Voice AI reduced our call costs by 58%: a case study
Voice AI reduced our call costs by 58%: a case study
A mid-size financial services client handling 22,000 inbound calls per month replaced their first-line human agent tier with a Voice AI system. Twelve weeks after go-live: call handling cost dropped by 58%, average handle time fell from 4 minutes 20 seconds to 2 minutes 48 seconds, and CSAT scores improved by 11 points. This post is the complete account of how we did it - the architecture, the numbers, the problems we hit, and the lessons that apply to any enterprise deploying Voice AI at scale.
The 58% figure gets attention. It is the number every CFO wants to hear and every vendor promises in a slide deck. What the slide decks do not show is the six weeks of SIP configuration, the VAD tuning sessions, the two UAT failures, the escalation logic that had to be rebuilt three times, and the moment at week eight when the client almost pulled the plug because CSAT scores dipped before they recovered.
This post is the honest version. The full account of a Voice AI deployment from business case to production results - with the real numbers, the real problems, and the real decisions that made the difference between a failed pilot and a system that now handles over 18,000 calls per month without a human agent on the first line.
The client and the problem
The client is a mid-size financial services company - I am keeping the name confidential but the details are real and shared with permission. They handle customer calls across three categories: account balance and transaction enquiries (41% of call volume), payment arrangement requests (33%), and general account support (26%). Total inbound volume at the time of engagement was 22,400 calls per month.
Their contact centre ran on a legacy PBX system with a traditional IVR front-end - press 1 for balance, press 2 for payments - routing calls to a team of 14 full-time agents. Average handle time was 4 minutes 20 seconds. Average cost per call, including agent salary, telephony infrastructure, and overhead, was £4.80. Monthly call centre operating cost: approximately £107,500.
The business case they brought to us was straightforward: could a Voice AI system handle the 41% of calls that were purely informational - balance and transaction enquiries - without a human agent, and could it do so without degrading the customer experience for a base of older, less tech-comfortable callers who had been using the same phone number for fifteen years?
The architecture we chose and why
The client's existing PBX used SIP trunking through a legacy carrier. Our first decision was whether to replace or integrate. Replacing the PBX would have added three months and significant capital expenditure to the project. Integrating the Voice AI platform alongside the existing PBX - routing AI-eligible calls to the new system while keeping the PBX in place for human agent handling - was the right call for a first deployment of this scale.
We used Vapi as the Voice AI orchestration layer, connected via a new Twilio SIP trunk sitting parallel to the existing carrier trunk. Inbound calls hit the PBX first. The IVR - which we kept in place - now had a fourth option added: callers pressing 1 for balance enquiries were routed directly to the Vapi-powered AI agent rather than the human queue. This meant the switchover was invisible to callers who chose options 2, 3, or 0 for operator.
| Component | Choice | Reason |
|---|---|---|
| Voice AI platform | Vapi | SIP integration, per-turn latency visibility, bring-your-own STT |
| SIP trunk | Twilio | Fastest integration, clean PBX handoff, existing team familiarity |
| STT engine | Deepgram Nova-2 | Lowest latency, financial services domain accuracy |
| LLM | GPT-4o | Accuracy on structured financial queries, function call reliability |
| TTS engine | ElevenLabs | Most natural voice for older demographic, low first-chunk latency |
| Backend API | Client's existing REST API | Balance and transaction data, account authentication |
The three problems we did not see coming
Every Voice AI deployment has surprises. This one had three that were significant enough to delay go-live by two weeks and required fundamental redesigns of parts of the system we thought were finished.
Before the AI could retrieve a caller's balance, it needed to authenticate them - verifying their account number and date of birth against the client's CRM. During testing, this API call averaged 340ms. In production, under load, it spiked to over 900ms on roughly 8% of calls. That 900ms of silence - while the AI waited for the authentication response - was causing callers to think the call had dropped and hang up. We built a bridging phrase system into the system prompt ("Let me just verify your details - one moment") and added a 5-second timeout with a graceful fallback to the human queue. Abandonment on those calls dropped from 12% to 2.4%.
Our UAT testing used colleagues in their 30s and 40s speaking in clear, complete sentences. The actual caller base - averaging 54 years old - spoke differently. They paused mid-sentence. They said "er" and "um" frequently. They read account numbers digit by digit with long gaps between each number. Our VAD threshold, tuned for clean speech, was cutting them off mid-utterance and triggering the AI response before they had finished speaking. We spent four days recalibrating the VAD threshold specifically for this demographic. The fix also required rewriting the system prompt to handle partial utterances more gracefully and to confirm account numbers read back to the caller before proceeding.
Our original escalation logic transferred to a human agent if the caller said "agent", "human", or "speak to someone". In production we discovered callers expressed this in dozens of ways we had not anticipated: "I don't want to talk to a computer", "Can I just speak to a real person please", "This isn't working", "I need help", and in one memorable case, just silence for eight consecutive seconds. We rebuilt the escalation logic around intent detection rather than keyword matching, adding a frustration signal based on turn count, repeated rephrasing of the same question, and rising sentence pitch patterns detected in the STT confidence scores. Unintended escalations dropped by 34%. Requested escalations were now being caught reliably.
