Fortay Connect | All content and resources

AI Virtual Agent Implementation: Where Mid-Market Teams Should Start

Written by Fortay Connect | May 22, 2026 8:55:03 AM

The conversation around AI virtual agents in UK contact centres has shifted from "should we?" to "where do we start?" That is progress. But for many mid-market operations teams, the answer they land on is still the wrong one.

The default ambition is broad: automate as much of the contact centre as possible, reduce headcount dependency, and modernise the customer experience in one programme. It is an understandable position, particularly when vendor demonstrations make deployment look straightforward. In practice, that framing is exactly what causes AI virtual agent projects to stall, overspend, or deliver results that cannot be attributed to anything measurable.

The firms seeing genuine returns are not starting with replacement. They are starting with one narrow, repetitive, high-volume service journey, proving economic value within a defined timeframe, and then earning the right to scale. The sequencing is not a minor implementation detail. It is the difference between a project that builds internal confidence and one that becomes a cautionary tale at the next budget review.

Key takeaways

  • Most AI virtual agent programmes underperform because scope is set before the business case is stress-tested
  • The strongest first use cases are repetitive, high-volume journeys with clear intent patterns and low regulatory sensitivity
  • Cost-to-serve and labour compression are more reliable success metrics than headline automation rates
  • A phased rollout reduces delivery risk and creates faster, more credible proof of value than a replacement-first strategy
  • The hidden costs of integration, data clean-up, and governance typically exceed licence fees in year one

Why the Wrong Starting Point Kills AI Virtual Agent Projects

"Organisations that capture significant value from AI usually start with narrow, high-impact use cases and scale in stages with clear metrics." — McKinsey, State of AI 2025

The ambition to replace large portions of customer service with AI is not irrational. The economics are compelling: cost per interaction for human agents runs at roughly £5 to £6, while AI-handled interactions can cost a fraction of that at scale. But the jump from that observation to a broad replacement programme skips several layers of operational complexity that determine whether those economics ever materialise.

Broad programmes fail for predictable reasons:

  • Scope is set on aspiration, not data. Without a baseline for current handle time, escalation rate, and cost-to-serve by journey type, there is no way to prioritise where AI will have the most impact or to measure whether it has worked.
  • Knowledge quality is underestimated. AI virtual agents are only as good as the information they can access. Fragmented knowledge bases, inconsistent policies, and undocumented exception handling are the most common reasons early deployments disappoint.
  • Stakeholder expectations diverge quickly. When the business case is built around broad automation, every delay or escalation becomes evidence of failure rather than a normal part of calibration.
  • Integration complexity is compressed. Connecting an AI virtual agent to CRM, billing, order management, and authentication systems takes time. Promising full-service automation before those integrations are scoped creates delivery risk that compounds.

Mid-market firms are particularly exposed here. Unlike large enterprises, they cannot absorb a twelve-month experimental cycle with uncertain outcomes. Only 20% of mid-market companies successfully scale AI deployments beyond pilot stage, and 60% of AI projects that attempt rapid scaling fail outright. The firms that do reach scale almost always started narrower than they originally planned.

Where AI Virtual Agents Should Fit First

The strongest first use cases share a common profile: they are repetitive, high-volume, low-ambiguity, and operationally well-understood. The AI virtual agent does not need to be impressive. It needs to be reliable in a bounded context, and it needs to remove enough workload that the impact is measurable within weeks, not quarters.

The right candidate journeys

The journeys that consistently deliver fastest time-to-value are those where intent is predictable, data is accessible, and the consequence of a poor interaction is low. These include:

Journey type

Why it works for AI first

Order status and tracking

High volume, clear intent, data is usually CRM or OMS-connected

Appointment booking and changes

Structured workflow, low ambiguity, containable without escalation

Billing and payment queries

Repetitive intent patterns, self-service resolution is often sufficient

Password resets and account access

Fully automatable, no regulatory sensitivity, immediate resolution

Simple service requests and FAQs

High containment potential, knowledge base-driven, easy to measure

Journeys involving complaints, vulnerable customers, complex financial decisions, or regulatory obligations are not good starting points. Not because AI cannot eventually assist with them, but because the cost of a poor interaction in those contexts is disproportionately high and the governance requirements are significantly more demanding.

The productivity test

The question to ask of any candidate journey is not "can AI handle this?" but "does AI removing this journey eliminate a meaningful step in the operational process?" BCG research on AI agent deployment is clear on this point: the biggest productivity gains come when AI eliminates entire workflow steps, not when it simply reduces wait time or deflects a small proportion of contacts.

