Vendor websites quote 90% cost reductions. Procurement teams quote 'show me the math'. This post is the math, from three production deployments we've shipped in the last 18 months.
The numbers
Average loaded cost per call (UK call center, 2026):
- Human agent: £3.20 per call (4-minute average, salary + benefits + supervision + facilities)
- AI voice agent: £0.18 per call (LLM tokens + STT + TTS + telephony)
- Per-ticket savings at full deflection: £3.02
- Realistic deflection rate: 40–60% of inbound
Where the AI cost actually goes
- LLM tokens: ~£0.04 per call (Claude / GPT-4o, 4-min average)
- Speech-to-text: ~£0.05 per call (Deepgram streaming)
- Text-to-speech: ~£0.06 per call (ElevenLabs natural voice)
- Telephony: ~£0.03 per call (Twilio inbound + carrier fees)
Where the human cost actually goes
If you only count salary, you're missing 60% of the real cost. Loaded cost includes:
- Salary + employer NI + pension (~£28/hour fully loaded for a UK tier-1 agent)
- Supervisor and QA cost amortised across the team
- Recruitment, onboarding, and turnover (12–18 month average tenure)
- Facilities, IT, training, holiday cover
- Out-of-hours premiums and weekend coverage
The break-even isn't about cost — it's about volume
A typical AI voice agent project costs £24k–60k to ship. At £3.02 per-call savings, that's 8,000–20,000 calls to break even. For a 100k-call-per-month operation, that's 1–2 months. For a 10k-call-per-month operation, it's 8–20 months. Below 5k/month, the maths usually doesn't work.
What this doesn't include
There's no number above for: customer satisfaction lift, NPS impact, agent burnout, employer brand, or the option value of being able to launch new services without hiring a team. These are real and they often dominate the call-cost arithmetic — they're just harder to put in a spreadsheet.
Written by
RMC Engineering