How to Pilot an AI Voice Agent: A Practical 6-Week Playbook for UK Contact Centres

Arkadas Kilic 8 min read · 2026-08-27

Most AI voice agent pilots fail before they start. Not because the technology does not work, but because the pilot is designed as a science experiment rather than a production deployment. Teams spend weeks on demos, skip compliance review, and never define what "good" looks like. Six months later, the initiative is shelved.

This playbook is different. It is the exact approach we use to take UK contact centres from kickoff to a live, enterprise-grade AI voice agent in 4 to 6 weeks. Every milestone is actionable. Every number is real.

Before You Start: Define the Problem, Not the Technology

The single biggest mistake contact centre leaders make is starting with the technology. Before you write a single line of configuration, answer these four questions:

1. Which call type are you automating? Pick one. Ideal candidates handle 500 or more calls per month, follow a predictable script, and have a clear resolution outcome (e.g. balance enquiry, appointment booking, payment arrangement, meter reading).

2. What does success look like at week 6? Set a number. A realistic target for a first pilot is 60 to 70% containment on the chosen call type, with a CSAT score within 5 points of your human agent baseline.

3. Who owns compliance sign-off? In UK regulated industries (financial services, utilities, healthcare), FCA, Ofgem, or CQC obligations apply to automated voice interactions. Identify your compliance stakeholder on day one, not week five.

4. What is your telephony stack? If you are already on Amazon Connect, your path to production is significantly shorter. If you are on a legacy PBX, budget an extra week for SIP trunk configuration.

Week 1: Discovery and Data Baseline

Pull 90 days of call recordings for your chosen call type. You need a minimum of 200 calls. Listen to 50 of them. This is not optional.

From this analysis, document:

• The 10 most common caller intents within this call type. Most teams discover their "simple" call type has 8 to 12 distinct sub-intents.
• The 5 most common failure points where calls escalate to a human agent. These become your containment risk register.
• Average handle time (AHT) for the call type. This is your cost baseline. A typical payment arrangement call in UK utilities runs 4 to 7 minutes. An AI voice agent should resolve the same interaction in 90 to 150 seconds.
• Peak call volume windows. Your pilot needs to handle real traffic, not synthetic test calls.

Deliver a one-page brief summarising these findings. This document governs every design decision for the next five weeks.

Week 2: Architecture and Compliance Design

For UK contact centres, we build on AWS. Amazon Connect is the telephony layer. Amazon Lex handles natural language understanding. AWS Lambda executes business logic. Amazon DynamoDB or your existing CRM stores session state. All of this runs inside your AWS account, in the eu-west-2 (London) region by default.

• FCA Consumer Duty (if applicable): Automated voice interactions must deliver fair, clear, and not misleading outcomes. Document how the agent handles vulnerable customer indicators and how it routes to a human agent within two interactions if resolution is not achieved.
• GDPR and UK Data Protection Act 2018: Confirm that no call audio or transcript data leaves the UK region. Confirm your data retention policy. Amazon Connect Contact Lens recordings can be configured with automatic deletion at 90 days.
• Ofcom regulations: If your agent makes outbound calls, confirm compliance with the Ofcom persistent misuse rules. Abandoned call rates must stay below 3% over any 24-hour period.
• Call recording disclosure: Your IVR must inform callers that the interaction is automated and may be recorded before any personal data is collected.

Draft your escalation logic at this stage. Every AI voice agent must have a clean handoff to a human agent. Define the exact triggers: caller requests a human, agent confidence score falls below threshold (we use 0.75 as a default), or three consecutive misunderstood utterances.

Week 3: Build Sprint One (Core Flow)

Do not try to build everything. Build the core flow that covers the top three intents identified in week one. In a typical UK utility contact centre, this might be: balance enquiry, payment arrangement, and direct debit update. Together, these three intents usually account for 55 to 65% of total call volume on the chosen call type.

