Case study
Cutting an AI voice bill by 90 to 95%
Our own project
Commercial AI voice services charge $20 to $200 a month for a mid-volume workload. We built a self-hosted alternative on open-source models and ordinary graphics hardware that produces the same workload for about $2 to $3 a month, a 90 to 95% cut.
The problem
- Established AI voice vendors price for enterprise volume. For a workload of about 150,000 characters a month, the commercial plans ran $20 to $200 a month, and the price climbs steeply with use.
- The quality was good. The price was the problem.
What was done
- Measured the real workload: how many characters, how often, and what quality bar it had to meet.
- Tested open-source voice models against that bar, side by side with the paid service.
- Built the pipeline on commodity GPU hardware and put it into daily production.
The result
- Monthly cost: $20 to $200 down to about $2 to $3 for the same 150,000-character workload.
- A 90 to 95% reduction.
- No per-character pricing, so the cost no longer climbs with use.
What it shows
Vendors price for their biggest customers. For a mid-sized workload, a purpose-built alternative can match the quality at a fraction of the cost. That is the build-versus-buy call we make on every engagement, and sometimes the answer is to buy.
Keep reading
Related case: 70% AI adoption, starting from zero
Guide: what inefficient processes cost, with the math