them to strategic priorities and business outcomes.
• Show the journey, not just the destination, taking stakeholders through the steps taken to arrive at the groupings.
• Invite dialogue, not defensiveness, engaging stakeholders in the interpretation of clusters, encouraging questions, and exploring the model’s limitations and opportunities together.
When we do this well, clustering becomes not just a tool—but a shared lens. And that’s when real adoption begins.
From Frustration to Stewardship
Moving from manual segmentation to clustering is as much a cultural journey as a technical one. And like any shift, it requires trust, translation, and time. If we want our stakeholders to embrace new ways of seeing their customers, we must first meet them where they are: in the comfort of manual segmentation—in the logic they understand and in the fears they may not voice.
That doesn’t mean we dilute our expertise. It means we steward it. We narrate the process, illuminate the patterns, and connect the dots to strategy. We stop blaming resistance and start designing for adoption.
In the end, data is not about models or algorithms. It’s about enabling decisions that matter. That only happens when people believe the story the data is telling—and when they trust those who tell it.
Notes
See, for example, Hanas, A. (2023, November 14). RFM analysis for customer segmentation using K-means. NeuroSYS. https://neurosys.com/blog/rfm-analysis-for-customer-segmentation-using-k-means; Kehinde, O. (2023, January 3). Customer segmentation using RFM analysis and K-means clustering in Python. Medium. https://medium.com/@omotolaniosems/customer-segmentation-using-rfm-analysis-and-k-means-clustering-in-python-d9d90fc24e18; and Rihaldijiran, V. (2025, January 16). RFM segmentation: Unleashing customer insights. Towards Data Science. https://towardsdatascience.com/rfm-segmentation-unleashing-customer-insights-da58deae4eb9/
Autonomous Author – Senior Data Professional in Australia
From Data to Strategic Action: Why Most Companies are Stuck at the Bottom of the Value Chain
