Voice agents: every call is the business speaking
Voice agents now handle support, sales, collections, KYC, scheduling and hiring calls at scale. Whatever an agent says or does on a call, the business that deployed it answers for it.
- Sector
- Voice AI agents in banking, insurance, lending, retail, healthcare and hiring
- Running agents like these?
- Talk to us
Transcriptlive
mera account balance bata do
sure, checking that now
account lookup
Syntrox · hard ruleBlock
No account lookup before the caller has passed verification.
Same rule in Hindi, Telugu or Englishother calls live
The setting
Voice agents now answer and place calls at high volume: support lines, sales follow-ups, payment reminders, KYC calls, clinic bookings and hiring screens. A single platform can handle millions of calls a month, in English, in Indian languages, and in the mix of the two that callers actually speak.
Three kinds of company build them:
- Agent platforms build and run agents for enterprises across support, sales, collections and HR.
- Journey builders deploy agents for specific high-volume journeys, such as reminders, renewals, drop-off recovery and KYC follow-ups, over the phone, the web and WhatsApp.
- Speech-model labs build the voices and the listening underneath, so an agent sounds native in Hindi, Telugu or Tamil.
Banks, insurers and hospitals are also starting to run agents of their own.
Whoever builds the agent, it speaks for the business that deployed it, and that business answers for every word and every action. So before a regulated business goes live, its security and risk review asks one question: what stops the agent saying or doing something we can't stand behind?
Where voice agents work, and what they must never do
| Journey | What the agent does | What it must never do |
|---|---|---|
| Support and IVR replacement | Answers account, order, booking and claim questions; looks up records; changes addresses, plans and bookings | Read out account details before the caller is verified; state a refund or coverage rule that does not exist |
| Sales, renewals and drop-off recovery | Calls new leads within minutes, qualifies them, books demos and visits, recovers abandoned carts and half-finished applications, renews policies | Promise a price, rate, discount or approval the business never offered; keep calling someone who opted out |
| Collections and payment reminders | Reminds about missed EMIs and due dates, takes a promise to pay, sends a payment link | Misstate what is owed; offer a waiver or settlement it is not allowed to; keep pressing a caller who says they are in hardship |
| Onboarding, KYC and verification | Follows up on incomplete KYC, confirms details, runs verification calls before a policy or loan is issued | Record consent the caller did not give; skip a required disclosure; mark a check complete when it is not |
| Healthcare scheduling and follow-up | Books, reschedules and reminds across clinics; follows up after visits | Book the wrong patient or slot; answer a clinical question it was never meant to answer |
| Hiring screens | Screens candidates and schedules interviews at volume | Ask a question the employer's policy forbids; tell a candidate something untrue about the role or the pay |
What goes wrong on a call, in the order it costs money
- Saying what the business never said. With nothing authoritative to retrieve, the agent fills the gap with plausible language: a refund rule, a coverage condition, a rate or a waiver that does not exist. Courts attribute an agent's statements to the business that deploys it. Air Canada was held liable for its chatbot's invented fare policy in 2024, and in May 2026 a German appeals court attributed a chatbot's statements to the business even where the system was carefully configured.
- Acting on a mishearing. Speech recognition gets a digit, an amount or a date wrong: "fifteen" heard as "fifty", one digit off in an account number. The transcript looks confident, and the agent books, charges or promises on it. Accents, noisy lines and mixed-language speech make this more likely.
- Doing what it is not allowed to do. A refund above the limit, a payment taken or an address changed before the caller is verified, a promise to pay accepted outside policy.
- Personal data in the wrong place. Account details read out to the wrong person; card and ID numbers left in transcripts and logs, or sent to outside services.
- Calling when it should not. Outside permitted hours, after an opt-out, or more often than the rules allow. In the US, the FCC confirmed in February 2024 that AI-generated voices count as "artificial" voices under the Telephone Consumer Protection Act, so its consent rules reach the outbound calls voice agents place.
- Missing the moment to hand over. A caller asks for a person, says they are in hardship or recently bereaved, or is plainly confused, and the agent keeps going.
- Being talked into it. A caller tells the agent to ignore its instructions, or talks it into reading out someone else's details.
- Runaway loops and silent drift. An agent redials or retries with no ceiling, burning spend with no outcome. Or its behaviour degrades as prompts, models and knowledge sources change. Nothing alerts, because nothing failed.
Why the builder's own tests don't catch it
- Tests run before launch, on scripted calls. Real callers interrupt, mumble, switch languages and try things nobody scripted. Fabrication, mishearing, injection and drift show up at runtime.
- Every business has different rules. A lender's limits are not a clinic's or a retailer's, and one generic filter cannot hold all of them.
- Guardrails check words. The costly failures are actions: the refund, the payment link, the booking, the dial.
What Syntrox does
Syntrox sits between the agent and the systems it can act on, independent of the agent itself.
- Checks each consequential action before it runs, whether a refund, a payment link, a promise to pay, a booking, a record lookup or an outbound dial, against that business's own rules, and names the rule behind each verdict.
- Judges what the agent is about to say against the business's approved policies before it reaches the caller.
- Works whatever language the call is in. The rules that matter most are about actions, not words: the largest refund the agent may issue, no account lookup before the caller has passed verification, no dialling outside permitted hours. These run as hard rules on the action itself, so they apply the same way in Hindi, Telugu or English.
- Masks personal data inline before it leaves the boundary.
- Stops runaway agents. Per-agent spend ceilings isolate a looping agent automatically; any other agent can be isolated by an operator without taking down the rest.
- Produces the evidence: a record of every consequential action, the verdict and the rule applied. It is the pack a builder hands to its customer's security and risk review.
Two numbers, always labelled: hard rules such as data masking run inline in under 20 ms; a judge verdict from Nucleus, our own 4-billion-parameter model, takes about 145 ms. Syntrox runs inside the builder's or the customer's own infrastructure.
Why it matters to each kind of voice company
- Agent platforms and journey builders: it is a sales asset more than a cost. Evidence that every consequential action was judged against the customer's own rules shortens security reviews and widens the set of regulated accounts they can win, such as banks, insurers, lenders and hospitals.
- Speech-model labs: it is one less reason for a regulated buyer to say no to agents built on their voices.
- Businesses running their own agents: the same checks and the same record, inside their own network.
What a first pilot looks like
One call flow, such as payment reminders or KYC follow-ups. A written list of what the agent may say and do. A shadow test on live or recorded calls that blocks nothing and measures three things: what Syntrox would have held, how often it was wrong, and how much time it added.