Healthcare Voice AI // For Compliance and Clinical Leaders
Cekura and CogniSwitch:
test how your voice agents behave,
and prove what they decided
CogniSwitch and Cekura are complementary, not competing. Cekura tests how your voice and chat agents behave, before production and on every live call. CogniSwitch proves the content of the decision they conveyed was correct against the record.
No, CogniSwitch is not a Cekura alternative. They operate at different layers of the stack, and they serve different buyers. Cekura tests and monitors how your voice agents behave on a call. CogniSwitch proves the content they conveyed was correct against the record. In a regulated voice deployment you may want both: it's an and, not an or.
Why this matters
Why run Cekura and CogniSwitch together?
Most health systems have voice agents stuck in pilots. A reliable call is not the same as a defensible one. The blocker is rarely whether the agent sounds right. It's that no one can prove, to a regulator or an internal audit, that what the agent told a patient or payer matched the record.
The problem it solves
You can move voice agents out of pilot and into production, because you can now stand behind the content of every decision they convey on a call.
The outcome
When an auditor asks why the agent told a caller what it did, you can pull the exact rule and the record it was checked against, and prove it. The decision is defensible.
The stack
What does a healthcare voice AI stack need to be complete?
A complete regulated AI stack has four layers, and CogniSwitch is the trust layer between the agents and the clinical data.
Evals & Observability
Test the agents before production, score their behavior, monitor live calls
Cekura (voice & chat agent QA), Braintrust, Langfuse
Agents & AI Applications
The voice agents themselves: benefits verification, post-discharge calls, intake
Vapi, Retell, LiveKit, Pipecat, ElevenLabs
Trust Layer
Check the content of each decision against policy at runtime, and give an auditor a reason they can re-derive.
CogniSwitch
Clinical Data
The source of truth the agent has to match
EHR, payer policies, discharge plans, clinical SOPs
Cekura answers
"Does the voice agent behave well on the call?"
CogniSwitch answers
"Can we prove what it decided was allowed?"
CogniSwitch sits above your voice-agent testing and completes the stack.
What does Cekura do on its own?
Cekura is a strong, purpose-built QA and observability platform for voice and chat agents. It simulates real conversations against your agent before production, evaluates the full session, and monitors live calls. The specialization is real, and so is the signal.
Cekura (formerly Vocera, YC F24) bootstraps a test suite from a description of your agent and mines real production conversations for new test cases. Synthetic users with varied accents, background noise, and conversational styles stress the agent against conditions that only show up in live voice traffic, while LLM-based judges and structured evaluators check whether it responded correctly. In production it runs voice-specific signals on every call: gibberish detection, interruption tracking, latency, and sentiment, plus instruction-following and drop-off tracking. For testing whether a voice agent behaves and stays reliable, Cekura does its job well.
What voice-agent testing can't do on its own
Testing and observability confirm how the agent behaves: it heard the caller, followed the script, sounded right, and completed the task. That is a different question from whether the content it conveyed was correct against the source of truth. An agent can pass every behavioral test and still state a benefit that does not match the payer policy. Three things follow from testing behavior rather than verifying the decision.
Behavior is not correctness
A call can be smooth, fluent, and on-script while the information conveyed is wrong against the record. Testing how the agent speaks does not check what it said against the source.
Simulation runs before the call, not on the content
Pre-production simulation stresses the agent against scenarios. It does not verify, on a given live call, that the specific guidance matched the patient's actual plan or record.
Reliability is not provenance
Live monitoring tells you the call completed and the agent stayed reliable. It does not produce a rule-named reason an auditor can re-derive for the one decision under question.
It cannot block a wrong answer on the record
Behavioral signals flag a poor call, not a non-compliant statement. To stop the agent conveying guidance that contradicts the record, you have to check the content against the source.
So with only Cekura you can test and monitor how the agent behaves. What you cannot do is prove that the content of one specific decision was correct and policy-compliant, with a reason you can repeat. That takes a verification step against the record. Better simulation alone does not get you there.
In practice // A health system running voice agents in production
How do Cekura and CogniSwitch work together for voice agents in healthcare?
Consider a health system running two voice agents in production: a benefits-verification agent that calls payers, and a post-discharge agent that calls patients. For each one, Cekura and CogniSwitch do different jobs in the same flow. Cekura tests and monitors how the call goes. CogniSwitch verifies what the agent conveyed against the record. Here is how that plays out.
Benefits-verification voice agent (calls payers)
The scenario
The agent calls a payer to verify coverage and prior-authorization requirements, then relays what is covered back into the workflow.
Cekura tests the behavior
Cekura simulates the call flow before production against varied accents, hold music, and interruptions, and monitors live calls for mishandled turns, mishearing, latency, and drop-off, so the conversation itself holds up.
Answers
"Did the agent handle the payer call well?"
CogniSwitch verifies the decision
CogniSwitch verifies the benefits and prior-auth information the agent states against the actual payer-policy version that applied, produces a verdict that names the rule that fired, and keeps an audit trail of what was conveyed and why.
Answers
"Can we prove the benefits it stated match the policy?"
