You never write a prompt
Your team explains the work and reviews what the agent says. The prompt itself stays ours to write and maintain.
Quality
Tested before it takes a call
You never write a prompt. We gather the material, build the tests from it, and release only what passes.
DRING Test Suite
The tests come from your own material
Prompt writing is not your job. Neither is imagining the awkward calls your agent will get.
Tests built from what we gather
Documents, recordings, policies and the comments you leave during review become the test set for that agent.
DRING approved, then live
An agent goes to production only after it clears that set. Until then it stays on a private line.
1,000 to 10,000 simulated conversations, built for your company, before the first real call. The number of simulated conversations depends on the agent, and every one of them is generated from your own workflow, documents and call recordings, not a generic benchmark set.
The same discipline applies to a change. Every rule you give us joins that agent's single test set, and a new rule ships only when every existing rule still passes beside it. See how the agent is built
Before launch
What happens before an agent launches
The same pipeline runs for every agent, however simple the use case looks from the outside.
Personas from your reality
Angry callers, wrong orders, policy traps, silence and interruptions, drawn from your own workflow.
Simulated conversations
Each persona runs through a full conversation with the agent, not a single scripted question.
Independent evaluation
Automated evaluators score each conversation. Where they disagree, a person audits the call.
Regression check
Results are compared against the previous release before anything ships, so a fix cannot quietly break something else.
Handover paths checked
Escalation to a person is verified explicitly rather than assumed to work.
Staged rollout
Live traffic increases in controlled stages, with scores watched at each step and the previous release ready.
Automated evaluation can drift or miss a nuance. Independent checks, plus human review of every disagreement, make the signal more useful than a single score would be.
After launch
Quality control keeps running
Testing does not stop at launch. Live agents go through regular quality control for as long as they run.
Every call scored
Resolution, policy adherence and response time are scored, and sentiment is labeled, the moment a call ends.
Regular sweeps
Every live agent is rerun against the previous release on a schedule, catching a quiet regression before a customer does.
Drift alerts
A score moving outside its expected range raises an alert on its own, without waiting for the next sweep.
Speech error tracking
Speech-to-text mistakes are tracked by name, place and product term, then corrected in the agent's dictionary.
| Measure | Definition | Why it matters |
|---|---|---|
| Resolution | Whether the customer's reason for calling was actually addressed. | The core measure of whether the agent did its job. |
| Sentiment | How the customer's tone changed from the start of the call to the end. | Catches calls that resolved on paper but frustrated the customer. |
| Policy adherence | Whether the agent's actions stayed within the rules configured for it. | Keeps the agent inside the boundaries the business set. |
| Response time | The gap between the customer finishing speaking and the agent replying. | A slow reply feels like a bad connection, even when the answer is right. |
| Handover accuracy | Whether the agent escalated to a person when it should have. | A missed escalation is worse than an unnecessary one. |
| Regression delta | The score change against the previous release on the same test set. | Catches a release that quietly makes the agent worse. |
Outbound A/B
The better opener wins on outcomes, not opinions
Which opening line, tone or voice works is settled on live outbound calls, not in a meeting.
Illustrative test · opener variantsPhase 1 · equal traffic
0 calls
0 calls
0 calls
Minimum sample not reached. Keep splitting traffic.
What gets tested
The opening line, the tone, the voice, the pacing, the order of questions, the way an objection is met. One variable per variant, so the result is attributable.
How the winner is chosen
On the campaign's own outcome, with an agreed minimum sample per variant. Sentiment and handover rate act as guardrails: a variant that wins on outcomes but loses on reception does not ship.
How it goes live
The winner is promoted in controlled stages, the other variants retire, and the change is recorded so the next comparison starts from a known baseline.
A promoted variant still passes the regression set before it reaches the whole campaign.
Guardrails
What an agent will never do
- It refuses what is out of scopeA request outside what the agent was built to handle is declined, not improvised around.
- It never invents policyIf it was not given a rule, it does not guess one and it does not promise on your behalf.
- It escalates sensitive topicsSensitive requests go to a person rather than being handled on the agent's own judgment.
- It always offers a personA human route stays available to the caller, whatever the agent is able to do itself.
FAQ
Testing questions
Do we have to write or maintain the prompt?+
No. You explain the work, share the documents and recordings, and review what the agent says. Writing and maintaining the prompt is our job, and the tests are built from the material you gave us rather than from a generic template.
Does testing happen before every release, not just launch?+
Yes. Regression sweeps run against the previous release on a regular schedule, and every change goes through the same staged rollout as the first launch did.
What happens when two evaluations disagree?+
The conversation is routed to a human reviewer and audited before the result is finalized. Repeated disagreements are a signal to refine the evaluation set itself.
Can we see the test scenarios for our agent?+
Yes. The personas and scenarios are built around your own policies and edge cases, and they are visible to you during onboarding.
What happens when a live agent's scores drift?+
A drift alert fires automatically, the calls behind it are reviewed, and the fix goes back through the test set before that call type returns to the agent.
See the test suite on your own use case
Walk through what a measured launch would look like, from the first workflow brief to the review loop after go-live.