Under pressure from increasing client needs, as well as cost and talent, many service organizations buy the most hyped and powerful AI on the market, only to watch it underdeliver. The reason is rarely the technology. It is the missing fuel: a well-defined target picture, data readiness, and system architecture that matches the scale. Learn from our Simon-Kucher Elevate experts why the ‘best’ AI technology is so often the biggest risk in customer service, and what works instead.
Let’s start with a common scenario for many leaders at customer service and contact center organizations: As in most companies, your cost pressure is relentless – qualified staff are harder to find every year, and customer expectations, shaped by the seamless experiences of the tech giants from Netflix to Amazon, have never been higher.
So imagine: To tackle these problems, you make a decision that feels both bold and prudent – you invest in the gold standard, the “Ferrari” among AI systems.
Then comes the disillusionment. After launch, the machine fails to deliver the efficiency and impact you were promised, not because the technology is faulty but because it is starving. You have dropped a high-performance sports-car engine into an environment whose data volume cannot supply the “octane” it needs: the critical mass of training data required for the system to synchronize with your content and your reality on the ground.
Why the biggest LLMs hit their limits fastest in service
Here is the crux. Many AI solutions in customer service are built on large language models (LLMs) with billions of parameters. Using such a system for a simple address change or a duplicate-invoice request is the technological equivalent of driving a Ferrari 400 meters to the bakery:
- Cost: For trivial, high-volume tasks, the token and computing expense is simply not economically viable – and it grows with every contact you automate.
- Reliability: Paradoxically, a highly intelligent system is more prone to hallucination on simple factual requests than a precisely defined, rule-based workflow. The reason lies in how LLMs work: they do not look up answers, they predict the most plausible text. A rule-based workflow retrieves the account balance from the database and is right every time; an LLM will occasionally generate a fluent, confident, wrong one.
This is not an argument against LLMs – they excel where language itself is the problem, from understanding an unstructured complaint to supporting an agent on a difficult call. It is an argument against using them for everything.
Overengineering, in 2026, is no longer the exception; it is on its way to becoming the default risk companies face. And the cause is rarely the technology itself. It is almost always a missing clarity on target picture and focus areas: which tasks, exactly, should the AI solve, for what volume, and on what data?
Octane decides: why AI cannot work without clear data architecture
This is the real blind spot. Most companies invest in technology and models, but not in the fuel. An LLM is only as good as the data it works on, and in the reality of many service organizations, exactly what the system needs to function reliably is missing:
- Knowledge bases are outdated, fragmented, or insufficiently structured.
- CRM data is patchy: customer records and interaction histories are incomplete.
- Tickets are poorly maintained or unlabeled, and customer context sits in silos.
- Process data, policies, and product information exist in formats no model can use without substantial preparation.
Just how real this gap has become was clear in a rapid audit our Simon-Kucher experts recently ran for an energy provider: 43% of the most important cost-driving KPIs in its customer service were not being measured at all, for sheer lack of a data foundation. But a business that cannot even see its own steering metrics has no foundations on which any AI, however powerful, could reliably operate.
The consequence is predictable. Without this depth of data, even the most capable LLM produces poor answers. It hallucinates where it should look up. It escalates where it could resolve. It generates loops instead of closures. “Intelligent automation” becomes an expensive text generator, one that creates more rework at Level 2 than there was before any AI at all.
On top of this comes a scale mismatch that many underestimate: the technology must fit the actual transaction volume. A company with 10,000 service contacts a month needs a fundamentally different architecture than one with 500,000. Deploy an enterprise-grade system for small volumes, and you not only overpay – you also get worse results, because the system never sees enough data to learn. Recognition rates stay low, resolution rates disappoint, and the promised ROI never arrives.
Lessons learned on the ground
We have watched this play out many times in our projects and discussions with clients. For example, one service organization invested heavily in a leading enterprise AI platform, and ultimately had to switch it off because its transaction volume was far too low to ever train the system to useful accuracy. The team went back to first principles, built a right-sized solution matched to its real volume and data, and today resolves 80–90% of relevant cases automatically. The CEO had wanted a Ferrari. What the business needed was the right car for its road.
The correct sequence, therefore, is this: first a clear target picture, then the data foundation, and then the tool, never the other way around. Ignore that order, and you have bought a Ferrari but filled the tank with tap water.
Liquid expectations and the danger of deliberate friction
Customer expectations today are liquid. They are set not by your industry, but by the best digital experience your customer has had anywhere: Amazon, Uber, a favorite app. And when an interaction disappoints, switching is one click away. Across industries, between 45% and 70% of consumers leave a provider after poor service experiences.