Together, in one flow
Cekura catches the call-handling failure: the agent mishears a code or stumbles on an interruption. CogniSwitch verifies the content against the record: the coverage the agent relayed actually matches the payer policy, and there is a trail to prove it.
Post-discharge voice agent (medication adherence)
The scenario
The agent makes outbound calls to recently discharged patients to confirm understanding of medications and the care plan.
Cekura tests the behavior
Cekura tests the conversation against edge cases such as accents, interruptions, and mishearing a medication name, and monitors task completion and call quality on live outbound calls.
Answers
"Did the patient call complete reliably?"
CogniSwitch verifies the decision
CogniSwitch verifies the clinical guidance the agent conveyed against the patient's discharge plan, matching every medication and instruction to the record, and flags any discrepancy with the reason it fired.
Answers
"Can we prove the guidance matched the discharge plan?"
Together, in one flow
Cekura catches the call-handling failure: the agent mishears the medication name on a noisy line. CogniSwitch verifies the content against the record: the dose and instructions it conveyed match the discharge plan, and a mismatch is flagged before it stands.
Both layers, together
Cekura catches the call-handling failure. CogniSwitch verifies the content against the record. Together, the voice agent is reliable on the call and defensible to an auditor.
What changes when you add CogniSwitch to Cekura?
What changes when you add CogniSwitch to the voice-agent testing you already run. The rows build from testing the agents to being able to deploy them in a regulated setting with confidence.
Yes = the stack can do this
The first three rows are Cekura doing its job well. The rest is what the verification layer adds.
| What you can do | With only Cekura | With Cekura + CogniSwitch |
|---|---|---|
| Simulate calls and catch behavioral failures before production | Yes | Yes |
| Monitor live calls for latency, interruptions, and drop-off | Yes | Yes |
| Score whether the agent followed instructions on the call | Yes | Yes |
| Verify the content the agent conveyed against the record | No | Yes |
| Name the exact policy rule that drove a decision | No | Yes |
| Reconstruct and prove one specific decision after the fact | No | Yes |
| Flag a statement that contradicts the record before it stands | No | Yes |
| Deploy voice agents in a regulated setting with confidence | No | Yes |
| Defend a decision to an auditor or regulator | No | Yes |
FAQ
Common questions from voice-AI teams that already run testing and observability and are deciding where the trust layer fits.
Q1Is CogniSwitch an alternative to Cekura?
No. They operate at different layers of the stack, and they serve different buyers. Cekura is automated QA, testing, and observability for voice and chat agents: it simulates conversations and monitors live calls to confirm the agent behaves well. CogniSwitch is the trust layer: it verifies the content of a decision against the record and policy. They barely overlap, so regulated voice teams run both.
Q2What does Cekura do that CogniSwitch does not?
Cekura tests and observes how a voice or chat agent behaves. It runs synthetic users against the agent before production, evaluates the full session with LLM-based judges, and monitors live calls with voice-specific signals like gibberish detection, interruption tracking, latency, and sentiment. CogniSwitch does not test conversational behavior. It adds a verification layer that checks what the agent stated against the authoritative record.
Q3What does CogniSwitch add to a Cekura stack?
Content verification and an audit trail. Cekura tells you the agent handled the call well: it heard the caller, followed the script, and completed the task. CogniSwitch proves whether the information the agent conveyed was correct against the payer policy, discharge plan, or clinical record, with a reproducible, rule-named verdict you can hand an auditor.
Q4Why isn't Cekura's testing enough for a regulated voice decision?
Cekura confirms the agent behaved as expected: it responded smoothly, followed instructions, and stayed reliable across scenarios. That is the right tool for shipping a dependable voice agent. It does not prove that the specific guidance the agent gave a patient or payer was correct against the source of truth. An agent can pass every behavioral test and still state benefits that do not match the policy. Verifying the content against the record is a separate, deterministic step.
Q5Do Cekura and CogniSwitch run together?
Yes, as complementary layers, each doing its job. Cekura simulates and monitors the conversation so the agent behaves reliably. CogniSwitch verifies the content of each regulated decision the agent conveyed against the record and keeps the audit trail. You keep your voice-agent testing and add the ability to prove what was decided.
Q6We already use Cekura. What changes if we add CogniSwitch?
Your voice-agent testing and observability stay exactly as they are. What you gain is content verification and provenance: every regulated decision the agent conveyed is checked against the source policy or clinical record, deterministically, producing an audit trail. Cekura keeps answering 'does the agent behave well on this call?'; CogniSwitch answers 'can we prove what the agent decided was allowed?'
Get your voice agents into production.
Keep your voice-agent testing. Add the layer that proves the content of a decision matched the record and flags the statement that does not, before it stands. It runs on a context graph, not another model in the scoring path.
Keep reading
CogniSwitch also completes Arize + CogniSwitch, Galileo + CogniSwitch, and Braintrust + CogniSwitch.
Author
Joshua Thomas
Co-Founder & CTO, CogniSwitch
Reading Time
~9 min read
References
- 1.Evaluating large language models for drafting emergency department encounter summaries — PLOS Digital Health, 2025
- 2.A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation — npj Digital Medicine, 2025