Optimise Your Voice Agent – Boost Performance and Improve AI Telephony
A Voice Agent that is just as good — or just as poor — after three months as it was on day one is a bad Voice Agent. The best deployments improve continuously — because a systematic optimisation process has been established that translates data into insights and insights into better configurations. This article shows you the optimisation methodology that successful SMEs apply.
The Fundamental Problem: Optimisation Without Method
Many companies "optimise" reactively: a customer complains → they try to fix the specific problem → they hope a similar complaint won't arise again. This is not an optimisation process — it is firefighting.
Systematic optimisation works differently: you continuously collect data, identify patterns, formulate hypotheses, test changes and measure the impact. This cycle — known as PDCA (Plan-Do-Check-Act) or the Lean Improvement Cycle — is the backbone of professional Voice Agent optimisation.
Companies that operate a systematic optimisation process improve their CSAT scores by an average of 28 points over 12 months. Companies without a structured process: 6 points.
Phase 1: Collect Data – What You Really Need to Measure
The Data Foundation
Optimisation without data is guesswork. The most important data sources for Voice Agent optimisation:
Conversation transcripts: The richest source. Every conversation is a data point about what customers want, how they phrase it, where misunderstandings arise and where the agent falls short.
Intent confidence scores: How confident was the agent in its intent classification? Low confidence scores signal training gaps.
Conversation drop-off points: Where does the conversation break off or get escalated? These points are optimisation priorities.
Post-call CSAT: Customer satisfaction immediately after the conversation. Best collected automatically via SMS or a short prompt at the end of the call.
CRM outcome data: Was the business objective achieved? (Appointment booked, problem solved, conversion achieved?)
Ensuring Data Quality
Poor data leads to wrong conclusions. Check:
- Are all conversation transcripts captured completely?
- Is CSAT data collected for all use cases?
- Are there data gaps (certain times of day, call types that are not captured)?
- Are integrations correct and complete?
Phase 2: Analyse – Recognising Patterns
The Five Analytical Lenses
Lens 1 – Volume analysis: Which use cases and intent categories have the highest volume? That is where optimisation is most worthwhile.
Lens 2 – Failure mode analysis: What are the most common error types? Ranked by frequency and impact.
Lens 3 – Funnel analysis: What does the conversation flow look like? At which point do customers leave the process prematurely?
Lens 4 – Segment analysis: Are there customer segments that systematically produce worse results? (Older customers, certain regions, certain call times?)
Lens 5 – Competitive benchmark: How do your metrics compare to industry benchmarks?
The Root Cause Analysis Method
When you have identified an anomaly (e.g. high escalation rate for a particular intent), proceed systematically:
- Collect data: All conversations with this intent in the last 30 days
- Identify patterns: What do the conversations have in common? Particular phrases? Particular conversation phases?
- Formulate hypothesis: "The high escalation rate arises because the agent does not process X correctly"
- Test hypothesis: A/B test with adjusted configuration
Phase 3: Formulate and Prioritise Hypotheses
The Hypothesis Framework
Good optimisation hypotheses follow this pattern:
"We believe that [change X] will improve [metric Y] by [expected effect Z], because [reasoning/evidence]."
Example: "We believe that adding a proactive appointment alternative ('If Tuesday doesn't work, we also have availability on Thursday') will improve the appointment conversion rate by 15–20%, because our data shows that 23% of customers hang up after the first appointment offer without asking for an alternative."
This structure forces evidence-based thinking and makes measuring results clear.
Prioritisation Matrix
Not all hypotheses are equally valuable. Evaluate on two axes:
Impact: How large is the expected effect? (Low/Medium/High) Effort: How complex is the implementation? (Low/Medium/High)
Prioritise: High impact + Low effort = Quick wins (address immediately). High impact + High effort = Strategic projects (plan and budget). Low impact = low priority.
Phase 4: A/B Testing – Validating Changes
Why A/B Testing and Not Just Making the Change?
The natural reaction: "If we know what the problem is, we just change it." The risk: a change that helps in one situation can harm another. Without A/B testing, you don't know whether an improvement is due to your change or to other factors.
A/B Testing for Voice Agents: Practical Implementation
Variant A: Existing configuration (control group) Variant B: New configuration (test group)
Traffic split: 50/50 for equal sample sizes, or 80/20 if you want to minimise the risk of a deterioration.
