Deepgram Nova-3 and a Zanus AI phone workflow answer different buying questions. Nova-3 is a speech-recognition model; a business evaluating a phone service also needs conversation handling, routing and saved actions. Decide which layer you need before comparing a component price with an operational service.
Identify the product layer
Deepgram’s model documentation lists Nova-3 among its speech-to-text models. That does not describe the whole of Deepgram’s product range or establish which speech technology another vendor uses. This comparison concerns Nova-3 specifically, not every Deepgram offering.
For Zanus AI, use the official product overview and a demonstration to establish the phone workflow available in your proposed configuration. Do not infer a Deepgram integration, underlying model or accuracy level without confirmation.
Separate transcript quality from business completion
| Layer | Evidence to request |
|---|---|
| Speech recognition | Correct words, names and corrections on representative audio |
| Conversation handling | Appropriate responses and interruption behaviour |
| Business action | The correct saved appointment, task or record update |
| Recovery | An owned exception when an action cannot finish |
A perfect transcript can still feed an incorrect action. A smooth conversation can also conceal a transcription error that changes an address or appointment time.
Use a representative, approved test set
Prepare internal test audio containing the names, locations and vocabulary your calls actually use. Include background noise, interrupted sentences and corrected details. Use the same audio when evaluating recognition, and keep the expected text available for review.
Then test the complete phone workflow separately. Ask for a booking, correct the requested date and inspect the final calendar record. Record where each failure occurred rather than attributing every problem to the speech model.
Budget the full responsibility
If buying a recognition component, include the work needed to connect telephony, conversation logic, speech output, integrations and monitoring. If buying a finished workflow, include setup, usage, review and exception handling. Obtain current written prices for the exact scope.
Choose based on the layer your team is equipped to own. A development team may want component control; an office may want an accountable service process. This guide proposes comparison criteria and reports no hands-on performance ranking of either vendor.
Questions and answers
Does using Nova-3 alone create a complete receptionist?
Nova-3 is a speech-to-text model. A receptionist workflow also needs the surrounding conversation, phone handling and business actions, whether built or supplied separately.
Mapping the decision before you contact a vendor
Before a business reaches the point of requesting a demonstration, it is worth establishing which operational problem is actually being solved. The distinction between a transcription component and a phone workflow service shapes every subsequent evaluation question, including which staff members should be in the room, which existing systems need to be described, and which failure modes carry the highest cost. A procurement conversation that starts at the wrong layer tends to produce a technically accurate answer to the wrong question.
Consider two illustrative scenarios that sit at opposite ends of this spectrum. In the first, a software team wants to add spoken search to an internal tool. They need a reliable way to convert speech to text and will write the surrounding logic themselves. Transcript accuracy on their specific vocabulary is nearly the entire evaluation. In the second, a small professional services office wants callers to be able to book, rescheduled or cancel appointments without staff involvement during busy periods. They have no development capacity. Transcript accuracy still matters, but it is one dependency inside a larger chain: the call must be answered, the conversation must reach the correct intent, a booking record must be written correctly, and an exception must be owned when something goes wrong. Evaluating a speech recognition model in isolation would leave most of the second team's buying question unanswered.
Neither scenario is more sophisticated than the other. They simply require different purchasing conversations. The practical discipline is to write down, before any vendor contact, which of the following layers your organisation needs to own, build, or buy already assembled.
The layers a phone workflow actually contains
Telephony infrastructure handles the inbound call — number assignment, routing rules, hold behaviour, fallback to a human, and call recording where permitted. Conversation management governs how the system interprets what a caller says, handles ambiguity, manages clarification exchanges, and decides when to escalate. Speech recognition converts spoken words to text and is one input into conversation management, not a synonym for it. Business action execution writes or updates a record in a calendar, CRM, ticketing system, or other store of truth. Notification delivery sends confirmations to callers or staff. Exception ownership determines what happens when any layer fails — who is informed, how quickly, and what fallback applies.
A business evaluating a finished workflow service should be able to get a clear answer about each of these layers from a vendor before signing a contract. A business buying a recognition component is taking explicit responsibility for all layers it does not purchase. Both are legitimate choices; neither should be made by accident.
