Two ways to buy
- Per bench. One use case, one bench: the twins it needs, the tasks, the simulated users, the runs and the reports, for a term.
- Annual platform licence. Several benches under one agreement, at a lower price per bench.
A deployment inside your own VPC is priced separately from SaaS, because it carries maintenance on both sides.
What sets the price
Scoping starts from a short sheet, and three things on it move the number.
- Twins. Which systems the agent's tools call, how they are reached, by API, by database or over MCP, and which are library twins as opposed to custom builds. The twins are most of the price.
- Tests. Where the tasks come from, in any mix of production logs, regulations and policies, hand-written scripts, and the agent's own prompts and tools; how many; and in which modalities. Voice calls carry real per-call costs.
- Where it runs. SaaS, or fully on-prem inside your VPC.
With the sheet filled in, we return a firm price, per bench or under an annual licence, with a twin-by-twin build plan and dates.
How soon
In SaaS, a first bench is ready in about two weeks from signing. Inside your VPC, the platform is deployed in your project first, usually a couple of days, and the two weeks start once it is running.
The commitment on your side in those two weeks: access to the agent in a dev or test environment, the tools and tables it reads and writes, seed data or a small masked sample, the test inputs, and one subject-matter expert to review the exam before the first run.
What a bench includes
- Twins of the systems the agent calls, from the library or built for you.
- Simulated users with the personas, and for voice the accents and background noise, the task calls for.
- Tasks with pass criteria your team signs before the first run.
- Runs with repeats, and a pass rate with a confidence interval for every version.
- The evidence behind every attempt: transcript, trace, the ledger of service requests, and each criterion's verdict with its reasoning.
- A report with a root cause for every failure, bucketed as prompt, tools, runtime or model, and a suggested fix.
- Console access for your team, and the CI hook for nightly or per-pull-request runs.