A virtual agent that handles 40% of order status queries but still requires agents to manage the remaining 60% through the same underlying process has not changed the operational model. One that handles 80% with clean containment, accurate data retrieval, and a well-designed handover for exceptions has materially compressed the labour requirement for that journey. That distinction determines whether the business case holds.

The Metrics That Matter More Than Headline Automation Rate

Automation rate is the metric vendors lead with. It is also the least useful one for building an internal business case. A 70% automation rate sounds strong until you discover that 30% of interactions are still escalating to agents, that the escalation handovers are poorly designed, and that customers are repeating themselves because context is not being passed cleanly. The headline number looks good. The operational reality does not.

Cost-to-serve versus cost per interaction

Cost per interaction is a useful starting point, and the gap between AI and human handling is significant. According to Freshworks data, cost per interaction drops from approximately £3.60 before AI implementation to around £1.15 after, when deployment is working well. But cost-to-serve is the more complete measure because it captures what happens across the full resolution journey, including escalations, agent rework, repeat contacts, and the human time spent cleaning up interactions the AI did not fully resolve.

A cheap AI interaction that generates a repeat call the following day is not cheap. It has simply moved the cost.

Labour compression as the right productivity measure

The more useful productivity metric for operations leaders is labour compression: the hours of human work displaced per 1,000 interactions handled by AI. This framing, supported by BCG's work on AI agent economics, shifts the question from "how much did we automate?" to "how much did we change the labour requirement?"

Metric

What it measures

Why it matters

Automation rate

% of contacts handled without human involvement

Useful but incomplete without containment data

Cost per interaction

Direct handling cost for AI vs human

Misses escalation, rework, and repeat contact costs

Cost-to-serve

Full cost of resolving a customer need end-to-end

Most accurate indicator of operational ROI

Labour compression

Human hours displaced per 1,000 AI interactions

Best measure of whether AI is changing the operational model

Containment rate

% of AI interactions fully resolved without escalation

Leading indicator of cost-to-serve improvement

The right target for a well-scoped first deployment: containment rates above 70% on the chosen journey, with cost-to-serve reduction visible within three months. Industry benchmarks suggest 63% of customer service AI deployments reach positive ROI in year one, with a median payback period of around four months, but only when the initial use case is tightly scoped and the baseline metrics are documented before go-live.

The Hidden Work Vendors Underplay

The licence fee is rarely the biggest cost in an AI virtual agent deployment. For most mid-market organisations, the work that determines whether the project succeeds happens before a single customer interaction is handled, and it is consistently underestimated in vendor proposals.

The costs that do not appear in the demo:

  • Data and knowledge clean-up. AI virtual agents retrieve answers from knowledge bases, CRM records, and policy documentation. If that information is inconsistent, outdated, or fragmented across systems, the virtual agent will produce unreliable outputs. Cleaning and structuring this data is often the longest task in any deployment.
  • Integration work. Connecting the virtual agent to live systems (order management, billing, authentication, CRM) takes development time that is frequently scoped too tightly in early proposals.
  • Governance and policy tuning. Regulated industries, and any organisation handling vulnerable customers, need documented escalation paths, reviewed response policies, and ongoing monitoring. This is not a one-time setup task.
  • Exception handling design. The interactions the virtual agent cannot handle need to transfer cleanly to human agents, with context intact. Poorly designed handovers are one of the most common causes of customer dissatisfaction in AI deployments.
  • Ongoing calibration. Intent models drift, customer language evolves, and new query types emerge. Virtual agents require regular review and prompt tuning to maintain containment rates after launch.

"Teams often undercount invisible costs, which can materially increase total cost-to-serve, especially in regulated workflows." — Teneo.ai, AI vs Live Agent Cost Analysis 2025

Human-in-the-loop design is not a sign that the AI has failed. It is a deliberate and necessary part of any responsible deployment. For complaints, vulnerable customers, and complex service exceptions, the question is not whether a human should be involved but how quickly and cleanly the handover happens.

A Phased Rollout Model for UK Mid-Market Teams

A phased approach is not a compromise on ambition. It is the most reliable route to sustainable scale, and it is what separates deployments that build internal momentum from those that quietly get deprioritised after the first review cycle.

