FAQs | Avoiding the hidden costs with AI in the contact centre.

Using an LLM on its own as a stand-alone AI language model is not the best solution for most contact centres.

The optimal approach for the nature and scale of a contact centre is using an “ensemble” of AI language models, combining LLMs alongside SLMs and MLMs. When such an ensemble is correctly orchestrated, each individual contact centre task can be handled by the language model that is most efficient and suitable.

LLMs should not be treated as a “one-size-fits-all” solution in the contact centre. While LLMs are capable of that, and do have a critical role to play in the contact centre, their size and complexity is inefficient for the majority of contact centre tasks, and therefore not cost-effective.

If an LLM is the sole AI language model deployed in the contact centre, both operational and commercial challenges will emerge, and the solution will struggle to achieve ROI.

In high-volume, structured environments like the contact centre, AI model architecture requires a combination of an LLM carefully orchestrated with more precise, smaller language models, with each handling the appropriate tasks and workflows.

If organisations rely primarily on LLMs alone for structured, high-volume contact centre tasks, they’re likely to introduce unnecessary operational costs, inefficiencies, and complexity.

Using frontier-scale LLMs for simple, repeatable, deterministic tasks will unnecessarily increase compute costs and energy consumption.

Using LLMs as a one-size-fits-all solution will also have negative consequences for customer experience, through problems such as excessive latency, increased average handling time, and lower first contact resolution.

These risks can be avoided by incorporating smaller task-specific language models to handle those simpler contact centre tasks, rather than using frontier-scale AI when it isn’t required.

Using frontier-scale LLMs in use cases where smaller, more efficient, task-specific language models would be more suitable could plausibly introduce between £200 and £500 million per year in additional operational costs.

In certain contact centre use cases, a frontier-scale LLM can also consume up to 95% more energy than a task-specific micro-language model.

The actual impact for each individual organisation will depend heavily on architecture, scale, model usage, and operational design.

At a national scale, using frontier-scale LLMs where small language models or micro-language models would be more efficient and suitable could introduce £1 billion or more in additional annual costs for the UK contact centre industry.

Every inbound call coming into your contact centre should first pass through some form of voice AI triage, before the system decides whether the interaction can be automated or routed to a human agent.

At UK scale, this means billions of interactions passing through AI systems every year, even if only a portion of those calls are ultimately automated.

If your AI solution relies on an LLM as a “one-size-fits-all” solution, every task will be handled by the most powerful, energy-intensive, and expensive type of AI model available.

However, LLMs are not the most suitable language model for the majority of contact centre tasks.

Using LLMs for high-volume, structured, deterministic tasks that require speed, precision, and accuracy will result in unnecessary computing costs. This comes from using, and paying for, frontier-scale LLMs for workloads that are better handled by smaller, less complex, more cost-efficient AI language models.

We advocate for a multi-model AI architecture, taking a right-sized AI strategy.

SLMs and MLMs should be used alongside LLMs in a multi-model “ensemble.”

This ensemble of language models should be carefully orchestrated, with each individual task being handled by the most suitable, efficient language model for that task.

• MLMs are suited to atomic, precision-based tasks
• SLMs are suited to orchestration and structured workflow management
• LLMs should be invoked selectively when deeper reasoning or contextual interpretation is required.

This AI architecture of multiple language models, with intelligent orchestration, will help achieve the required operational and commercial performance with AI in the contact centre.

A language model is an AI model trained to understand and generate human language by learning patterns across large volumes of text data.

A parameter is a learned weight inside the AI language model, used to measure its capacity.

More parameters generally mean that a model has greater expressive capacity and deeper reasoning ability. However, higher parameters also correlates with greater computational demand, energy consumption, and costs.

Models differ primarily by scale and specialisation:

• LLMs typically exceed 20 billion parameters
• SLMs operate in the low billions
• MLMs are under 100 million parameters and are trained for highly specific tasks, such as intent detection or entity extraction.

Every interaction with a language model is broken into “tokens” (fragments of words and numbers). The more context you provide to the model, the more tokens it must process.

Larger language models make each of those tokens more computationally expensive. At scale, that combination translates directly into infrastructure cost and energy consumption.

AI implementation in the contact centre should be approached as a strategic commercial initiative, not a purely technical one.

That means AI model selection and architecture should be aligned with the strategic outcomes required from the contact centre.