The week eight crisis - and what saved the project
At week eight, the client's operations director called me directly. CSAT scores for AI-handled calls were sitting at 61 - fourteen points below the human agent baseline of 75. The board had seen the numbers and were asking whether to revert. I had 72 hours to show improvement or the project would be rolled back.
I pulled every call recording from the previous week and listened to 40 of the lowest-rated calls personally. The pattern was clear within the first ten calls. The AI was giving technically correct answers in a tone that felt transactional and cold. When a caller said "I'm worried about this transaction I don't recognise", the AI was reading back the transaction details accurately but without any acknowledgement of the caller's concern before doing so. It sounded like a database query, not a conversation.
We rewrote the system prompt over 36 hours. We added explicit empathy instructions for financial concern scenarios. We added a pause before reading sensitive information. We changed the opening acknowledgement from "I can help you with that" to responses that mirrored the caller's emotional register - "I understand, let me check that for you right away." We also shortened the AI's responses by 30% - callers in a financial anxiety state do not want paragraphs, they want their answer.
The result: CSAT scores moved from 61 to 74 in nine days - one point below the human baseline. By week twelve they reached 86, exceeding the human agent score for the first time. The single biggest lever was not the technology. It was the system prompt. A well-engineered Voice AI system with the wrong conversational design will always underperform a human agent. A well-engineered system with a thoughtfully written prompt can exceed one.
The numbers at week twelve
Here is the full before-and-after breakdown at the twelve-week mark, measured against the same call category (balance and transaction enquiries) handled by human agents in the prior quarter:
| Metric | Before (Human) | After (Voice AI) | Change |
|---|---|---|---|
| Cost per call | £4.80 | £2.02 | -58% |
| Average handle time | 4m 20s | 2m 48s | -35% |
| CSAT score | 75 | 86 | +11 pts |
| First-call resolution | 81% | 88% | +7 pts |
| Call abandonment rate | 6.2% | 3.8% | -2.4 pts |
| Monthly operating cost | £107,500 | £68,200 | -£39,300/mo |
The monthly saving of £39,300 represents a full return on the implementation investment in under four months. The 14 agents previously handling the full call volume were not made redundant - they were redeployed to handle the payment arrangement and complex account support calls, which require genuine human judgement and where the client was most uncomfortable with AI handling. Average agent handle time on those remaining calls actually improved because agents were no longer fatigued from handling repetitive balance enquiries all day.
"The CSAT improvement was the result that mattered most to the client. The cost saving was expected. The quality improvement was not. That is now the first thing I show in every new business case - not the cost slide, the CSAT slide."
- What I say when clients ask whether Voice AI will hurt their customer experienceFive lessons that apply to every Voice AI deployment
Start with high-volume, low-complexity call types - balance enquiries, appointment confirmations, status checks. These calls are highly structured, the caller intent is narrow, and the success criteria are easy to define. Getting these right builds internal confidence and gives you operational data before you tackle the harder problems.
Our VAD and escalation problems would have been caught in UAT if we had tested with a representative sample of actual callers rather than project team members. For B2C deployments, the demographic of the real caller base is almost always different from the team doing the testing. Build a beta group of 20–30 real callers and run them through the system two weeks before go-live.
More engineering time was spent on system prompt iteration than on any other component of this deployment. The prompt determines tone, escalation behaviour, how the AI handles uncertainty, and how it responds to emotional callers. Treat it with the same rigour as production code - version control it, test changes systematically, and never push a prompt update to production without an A/B test against a controlled call sample.
We started collecting post-call CSAT surveys on day one of go-live. This gave us eight weeks of data to diagnose the quality problem before the client escalation. If you wait until the client asks for a review to start measuring quality, you will always be behind the problem. CSAT measurement infrastructure should be part of the go-live checklist, not an afterthought.
The internal rollout of this project was significantly smoother because we framed it to the client's operations team as redeployment rather than headcount reduction. The 14 agents moved to higher-value work. Agent satisfaction scores actually improved - people prefer handling complex problems over answering the same balance enquiry 80 times a day. If your deployment plan includes involuntary redundancies, expect internal resistance that will slow every phase of the project.
What 58% actually takes
The 58% cost reduction is real. It took twelve weeks, three significant problem-solving sessions, one near-cancellation, and more system prompt iterations than I can count. It was not delivered by the technology alone - it was delivered by a team that understood both the technology and the callers it was serving well enough to close the gap between them.
If you are evaluating Voice AI for a contact centre deployment, the numbers in this case study are achievable. But they require the same rigour you would apply to any significant technology change: real-caller testing before go-live, CSAT measurement from day one, a system prompt treated as a first-class engineering artefact, and escalation logic built around intent rather than keywords.
The technology will not save a poorly designed deployment. A well-designed deployment will produce results that surprise even the most sceptical CFO.
Planning a Voice AI deployment?
I write every week about what it actually looks like to ship Voice AI in production - the architecture decisions, the things that go wrong, and the results when it works. All free, no paywall.
Follow with your Google account and get new posts in your Blogger reading list automatically.
Comments
Post a Comment