The Hidden Costs of Building AI for CX

The Hidden Costs of Building AI for CX

James Darrall James Darrall

James Darrall

Global Retail Industry Lead at Zendesk | AI at Oxford

Published Jun 11, 2025
+ Follow

AI is transforming customer experience — but building it is not as simple as flipping a switch. Building AI yourself means grappling with massive data needs, complex model choices, ongoing human oversight, and costly maintenance. Meanwhile risk increases, customer expectations rise and competitors invest in true innovation. The real challenge is execution — navigating complexity without losing focus on what truly drives growth.

📈 More Data, More Problems

Every AI model is only as good as the data it’s trained on — and most teams underestimate how much is actually needed. In CX, that means billions of real, multi-intent conversations captured consistently across channels.

🔢 It’s Not Just Volume — It’s Complexity

The volume required depends on the problem type, model complexity, and number of features. A common rule of thumb? You need at least 10× more examples than model parameters — and thanks to Bellman’s curse of dimensionality, each additional feature exponentially increases the data needed (Bellman, 1961). Take a simplified example: 120,000 chat conversations with just 29 binary features (e.g. positive–negative, refund–purchase, sofa–bed, customer-prospect, english-french etc.) already produce over 5 million possible combinations.

Applying the 10× rule, you’d need 50 million labelled conversations to train a reliable model. And that’s before accounting for real-world variables like hundreds of languages, millions of SKUs, layered intents, shifting sentiment, and multiple customer segments.

📬 Real CX Data Is Messy

Customer experience data isn’t neat or uniform — it spans multiple channels all with their own nuances (email, chat, WhatsApp), fluctuating sentiment, edge cases, and multi-intent conversations that don’t follow predictable patterns.

Before any training can begin, teams need to collect it from every system, clean it for consistency, deduplicate noisy records, scale and format it for model compatibility and split it into training and testing sets whilst ensuring balanced coverage across every scenario the model might encounter.

🧪 Rubbish In, Rubbish Out

There’s also hidden bias risk: for example if your data over represents peak-season interactions (e.g. winter sales), your model may lean disproportionately toward patterns from that period — skewing forecasts and undermining reliability. As Sandra Stanley Chief Data & Science Officer at dunnhumby, puts it (Retail Week, 2025):

“AI without good data is like a car without fuel."

Even massive datasets fail when they’re poorly labelled, inconsistent, biased or unbalanced. And while your engineers are stuck debugging models or correcting false positives, they’re not building what actually drives customer value — like subscription flows, wellness journeys, or personalised recommendation engines.

🛤️ Choosing the Right Model

AI isn’t one-size-fits-all. Selecting the optimal model type depends on the data available, the complexity of the task, and business goals.

  • Supervised learning excels when with labelled data. Tasks like CSAT prediction, ticket classification, and volume forecasting - historical data guides the model.

  • Unsupervised learning shines with unlabelled data. It uncovers hidden patterns through clustering, identifies anomalies, and surfaces emerging trends.

  • Reinforcement learning uses trial and error to optimise, making it powerful for adaptive bots. However, it demands substantial compute and careful design, making it expensive and complex to train effectively.

The real value lies in matching the right approach to the right problem — and that requires informed human judgment. Too often, teams default to whichever model is easiest or fastest to deploy, rather than the one best suited to the problem.

This shortcut leads to suboptimal performance, wasted investment, and AI that ultimately fails to deliver meaningful business value.

⚠️ The Human Cost

AI isn’t a one-and-done effort. Humans aren’t just needed to collect and prepare data or choose a model and walk away — they’re needed throughout the lifecycle. Training a model typically takes 6–9 months before deployment. And even then, the work isn’t over...

Models need to be monitored, maintained, and continuously refined to remain accurate and relevant. What worked six months ago might not work today. Data drifts, biases creep in, and performance degrades if left unchecked (Sculley et al., 2015).

💬 It’s About People, Not Just Performance

This isn’t just about operational cost — it’s about the employee and customer experience. As B&Q 's Andy Moat put it (Retail Week, 2025):

"No growth strategy, tech investment or innovation can compensate for poor human interaction"

That’s why the best technology doesn’t just serve customers — it empowers teams. Tesco's Technology Director Jane Mustoe shared at the World Retail Congress that their investments prioritise tools that put people first, giving colleagues more time to deliver better service (Retail Week, 2025). And the payoff is clear. As Andy Moat said:

"As retention goes up, sales go up, churn goes down, cost of hiring and onboarding goes down."

The opposite is also true. Poorly designed systems and ineffective AI increase frustration, damage trust, and erode team morale — driving up churn, recruitment costs, and ultimately lost revenue.

💼 The Opportunity Cost

Most companies don’t need to build conversational AI. They need to integrate to it — securely, scalably, and without distracting from what truly drives growth.

The goal isn’t to reinvent foundational tools — it’s to free up time and talent to build what sets your brand apart. Initiatives such as:

  • Pets at Home launched Easy Repeat, a subscription platform now onboarding 1,000 customers daily — increasing frequency by over 50% (Retail Gazette, 2025).

  • Holland & Barrett released H&B&Me, a wellness app using biological age data and behavioural nudges. 87% of trial users reported major gains in energy, mood, and sleep (Retail Systems, 2025).