Metrics for the test:
- Primary metric: What you want to improve (e.g. Conversion Rate)
- Guardrail metrics: What must not get worse (e.g. CSAT, Escalation Rate)
Duration: At least 7 days, ideally 14 days, to account for day-of-week effects. Minimum sample size: 100 conversations per variant for statistically robust results.
Interpreting Results
- Variant B is statistically significantly better (p < 0.05) on primary metric with no guardrail violation → Roll out
- No statistically significant difference → Hypothesis rejected, formulate new hypothesis
- Variant B improves primary metric but worsens a guardrail → Revision needed
Phase 5: Prompt Optimisation in Detail
The Most Common Prompt Problems
Problem 1: Instructions too vague Symptom: Agent behaves inconsistently, sometimes correctly, sometimes incorrectly. Solution: More explicit if-then rules, concrete examples in the prompt.
Problem 2: Too many instructions at once Symptom: Agent "forgets" instructions or prioritises incorrectly. Solution: Modularise prompts, set priorities explicitly.
Problem 3: Missing fallback definitions Symptom: Agent gives useless or incorrect answers for unknown requests. Solution: Explicit out-of-scope handling instructions with a clear escalation path.
Problem 4: Static persona without emotional adaptability Symptom: Agent sounds robotic in emotional conversations. Solution: Sentiment detection hooks and different conversation modes (neutral/empathetic/solution-oriented).
Prompt Optimisation Workflow
- Identify conversations with low scores
- Locate the section of the conversation causing the problem
- Analyse the current prompt for that section
- Formulate an optimised prompt
- Set up an A/B test
- Evaluate after 14 days
- Roll out or iterate
Phase 6: Conversion Funnel Optimisation
The Voice Agent Conversion Funnel
For outbound Voice Agents, conversion optimisation is particularly relevant:
Step 1 – Connection Rate: How many of the dialled numbers are actually answered? Optimisation: Optimise call timing (day of week, time of day, season).
Step 2 – Engagement Rate: How many of those who connect listen for longer than 20 seconds? Optimisation: The conversation opener — the first 10 seconds are decisive.
Step 3 – Intent Qualification: How many of those engaged have a relevant concern/interest? Optimisation: Improve targeting, work with more qualified lists.
Step 4 – Desired Action Rate: How many book an appointment, give consent, place an order, etc.? Optimisation: Value proposition, offer framing, urgency elements.
Step 5 – Post-Call Follow-Through: Do those who booked actually show up for the appointment? Optimisation: Reminder sequences, confirmation SMS, re-communicating the value of the appointment.
Latency Optimisation: Speed as a Quality Feature
According to research, users perceive response latencies of more than 1.5 seconds as a clearly noticeable delay. Above 2.5 seconds, CSAT drops significantly.
Latency sources and optimisation measures:
ASR latency (speech recognition): Depends on the provider. Typically 300–800 ms. Barely optimisable by the user.
NLU processing time: 100–300 ms. Optimisable through more efficient prompts and intent definitions.
API call latency: Variable (50 ms to 2 seconds). Optimisable through caching of frequent requests, regional server locations.
TTS latency (speech output): 200–600 ms. Optimisable through TTS streaming rather than full pre-generation.
Measuring Optimisation ROI
Every optimisation measure should demonstrate its ROI. Calculation:
ROI of an optimisation measure:
- Effort: Internal hours × hourly rate + provider costs for adjustment
- Benefit: Delta in primary metric × monetisation factor
Example: Optimising the conversation opener increases the conversion rate from 18% to 23% with 500 outbound calls/month. Average order value: €400.
- Additional conversions: 500 × 5% = 25/month = 300/year
- Additional revenue: 300 × €400 = €120,000/year
- Optimisation effort: 8 hours × €80 = €640
- ROI: 18,650%
This example is extreme — but it illustrates why conversion optimisation for Voice Agents is among the most profitable measures available.
Conclusion: Optimisation as a Core Discipline
A Voice Agent is not a static product — it is a learning system that benefits from systematic human guidance. Companies that establish optimisation as a permanent practice build an insurmountable lead over time: their agent keeps improving while competitors stagnate with their initial configuration.
Discover how anicall.io accompanies you through the entire optimisation process.