Contact centres are measured in an organisation based on a set of strategic outcomes:

• Cost-to-serve (CST)
• Average handling time (AHT)
• First contact resolution (FCR)
• Containment rate
• Customer advocacy (CSAT / NPS)
• Compliance and operational risk
• ESG and sustainability performance.

Selection and architecture of AI language models will have a direct influence on each of these outcomes, which means it has a strong influence on commercial performance. So, stakeholders such as the CFO, the COO, the CX director, and even ESG leads should be involved in the decision of how AI is implemented in the contact centre.

More specifically, a commercially successful approach combines typical LLMs with smaller task-specific language models, ensuring each individual contact centre task is handled by the most efficient size of model.

The objective in this decision making should be to ensure the organisation only pays for frontier-scale AI when it’s genuinely required, otherwise unnecessary operational costs will begin to erode the ROI.

When AI model architecture is decided purely based on technical criteria, in isolation from the rest of the organisation, it often leads to an AI deployment that fails to deliver on its commercial promise, even though it may be technically sound.

To identify these pitfalls in the contact centre, assess whether the chosen architecture improves core contact centre metrics in proportion to the cost and complexity it introduces.

A contact centre’s AI solution should be having a positive effect on metrics like:

• Cost-to-serve (CST)
• Average handling time (AHT)
• First contact resolution (FCR).

Using frontier-scale LLMs as a one-size-fits-all solution will have a negative impact on these metrics due to the scale, nature, and complexity of the language model in proportion to the contact centre tasks its handling.

When contact centre automation scales, marginal inefficiencies in precision, latency, energy consumption, or governance discipline compound quickly.

The impact typically emerges across a small number of primary drivers, each of which maps directly to measurable business performance.
The key principles of a successful contact centre AI solution are:

• Precision
• Latency
• Energy efficiency
• Governance
• Orchestration and task-specific usage.

Good examples of where a SLM or MLM may be more appropriate and efficient than LLMs include identity verification, balance enquiries, appointment changes, and policy updates.

These are structured, deterministic workflows that require speed, precision, and reliability, rather than open-ended reasoning.

For example, with identification and verification (ID&V), LLMs can perform the task, but they’re not optimised for atomic precision. If one digit is misinterpreted in a 10-digit account number, the system will re-prompt. This repetition of the task with an LLM increases average handling time, and can reduce containment and first contact resolution, each of which will have a negative impact on the customer experience and the operational costs of the AI model.

Task-specific MLMs, conversely, are purpose-built for this workload.

Translated commercially: Fewer re-asks, fewer escalations, higher containment, and more predictable first contact resolution.

MLMs handle atomic, precision-based tasks such as intent classification, entity capture, validation, and ID&V. These models are optimised for speed, determinism, and cost efficiency.

SLMs manage workflow orchestration, routing logic, summarisation, and structured journey management. They coordinate the interaction, determine next best action, and ensure procedural consistency.

LLMs should be invoked selectively when deeper reasoning, contextual interpretation, or emotional nuance is required.

An F1 score is a combined measure of precision and recall for AI language models.

In benchmark testing, specialist intent and entity AI language models routinely achieve F1 scores in the mid-90s.

F1 scores above 90% indicate strong production reliability of AI contact centre solutions.

As a point of reference, Connect’s intent models, powered by Elerian AI, operate at approximately 96% F1 with inference times measured in milliseconds.

Yes, SLMs and MLMs usually provide faster response times for structured contact centre workflows.

Smaller language models generally process requests faster because they require less compute per token.

How does this work in simple terms?

As these language models require less open-ended reasoning and context processing than LLMs, they can complete tasks with greater precision and speed. This results in faster, more accurate interactions for the customer.

Even small delays in voice automation can disrupt natural turn-taking in these interactions. Additional seconds during authentication or workflow confirmation increase average handling time (AHT).

In structured customer journeys, that speed advantage translates directly into shorter interactions and a more efficient, satisfying customer experience.

The larger the AI language model, the greater the computational demand and energy consumption.

Energy benchmarks illustrate the scale of the difference:

• MLMs: ~0.1–0.4 watt-hours per million tokens
• SLMs: ~1–4 watt-hours
• LLMs: often 10–100+ watt-hours

In commercial terms, using a frontier large language model (LLM) for a high-volume, low-complexity journey can consume up to 95% more energy than a task-specific micro-language model.