Designing a test set that reflects your actual calls
Generic demonstration audio rarely predicts performance on calls a specific business actually receives. A useful test set is built from the vocabulary, names, and conversational patterns that appear in real operational calls. If the organisation handles calls about specialist equipment, medication names, place names with unusual pronunciation, or surnames common in the communities it serves, those should appear in the test audio. If callers frequently correct themselves mid-sentence, the test audio should include self-corrections. If there is background noise from a busy reception area or a vehicle, that should be represented.
Preparing this material serves two purposes. First, it gives a more informative signal about recognition accuracy than any published benchmark. Second, it forces the evaluation team to articulate what a correct result looks like, which is a necessary step regardless of which vendor is being considered. A test set without a documented expected output is not a test set; it is a listening session.
Structuring individual test cases
Each test case in a well-constructed set has an input, an expected output, and a field for the observed output. For recognition evaluation, the input is an audio clip and the expected output is the correct transcript. For workflow evaluation, the input is a complete fictional caller interaction and the expected output is the correct record in the connected system. Keeping these two evaluation tracks separate matters because a failure in the second does not necessarily indicate a failure in the first.
Illustrative fictional test cases might include a caller booking an appointment for a patient named Siobhan Kowalczyk, then correcting the requested date from Tuesday to Thursday; a caller whose connection quality drops mid-sentence while providing a reference number; a caller who gives a street address and then immediately gives a different one; and a caller who asks a question the system is not designed to handle. Each of these exercises a different part of the workflow and surfaces a different category of failure.
For the correction case, the observable evidence of success is not that the system heard the correction but that the final saved record reflects Thursday, not Tuesday. For the dropped connection case, the observable evidence is that the system either recovered gracefully or escalated to a fallback rather than silently writing a partial record. For the out-of-scope question, the observable evidence is a defined behaviour — escalation, a clear decline, or an offered alternative — rather than a confident but wrong answer.
Assigning ownership before go-live
One of the more consequential pre-deployment conversations is the one about who owns each failure mode. This conversation tends to be deferred until after a problem occurs in production, at which point it becomes a dispute rather than a plan. Running it in advance, as part of evaluation, changes the nature of the vendor relationship and surfaces mismatches in expectation before they become operational incidents.
Ownership questions worth documenting explicitly include the following. If a caller's booking is written to the system with an incorrect date because of a recognition error, who is responsible for identifying it, correcting the record, and notifying the caller? If the telephony layer drops a call before the booking is confirmed, what does the caller experience and what record, if any, exists of the attempt? If the connected calendar or CRM is unavailable at the moment the system tries to write a record, does the system retry, queue the action, escalate to a human, or fail silently? If the system interprets an ambiguous intent incorrectly — for example, treating a rescheduling request as a new booking — how is that detected and by whom?
For a finished workflow service, the answers to these questions should come from the vendor and be documented in the service agreement. For a component-based build, the answers come from the team designing the surrounding system. In either case, a business that cannot answer these questions before go-live is accepting an undefined failure state rather than a managed one.
Staff handover and escalation design
Even a highly capable phone workflow will encounter situations it cannot resolve. The design of the handover moment — when the system decides to involve a human, what it communicates to the caller during that transition, and what information it passes to the staff member — has a significant effect on caller experience and on staff workload. A handover that arrives with a summary of what the caller said and what the system attempted is qualitatively different from one that delivers a caller who must repeat everything from the beginning.
When evaluating a workflow service, it is worth running a test specifically designed to trigger escalation and then observing what the receiving staff member actually sees. Is there a record of the conversation? Is the caller's stated intent captured? Is the system's last action noted? These details are not decorative; they determine whether the escalation saves time or creates additional work.
Criteria that change a purchase decision
The following criteria are proposed as a structured framework for distinguishing between a component purchase and a workflow service purchase. Each criterion is paired with the reason it matters operationally, because a list of factors without reasoning does not support a decision.
Development capacity
If the organisation has staff who can build and maintain a conversation layer, handle telephony integration, and monitor recognition quality over time, a component approach is technically viable. If it does not, buying a component creates a dependency on future technical work that may not be resourced. This matters because the cost of a component is not the price of the component; it is the price of the component plus the cost of everything built around it, including the ongoing maintenance when the surrounding systems change.
