Build the agent. Then argue with it.
The Studio is where an agent stops being a prompt and becomes an artefact with an owner. You compose eight node types into a graph, attach guardrails written in your own policy language, dress it in a voice, and hand it to a reviewer before a single customer hears it.
Eight nodes. Every call walks them in order.
A graph is not a flowchart — it is the runtime's instruction set. Each node declares its inputs, its latency budget and its failure behaviour, and the runtime refuses to publish a graph with an unreachable node or an unhandled error path.
Trigger
Fires on an inbound DNIS, a campaign list row, a webhook or a missed-call callback. Declares business hours, maximum concurrency and the fallback number if the runtime cannot admit the call. Budget: 4 ms.
Fails how → routes to the carrier's own overflow rather than dropping.
Identity check
Matches the calling number against stored consent and grant-state, then steps up with a spoken date-of-birth or last-four challenge where the journey demands it. Three attempts, then a human.
Fails how → the conversation continues but everything downstream is read-restricted.
Intent resolve
Classifies the opening turns into one of your declared intents, with a confidence floor you set per intent. Below the floor the agent asks one clarifying question rather than guessing.
Fails how → falls back to the generic triage script and flags the call for review.
Knowledge lookup
Retrieves from the sources you nominate — product catalogue, price list, policy PDFs, order API — with citation attached to the answer. Uncited answers can be blocked outright.
Fails how → the agent says what it cannot confirm and offers a callback.
Policy guardrail
The deterministic pass. Every candidate response is checked against the rules attached to this node before it is allowed to be spoken. Denials are logged with the rule id and the offending phrase.
Fails how → the response is replaced, never softened; the caller hears a safe alternative.
Response plan
Chooses register, length and pacing for the reply: one sentence for a distracted caller, a structured three-part answer for a compliance call, jargon-free wording for a first-time buyer.
Fails how → defaults to the shortest safe variant.
Tool call
Typed, validated calls into CRM, calendar, payments and ITSM. Each tool declares a timeout, an idempotency key and a retry ladder; a failed write is surfaced to the caller, not swallowed.
Fails how → three retries, then human handoff with the partial state attached.
Wrap-up & CRM write
Writes disposition, transcript link, qualification score, sentiment and the exact sentences that produced them. Summarises in the CRM's own field shape, so your reps read it as if a colleague wrote it.
Fails how → queues the write durably and reconciles within 60 seconds.
What the compiler enforces
No unreachable nodes. No node without an error edge. No tool call before identity where the journey touches personal data. No publish without a rehearsal run against the current graph hash.
How runs are scoredA diff a reviewer can actually read.
Line-by-line prompt diffs are useless when the prompt is four thousand tokens. The Studio diffs behaviour: which intents moved, which guardrails tightened, which tools were granted, which sentences a caller would hear differently.
- Branching
- Any version can be branched. Branches inherit the parent's guardrails unless explicitly loosened, and loosening is highlighted in red on the review screen.
- Semantic diff
- Ghost runs replay 400 historical calls through both versions side by side, so the reviewer reads transcripts instead of configuration.
- Sign-off
- Two named roles: the agent owner and a compliance reviewer. Either can block; only the owner can request the canary.
- Rollback
- Any prior version is one click away and stays deployable for 400 days — the lifetime of the shortest retention policy we support.
- Export
- The whole version, guardrails included, exports as readable YAML for your change-management system of record.
Rules written the way your risk team already writes them.
428 rules ship with the platform, mapped to common regulatory and commercial obligations. You attach them to a node, set a severity, and cite the clause they implement. Eight examples, verbatim from the library:
“Before any price, interest rate or EMI figure is spoken, state the full cost of credit and the cooling-off period in one uninterrupted sentence. If the caller interrupts, restart the disclosure.”
Maps to: RBI fair-practices code §4.2 · Evaluated in 9 ms
“If the call reaches an answering machine, a minor, or a third party, end the call within one turn and mark the record no-contact-made. Never leave a callback number on a voicemail.”
Maps to: TCPA §227(b) · TRAI UCC 2018 · Evaluated in 6 ms
“Never state that an order has shipped, been approved or been refunded unless the upstream system returned a terminal status in this call. Pending is pending, and must be described as pending.”
Maps to: internal SLA-CX-7 · Added after incident NI-2291
“Do not give medical, legal or tax advice, even when asked directly. Acknowledge the question, state the limit of the agent's remit, and offer a specialist callback within one business day.”
Maps to: responsible-AI policy §2.1 · Triggers QA sampling
“If the caller's sentiment score falls below −0.4 twice in the same call, stop persuading. Move to acknowledgement, then offer a human. Do not repeat an offer that has already been declined.”