Avon slashed 70% of SKUs to refocus on personalised beauty and omnichannel innovation (Reboot Chronicles, 2025). As Avon CEO Kristof Neirynck recently said:

“You have to focus where you can win"

The takeaway? Build where it matters. Cut what doesn’t. Buy what doesn’t differentiate.

⚙️ Complexity and Compute Don’t Scale Easily

Some problems aren’t just hard — they’re NP-hard (the toughest puzzles in computing). Take the Travelling Salesman Problem (TSP): a classic optimisation challenge that asks:

"Given a list of cities and the distances between each pair of cities, what is the shortest possible route that visits each city exactly once and returns to the origin city?"

It sounds simple. But every additional city exponentially increases the number of possible routes (Garey and Johnson, 1979).

Travelling Salesman Problem

That same combinatorial complexity applies in CX conversations. As you try to optimise across more variables — product, intent, sentiment, priority, channel, customer profile, language, agent availability — the routing logic becomes increasingly unmanageable. Every edge case adds exponential complexity.

🖥️ Scale Is Expensive — Really Expensive

And compute isn’t cheap. OpenAI (2018) found that training compute once doubled every 3.4 months; more recent analysis puts that rate closer to every six months (Sevilla et al., 2022).

Training a model like GPT-4 costs tens of millions — with frontier models projected to exceed $1 billion by 2027 (Epoch AI, 2024).

Trying to replicate that internally is like trying to build AWS before launching your app.

⚖️ Reputational and Legal Costs

The EU AI Act raises the stakes. Serious violations — such as deploying prohibited AI or breaching obligations for high-risk systems — carry fines of up to €35 million or 7% of global turnover (European Parliament and Council, 2024). If personal data is also mishandled, GDPR fines of €20 million or 4% apply on top (European Commission, 2016).

Worst-case scenario: €55 million or 11% of global turnover — not including lawsuits, disruption, or reputational damage.

🛡️ Build Alone, Risk Alone

In April, Marks and Spencer was hit by a major cyberattack. The breach is expected to cost £300 million and forced a rapid overhaul of M&S’s digital roadmap (Retail Week, 2025).

Under UK GDPR, data controllers like M&S remain accountable unless risks are contractually managed. Had the system been built in-house, M&S would have borne full liability and reputational fallout—with no vendor to share risk. CEO Stuart Machin called it:

“the most challenging situation we’ve encountered.”

The risk of going solo is stark: Australia’s Robodebt, an in-house automated system, led to $1.2 billion in compensation (Royal Commission, 2023). Without proper safeguards, building AI means accepting full liability. That’s why mature CX platforms offer shared responsibility, built-in security, and resilience, not just features.

💡 A Smarter Way Forward

Platforms like Zendesk remove the hidden costs of building AI by doing the heavy lifting for you — training on trillions of real CX interactions and embedding AI directly into the systems your teams and customers already use. It’s fast, flexible, and built for customer experience from day one.

As Zendesk CTO Adrian McDermott puts it:

“AI needs to be built into every channel, in every moment”

🔧 No-Code, All Control

Instead of manually scripting flows for every channel, modern AI tools can now interpret your text-based business policies to automatically generate resolution paths — aligned with your brand tone and service standards. This empowers business users to make changes directly, without waiting on developer queues or IT support — increasing agility.

⚙️ Integrate in Hours, Not Months

No-code integration capabilities allow teams to connect systems and deploy workflows in just a few hours. This doesn’t just accelerate time-to-value — it also frees internal teams to focus on building high-impact models, such as churn prediction, product recommendations, or personalisation engines, and makes it easier to embed those models seamlessly into the customer journey.

💸 Built for Outcomes

The economics are evolving too. Outcome-based pricing — where cost is tied to successful resolutions rather than seats or interaction volumes — is helping businesses align automation investment directly with measurable value (Futurum Group, 2025).

🧑🔧 Tools + Teams = Transformation

Crucially, Zendesk doesn’t just deliver tools. We deliver expertise. Because, as Andreessen Horowitz (2025) reminds us:

“companies that succeed don’t just ship software — they roll up their sleeves and help customers implement it. Enterprise AI demands forward-deployed teams, expert onboarding, and hands-on integration that turns models into outcomes”.

CEO Tom Eggemeier reaffirmed our approach: speed, safety, and scale — delivered through embedded AI, not stitched-together silos (Retail Week, 2025).

🔚 The Bottom Line

Customer service is no longer a back office — it’s your brand. And AI is fast becoming the front door. Building conversational AI from scratch means taking on the full weight of infrastructure, compliance, security — and the opportunity cost. It diverts focus from what actually drives growth.

The winners won’t be the ones rebuilding plumbing. They’ll be the ones doubling down on differentiation — not duplication. Because your next breakthrough won’t come from reinventing a generic CX model. It will come from building the experiences that set you apart — seamlessly integrated with a secure, scalable, and easily customisable platform like Zendesk .

References

CX, Reimagined CX, Reimagined

CX, Reimagined

1,197 follower

+ Subscribe

To view or add a comment, sign in

More articles by this author

See all

Others also viewed

Explore topics