• Amazon Connect contact flow with initial greeting, intent routing, and human escalation path
• Lex bot with three intents, 15 to 20 sample utterances per intent, and slot filling for required data (account number, postcode, payment amount)
• Lambda functions for CRM lookup and write-back. If your CRM is Salesforce or Dynamics 365, plan for 2 to 3 days of API integration work.
• A test harness with 30 scripted test calls covering happy path, edge cases, and escalation triggers

By end of week three, your agent should pass 85% of scripted test cases. If it is below that, do not move to week four. Spend an extra two to three days on utterance tuning.

Week 4: Build Sprint Two (Edge Cases and Integrations)

This is the week most pilots skip, and it is the reason most pilots fail in production.

Work through your containment risk register from week one. For each of the five failure points you identified, build explicit handling:

• Caller cannot verify identity: After two failed verification attempts, route to a human agent with a warm handoff message that includes the caller's phone number and the intent they stated.
• CRM lookup fails: Do not let the agent hang. Return a graceful fallback message and offer a callback or human transfer within five seconds.
• Caller is in financial hardship: In UK financial services and utilities, agents must follow vulnerability protocols. Build a keyword detection layer (using Amazon Comprehend or Connect Contact Lens real-time analytics) that flags distress indicators and triggers immediate human escalation.
• Out-of-hours calls: Define your after-hours behaviour explicitly. Options include voicemail capture, callback scheduling, or a static information message. Do not leave this as a default.
• Repeat callers: If a caller has contacted you three or more times in seven days on the same issue, route to a specialist queue. This is both a CX best practice and a Consumer Duty consideration.

Complete your end-to-end integration testing this week. Run 100 test calls across all intents and edge cases. Target: 90% pass rate before moving to live traffic.

Week 5: Controlled Live Traffic

Route 10 to 15% of your target call type to the AI voice agent. Keep 85 to 90% on your existing human agent queue. This is not a soft launch. It is a controlled production deployment with active monitoring.

• Containment rate: Percentage of calls fully resolved by the agent without human transfer. Target: 55% or above in week five. Below 45% is a red flag requiring immediate investigation.
• Average interaction duration: Should be running 90 to 180 seconds for simple resolution intents.
• Escalation reason codes: Log every escalation by reason. If more than 20% of escalations share the same reason, that is a fixable gap in your agent logic.
• CSAT on AI-handled calls: Send a post-call SMS survey (one question, five-point scale). Target: 3.8 or above out of 5.0 in week five.
• False positive escalations: Cases where the agent escalated unnecessarily. These represent containment you are leaving on the table.

Hold a daily 15-minute standup with your technical lead, operations manager, and compliance stakeholder throughout week five. Issues surface fast on live traffic. You need a decision-making loop that is equally fast.

Week 6: Scale and Handover

If your week five metrics hit target, scale to 100% of the chosen call type in week six. If you are 5 to 10 percentage points below containment target, spend the first three days of week six on utterance and flow tuning before scaling.

• Production runbook: Documents every contact flow, Lambda function, Lex bot version, and integration endpoint. Your operations team must be able to make minor changes without calling your implementation partner.
• Monitoring dashboard: A CloudWatch dashboard with the five metrics above, plus error rate on Lambda functions (target: below 0.5%) and Amazon Connect service health.
• Escalation playbook: A one-page document for your agent supervisors covering how to identify AI-handled calls in your CRM, how to flag a call for model review, and who to contact if the agent behaves unexpectedly.
• Compliance sign-off document: Signed confirmation from your compliance stakeholder that the deployed agent meets FCA, Ofgem, or relevant regulatory requirements. Do not go to full production without this.
• 30-day optimisation plan: A prioritised backlog of the next three intents to automate, based on call volume data from weeks one through six.

What Realistic Results Look Like

Here is what a well-executed 6-week pilot delivers for a UK contact centre:

These numbers are based on deployments we have run in UK financial services and utilities. They are not projections. They are outcomes.

The Three Mistakes That Kill Pilots

Ready to Run Your Pilot?

We build enterprise-grade AI voice agents for UK contact centres. Our team has delivered production deployments in financial services, utilities, and healthcare, all on AWS, all with compliance built in from the first line of configuration.

If you want a structured assessment of your contact centre's readiness and a clear path to production in 4 to 6 weeks, start with our scorecard.