Maps to: conduct policy §7.3 · Escalation SLA 45 s
“Mirror the caller's language within two turns. If they code-switch mid-sentence, answer in the language of the last complete clause and keep proper nouns in their original form.”
Maps to: CX standard §1.4 · Applies to 42 locales
“Never read a full card number, CVV or UPI PIN aloud, and never accept one spoken. Route to a DTMF capture channel and confirm only the last four digits and the amount.”
Maps to: PCI DSS SAQ-A · DTMF channel enforced
“On the first unambiguous request for a human, transfer. Do not attempt one more qualification question, do not offer a callback as a substitute, and pass the full context to the receiving agent before the transfer completes.”
Maps to: conduct policy §2.9 · Warm transfer, 12 s target
The voice is a product decision, so we made it a dial.
Tone, pace, accent and pronunciation are separate controls with separate owners. Nobody has to retrain a model to stop the agent sounding like a call-centre robot at 9 a.m. and a motivational speaker at 9 p.m.
- Tone
- Six presets — warm-neutral, brisk-professional, reassuring, apologetic, advisory, celebratory — with a per-intent override. A collections call and a delivery confirmation should not sound alike.
- Pace
- 180–255 words per minute in five-point steps, plus automatic slowdown when the caller's own speech rate drops. Numbers are always read 15% slower than prose.
- Accent
- Nine English accents and five Indian-language registers. Set per journey; the agent never switches accent mid-call, because callers notice.
- Pronunciation
- A per-workspace lexicon for brand names, SKUs, place names and acronyms. 4,120 entries ship by default and overrides win.
- Disclosure
- A configurable AI-disclosure line, spoken once at the top of the call in the caller's language, recorded in the audit spine as a boolean.
These sixteen are production-grade with published WER and MOS per locale. The remaining 26 languages are available in beta and carry a “beta” flag in the audit trail — useful for pilots, flagged for anything regulated.
Rollout is a sequence, and every step can stop it.
A publish is not a deploy of code; it is a change in what your customers hear. Nivākya moves traffic in named stages with thresholds that halt automatically, so the person who pressed the button is not also the person watching the graph at midnight.
| Stage | Traffic | Duration | Automatically watches | Halts if |
|---|---|---|---|---|
| Shadow | 0% — mirrored audio | 20 min | Latency, ASR confidence, policy output on the same calls the incumbent takes | p95 turn rises above 1.1s or policy denial rate triples |
| Canary | 5% | 45 min | Task completion, escalation rate, sentiment in the last third of the call | Completion drops >4 points vs incumbent |
| Quarter | 25% | 30 min soak | CRM write success, tool-call error rate, repeat-call rate within 24h | Any tool error rate >1.2% |
| Half | 50% | 30 min soak | Queue pressure, human handoff latency, DNC-spike detection | Handoff SLA breached twice in ten minutes |
| Full | 100% | — | Everything above, for 72 hours | Rollback remains one click for 400 days |
| Rollback | Previous version | < 20 s | Confirming the caller experience did not visibly change mid-call | — |
Publishes to regulated agents are refused outside declared change windows unless an incident flag is open.
A version can be pinned to a single region, a single client account or a named test number without exposing it to everyone.
Every stage emits a receipt — who approved, what the metrics were, why it advanced — attached to the version for good.
Start from something that already worked.
Site-visit booker
Qualifies budget, timeline and location, then books a slot against the sales team's live calendar. Declines politely when the caller is three months out.
Typical completion 78% · Median call 2m 41s
COD confirmation
Confirms cash-on-delivery orders, repairs incomplete addresses, reschedules twice, then cancels with a reason code your ops team can act on.
Failed deliveries down 22% · Handles 6 languages
Appointment reminder
Confirms, reschedules and delivers prep instructions without ever reading clinical detail aloud. HIPAA-scoped storage from the first second.
No-show reduction 31% · BAA included
Soft collections
Identity-stepped reminders inside a calling window, with settlement authority capped by policy and every disclosure logged. Escalates to a human on the second refusal.
Promise-to-pay rate 41% · TCPA + DPDP ready
Renewal risk call
Opens with usage facts, asks one diagnostic question about the drop-off, and offers exactly one remedy. Never discounts without a signal.
Save rate 27% · Writes to CRM + CS platform
Post-service survey
Four questions, adaptive branching, and a verbatim capture that routes straight into Intelligence. Under ninety seconds or it closes itself.
Response rate 3.4× email · Median call 1m 12s
Bring one journey. Leave with an agent graph.
We will sit with your team for ninety minutes, map one real journey into the eight nodes, attach the guardrails it needs, and rehearse it overnight against your own scenarios.
Templates are free to fork. Rehearsal is included. Nothing reaches a customer until you say so.