Vocabulary specificity
Organisations with highly specific vocabulary — specialist fields, unusual proper nouns, non-standard addresses — should test recognition on that vocabulary specifically rather than on general language. The gap between general performance and domain-specific performance can be significant, and it may differ between vendors in ways that are not visible from published documentation. This criterion affects which vendor is worth pursuing further, not just how the evaluation is structured.
Call volume and pattern
A business with predictable, high-volume, structurally similar calls — for example, appointment bookings that follow a consistent conversational pattern — presents a different workflow design challenge than one with varied, complex, or emotionally sensitive calls. The former is more amenable to automation; the latter may require a more carefully designed escalation path. Understanding which category applies before vendor conversations begin helps set realistic expectations about what automation can usefully handle.
Record system integration
A workflow that does not write to the system of record the business already uses creates a reconciliation problem. Before evaluating any vendor, document which system or systems would need to receive the output of the workflow, and establish whether each vendor can connect to them in the proposed configuration. Do not assume a connection exists; confirm it for your specific setup and obtain that confirmation in writing.
Exception visibility
Some workflows surface exceptions clearly — a dashboard, an alert, a queued item — while others may fail without producing an obvious signal. For a business where a missed booking has a meaningful operational consequence, the visibility of exceptions is a criterion that should be evaluated explicitly rather than assumed. Ask, during a demonstration, how a failed booking action appears to the staff member responsible for reviewing it.
Caller identity versus request
In some operational contexts, the identity of the caller and the content of their request are handled by the same system. In others, a caller's identity may need to be verified before their request is acted upon. These are distinct requirements, and a workflow that handles one does not necessarily handle the other. If the business needs to verify that a caller is who they claim to be before making a change to a sensitive record, that requirement should be stated explicitly during evaluation rather than inferred from a general description of the product.
An acceptance exercise before going live
An acceptance exercise is a structured test conducted on the production or near-production system, using the agreed test set, before the workflow handles real callers. Its purpose is to confirm that the system as configured — not as demonstrated in a vendor environment — behaves as expected on the specific calls the business intends to process.
A practical acceptance exercise covers at least the following: a standard successful transaction from start to confirmed record; a transaction that includes a caller correction; a transaction that includes an out-of-scope request; a transaction that triggers escalation; and a transaction attempted when a connected system is unavailable. For each, the evaluating team records the input, the expected output, and the observed output, and signs off or raises a remediation request.
Observable evidence worth checking in each case includes the final state of the connected record, the content of any confirmation sent to the fictional caller, the information available to a staff member at the point of escalation, and the audit trail showing what the system attempted. A system that produces the correct outcome without a legible audit trail is harder to operate safely over time, because staff cannot distinguish between a case that completed correctly and one that completed incorrectly but appears similar.
Running this exercise before go-live also establishes a baseline for future comparison. If recognition quality or workflow behaviour changes over time — because the underlying model is updated, because the connected system changes, or because call patterns shift — the acceptance exercise provides a reference point against which current behaviour can be measured. Retaining the test audio, the expected outputs, and the observed results as internal documentation is a straightforward operational discipline that pays for itself the first time a regression needs to be investigated.
The most durable outcome of a careful component-versus-workflow evaluation is a written internal document that records what the organisation decided to own, what it decided to purchase, what the agreed scope of each vendor's responsibility is, and what the escalation path looks like when something falls outside that scope. That document, reviewed at the point of renewal or when operational needs change, is more useful than any benchmark score because it reflects the actual context in which the system operates.
Sources and next steps
This independent Sysevo guide proposes evaluation steps and reports no hands-on vendor test or compliance certification. Check the linked primary sources and Zanus AI’s current Front Office overview, then confirm the scope of your proposed deployment in writing.
Read Zanus AI voice agents for related context. To compare providers using the same requirements, explore Sysevo AI voice agents or book a Sysevo demo. Sysevo publishes this guide and offers its own business software.
Related workflow guides
- Zanus AI Voice Latency: Test Interruptions and Slow Actions
- FreJun Teler vs Zanus AI: Build Voice or Buy a Workflow?
For current Sysevo subscriptions and usage options, compare plans and pricing.