Too often, technology, and AI in particular, is misused as a digital firewall: deliberate friction designed to keep customers away from costly human contact. The result is trapped customers, caught in chat loops. It is the most short-sighted form of cost optimization imaginable because it erodes the foundation of every brand: customer satisfaction and trust.
More than 90% of customers expect their issue resolved on first contact – and with every additional interaction it takes, not only does their patience drain away, but their satisfaction and their customer lifetime value with it.
This is precisely why the data question is not an IT topic but a matter of business strategy and customer value. Deploy AI on a flawed data foundation and you do not merely create inefficiency – you actively manufacture friction. And in a market of liquid expectations, friction is the most expensive risk there is.
The 12-minute opportunity: service as a sales lever
In sectors such as energy, insurance, banking or telecommunications, the direct interaction time per customer often adds up to just ten to twelve minutes a year. In practice, most of that is consumed by routine matters (address changes, meter readings, billing queries), leaving only one or two minutes for genuine value creation.
And here lies a lever that many companies underestimate: highly efficient contact center and customer service processes are not only cost factors, but also powerful sales and retention channels. They create the room to make customers individual, precisely targeted up- and cross-selling offers, without it ever feeling “like sales.” The precondition is a strategically designed service architecture beneath it all.
But the decisive point is this: the lever only works if the routine cases are genuinely, reliably automated. Which brings us back to the data question. If the knowledge base is patchy, if CRM data is not available in real time, if process logic was never cleanly translated into automation, then those twelve minutes do not become “eleven minutes of routine plus one minute of value.” They become ten to twelve minutes of frustration.
H2: Value-based empathy: no AI hold loop when it matters
Innovation here means using real-time data to decide which case needs a human immediately, and which can be resolved quickly and digitally. In practice, a two-level structure proves itself.
Level 1 (digital and automated): Standard cases in customer service run through highly efficient and intuitive processes – enhanced with intelligent software solutions. Here, some 85–95% of requests are resolved directly and automatically, with minimal effort for the customer and minimal cost. The precondition: clean client and process data, a well-maintained knowledge base, and the right tool matched to the right task.
Level 2 (human, with an AI copilot): When a customer is frustrated, or a case turns business-critical or genuinely complex, a human should take over at once, supported by context data and real-time AI recommendations. This is the remaining 5–15% of cases, where human judgment makes the difference. Here the service contact becomes a moment of loyalty, and often the starting point for upselling – customer experience not as a sales pitch, but as genuine client support.
What the right architecture makes possible
Resolve the Ferrari Paradox by matching technology to the problem, the contact volume, and data, and a whole set of applications moves from PowerPoint promise to working reality. Each rest on the same precondition this article has discussed throughout: data maturity, contact volume, and target picture for using AI are aligned.
The two foundations for the model are the two levels themselves:
- On Level 1, conversational AI and self-service turn intuitive portals into genuine 24/7 resolution, as long as the underlying knowledge base and data are structured, current, and complete. Without it, the bot produces hallucinations instead of answers.
- On Level 2, an AI copilot on the call adds live sentiment analysis and proactive response suggestions, turning the agent into a strategic advisor, but its quality stands or falls with the depth of CRM data and interaction history. Without context, the copilot is blind.
Around these sit three more important factors for successful AI in service:
- Intelligent case routing classifies incoming requests and sends the complex ones straight to the expert with the right skill profile, instead of losing the customer in a phone menu or in the IVR. It works only on cleanly labeled tickets and structured process data; cut corners here, and you have simply moved the problem from the hold queue to the misrouted ticket.
- Demand forecasting draws on market and weather data to anticipate request peaks and reach customers proactively, before they ever pick up the phone, which means looking beyond the classic CRM horizon for data.
- Next best action delivers offers timed precisely to the moment in the customer journey, closing the circle back to the twelve-minute opportunity: only when the routine is automated and the customer context is live does the room open for the right offer at the right moment.
The thread connecting all five is the one we began with. None of them is a product you buy. Each is a capability you earn by getting the fuel right first.
Conclusion: the end of the prestige investment
AI excellence in customer service in 2026 is not only a software purchase. It is a deep, strategic architecture decision. Avoid the scale mismatch, deploy empathy where it actually works, and service becomes your strongest growth driver, lifting revenue through up- and cross-selling that finally feels like service rather than sales.
The Ferrari in the garage is impressive. But the race is won by the right car, with the right fuel, on the right track.
Want to make your service architecture future-proof? Talk to our Simon-Kucher Elevate experts.