Energy consumption has a direct impact on operational costs for the AI solution.

More specifically, at scale, this will have a negative impact on cost-to-serve, margin performance, cost predictability, ESG, and sustainability commitments.

SLMs and MLMs can often provide organisations with greater infrastructure control and governance flexibility than LLMs.

Frontier LLMs are typically accessed via shared cloud APIs. While often secure, this introduces external dependency and limits infrastructure control.

SLMs and MLMs can often be deployed within your own environment on-premise, private cloud, or tightly controlled virtual private cloud VPC environments, because they require significantly less computational resource.

For highly regulated industries, this can support:

• Data sovereignty
• Reduced external data exposure
• Stronger auditability
• Alignment with regulatory frameworks.

Governance directly influences compliance risk and brand protection, so it should be a priority in the decision making when designing and architecting contact centre AI solutions.

Mature organisations carefully orchestrate an “ensemble” of different sized AI language models in the contact centre.

Rather than relying on frontier-scale LLM to handle every task, mature AI strategies distribute workloads, according to task nature and complexity, to smaller, more efficient language models.

LLMs should only be used for tasks that require the highest level of reasoning and non-deterministic capabilities.

SLMs and MLMs should be used for simpler, structured, deterministic tasks, which reflects the majority of contact centre workloads.

This is where capital efficiency emerges.

Instead of paying frontier-scale compute costs for every single customer interaction, organisations only pay for deeper reasoning and context processing from LLMs when it’s actually required.

ROI and success with AI in the contact centre may be determined by whether stakeholders concerned with that commercial performance are in the room when AI model selection and architecture decisions are made.

From the customer’s perspective, a multi-model, right-sized AI approach should still feel seamless.

Some key benefits include:

• Faster average handling time (AHT)
• Improved containment and first contact resolution (FCR)
• More efficient customer journeys.

But from the organisation's perspective, every single interaction is completed as efficiently as possible in terms of energy consumption and computing costs.

The investment in AI should be measured by commercial outcomes, not technical capability or parameter count.

Using the appropriate AI model for the appropriate workload can help organisations:

• Improve containment and first contact resolution
• Reduce unnecessary latency and operational inefficiency
• Improve cost effectiveness and infrastructure utilisation
• Reduce energy consumption and support ESG goals
• Strengthen governance and operational resilience
• Enhance customer experience and customer satisfaction.

Connect advocates for a right-sized AI strategy that aligns language model selection and AI architecture to commercial goals and outcomes.

A right-sized AI approach brings together MLMs, SLMs, and LLMs in a carefully orchestrated ensemble, and uses the context of the interactions to route each individual task to the most efficient language model.

The objective is not to eliminate LLM usage, but to ensure organisations only use (and pay for) frontier-scale AI where it’s required. This is the key to achieving the desired ROI and commercial performance with contact centre AI solutions.

Get in touch with Connect to discuss how a multi-model AI architecture approach could help optimise your contact centre AI strategy.

The biggest misconception regarding AI in the contact centre is that AI language model selection and architecture should be a technical decision made only by the technical team.

In reality, however, AI model architecture should be a commercial business decision, aligned with the commercial outcomes required from the contact centre.

Selection and use of AI language models will influence key contact centre metrics that drive commercial performance, such as:

• Cost-to-serve (CTS)
• Average handle time (AHT)
• First contact resolution (FCR)
• Customer satisfaction (CSAT)
• Energy consumption
• Data sovereignty.

These metrics have direct regulatory and commercial consequences for the organisation. Therefore, stakeholders such as the CFO, the COO, the CX director, and even ESG leads should be involved in the decision of how AI is deployed in the contact centre.

When AI model architecture is decided purely based on technical criteria, in isolation from the rest of the organisation, it often leads to an AI deployment that’s unable to deliver on its commercial promise, even if it is technically sound.

In those cases, the most common selection is a frontier-scale large language model (LLM) deployed as a general-purpose, one-size-fits-all contact centre AI solution. However, this is a problematic approach.

Using LLMs alone in this way creates a hidden layer of cost for the contact centre that quietly erodes the return on AI investment.

This is known as the hidden “tax” of LLMs, which many organisations are unknowingly paying as a result of making technically-focused AI architecture decisions, rather than commercially-focused ones.