What we deploy before we sell it.
Teknovat runs its own agents in production — on Mawered, our multilingual HR/payroll SaaS, and on our own internal processes. What we propose to you, we measure on ourselves first.
An agent is not deployed before it has been measured.
Every agent goes through three stages before reaching a client: a benchmark on real data, a precision evaluation on edge cases, and a week of internal production with full observability. If the error rate exceeds our threshold, the agent does not ship.
- 01 Benchmark
Real data, edge cases identified manually, baseline precision measured.
- 02 Evaluation
Production load simulation — speed, consistency, anomaly handling.
- 03 Internal production
Deployed on our own systems for one week. Full observability, alert thresholds active.
What we measure per request.
- Latency
End-to-end response time, p50 and p95.
- Cost per request
Tokens consumed × model price. Calculated on every call.
- Error rate
Incorrect outputs caught by consistency checks.
- Coverage
Share of cases handled without human intervention.
Why cost per request matters.
A payroll agent that costs €0.04 per payslip produced is profitable from 50 employees. An agent whose cost exceeds €0.50 per request does not scale. Teknovat optimises cost at design time — not by cutting corners on precision.
Exact figures vary by model, volume and complexity. We document the observed cost for every client deployment.
You buy the production-hardened version, not the prototype.
The agents we deploy with clients have passed our own quality controls. The methodology, precision thresholds and observability measures are the same as on our internal systems. This is not rhetoric — it is what we commit to.
See the payroll agent in 20 minutes