  1. Phase 1: Prove value on one journey. Select a single, high-volume service journey that meets the use-case criteria above. Document the baseline: total monthly contact volume for that journey, average handle time, current cost-to-serve, and escalation rate. Deploy the AI virtual agent against that journey only. Set a 90-day window to demonstrate containment above 70%, measurable cost-to-serve reduction, and stable customer satisfaction scores. Do not expand scope until these are confirmed.
  2. Phase 2: Extend into adjacent intents. Once Phase 1 metrics are proven and the operational model is stable, identify the next two or three journeys that share similar intent characteristics. These are typically in the same service domain (for example, if Phase 1 covered order status, Phase 2 might extend to returns initiation and delivery exception handling). Reuse the knowledge base and integration work from Phase 1 where possible. Set the same baseline and containment targets before signing off on expansion.
  3. Phase 3: Connect to broader orchestration. At this stage, the AI virtual agent moves from handling isolated journeys to participating in multi-step workflows connected across CRM, billing, field service, or logistics systems. This is where agentic AI capabilities become relevant: the virtual agent is not just answering questions but initiating actions, updating records, and coordinating with other systems. Industry data suggests UK contact centres could automate 60 to 80% of common enquiries by 2029, but the organisations reaching that level are those that built the operational and data foundations in earlier phases.

The critical discipline at each phase: resist the pressure to expand before the current phase is stable. The most common failure mode is not starting too small. It is expanding too quickly before containment, knowledge quality, and integration reliability have been confirmed.

What Operations Leaders Should Ask Before Signing Anything

Vendor proposals for AI virtual agents are almost always optimistic on timeline, conservative on integration effort, and vague on how success will be measured after go-live. Before committing budget, operations leaders should be able to answer the following questions clearly, and so should their prospective vendor.

On business case and sequencing:

  • Which specific journey are we targeting first, and what is the current baseline for volume, handle time, and cost-to-serve?
  • What containment rate is the vendor committing to, and under what conditions?
  • What is the escalation rate threshold at which we would pause or rework the deployment?

On integration and data readiness:

  • Which systems does the virtual agent need to connect to in order to resolve this journey end-to-end, and what is the integration timeline?
  • What is the current state of the knowledge base for this journey, and who owns the clean-up work?
  • How will context be passed to human agents when escalation occurs?

On governance and risk:

  • How does the deployment handle vulnerable customers or regulated interactions?
  • Who is responsible for ongoing calibration, intent model review, and policy updates after launch?
  • What does the monitoring and reporting model look like in the first 90 days?

The right question to anchor the entire evaluation is not "how much can we automate?" It is: where can we reduce cost-to-serve without damaging the customer experience, and how will we know within 90 days whether it is working?

Any vendor who cannot answer that question with specifics is not ready to deploy in a mid-market contact centre environment.

Start Narrow, Scale With Evidence

AI virtual agents will reshape how UK contact centres operate. The economic case is not in doubt: Gartner projects a reduction of $80 billion in global agent labour costs by 2026, and the organisations already seeing those returns are not the ones that launched the biggest programmes. They are the ones that started with the clearest use case, measured the right things, and used early proof of value to justify the next phase of investment.

For UK mid-market operations leaders, the competitive risk is not moving too slowly on AI. It is moving too broadly too soon, absorbing the cost and complexity of a large-scale deployment before the operational foundations are in place to make it work.

Start with one journey. Measure cost-to-serve before and after. Prove containment. Then scale.

If you are building the business case for AI virtual agents and want to work through use-case selection, baseline metrics, and rollout sequencing with an independent perspective, book a discovery workshop with the Fortay Connect team.

FAQs

What is the best first use case for AI virtual agents? The best first use case is a repetitive, high-volume journey with clear intent, low ambiguity, and limited regulatory risk. Order status, appointment changes, billing queries, and password resets usually fit best because they offer quick proof of value and measurable containment.

Why do so many AI virtual agent projects fail? They usually fail because the scope is too broad at the start. Teams often try to modernise the whole contact centre at once, which hides the real work in knowledge quality, integrations, escalation design, and governance.

How should operations leaders measure AI virtual agent success? Cost-to-serve and labour compression are stronger measures than headline automation rate. They show whether AI is reducing the full effort required to resolve a customer need, including escalations, rework, and human clean-up.

What should a phased AI virtual agent rollout look like? Start with one journey, prove containment and cost-to-serve reduction, then expand into adjacent intents only when the first phase is stable. After that, connect the virtual agent to broader workflows across CRM, billing, or order management.

How do you choose the right first AI virtual agent journey? Choose a journey with predictable intent, high volume, accessible data, and low emotional or regulatory sensitivity. The right first use case is the one where automation removes a meaningful operational step, not just a few customer contacts.