Platform

Newo Zero-Hallucination Architecture: the orchestration layer that keeps a Voice AI Agent grounded in your business

Two control systems run alongside every conversation. One verifies what the agent actually knows and directs the next step. The other detects what the agent committed to and executes it through gated, checked actions.

Published 11 September 2026 · Updated 11 September 2026

A Voice AI agent hallucinates when a language model is left to answer on its own. The model produces fluent language and has no way to check whether the answer is true for your business. The fix is orchestration: a control layer around the model that verifies each turn against approved business information, tracks what has actually been confirmed, and routes real actions through gated tool calls instead of leaving them to the conversation. Newo Zero-Hallucination Architecture is Newo’s implementation of that control layer inside the Newo Agent Framework.

Why does Voice AI hallucinate in the first place?

A large language model on its own is an unreliable foundation for a business phone line, because it is optimized to produce a plausible answer rather than a verified one.

Three limits matter here. A model has no built-in knowledge of your prices, your policies, or tomorrow’s availability, so it will fill the gap with something that sounds right. It does not hold state across a conversation on its own, so it can lose track of what the caller already told it. And it cannot reach your live systems unless something puts them in its path.

That is acceptable for a draft. It is not acceptable on a call where someone is deciding whether to book.

So Newo does not put a model on the phone and hope. The model handles language. A separate control layer handles what is true, what has been confirmed, what happens next, and what the agent is permitted to do.

What is Voice AI orchestration?

Voice AI orchestration is the layer that sits between a language model and a live conversation, deciding what the agent should say next, what it is allowed to do, and when a person should take over.

Most Voice AI products are a pipeline: speech in, model, speech out. Whatever the model produces reaches the caller. Orchestration replaces that straight line with a loop. Every turn is analyzed while the conversation is still running, and the result of that analysis shapes the next turn.

Newo Zero-Hallucination Architecture is that loop. It hangs off a single platform signal: when the agent finishes a response, the framework fires a turn-analysis event, and two independent control paths pick it up and run in parallel against the turn that just completed.

Agent responds
The turn completes
Turn analysis event
Fired by the framework
In parallel
Observer
Verify and direct

Supervisor
Detect and execute
Next turn is steered
The loop repeats

How does the Observer keep answers grounded?

The Observer is a parallel reasoning system that extracts what the caller has said, checks it against what the agent still needs, and produces the specific next steps the agent should take.

The grounding device that matters most is field tracking. Every piece of information the agent is expected to collect sits in one of four states, and the difference between them is the difference between knowing something and assuming it.

State	What it means
Not known	No value yet.
Detected	The value came from context, such as a phone number from caller ID, rather than from the caller saying it.
Gathered	The caller stated it explicitly, as the step required.
Reconfirmed	The value was read back and confirmed again.

A detected value is tracked separately from a gathered one, so the agent is pushed to confirm a detail rather than proceed on an assumption. Acting confidently on an assumption is one of the most common ways an AI conversation goes wrong, and it is closed off by design here.

Around that, the Observer assembles the working context for the turn: your business information, operating constraints, schedule, procedures, availability, and the current state of the task. From that it produces the next sequential steps, referenced to the specific scenario and procedure, in the language the caller is speaking.

The Observer also scores conversation quality as it goes. When a conversation degrades, that score triggers a recovery path such as a handoff to a person or an offer to call back, rather than letting the agent continue to struggle.

How does the Supervisor stop the agent from taking the wrong action?

The Supervisor detects what the agent has actually committed to and executes it through explicit, gated tool calls, so a real action is never left to the text of the conversation.

Booking an appointment, sending a text, transferring a call, checking coverage: these have consequences. The Supervisor compiles the tools available to that agent, evaluates each one against what the agent just said, resolves conflicts when more than one could apply, and runs the ones that qualify.

The point is to act only on a genuine commitment. Five exclusion rules keep the agent from firing an action it merely mentioned.

Rule	What it blocks	Example the agent said
In progress	Something still underway	I’m just checking that now.
Completed	Something already done	I’ve sent that over.
Future step	Something described as a later step	And then I’ll email the confirmation.
Offer	A suggestion or a question	Would you like me to book that?
Subject check	Anything where the agent is not the one acting	You’ll receive a text shortly.

There is a second guardrail underneath the first. Each built-in tool is gated behind a setting your business controls, so an action can run only if you enabled it for that agent. The per-turn check decides whether the agent meant it. The configuration gate decides whether it was ever permitted.

How does this work on a live phone call?

On voice-to-voice conversations the direction is applied mid-call, by injecting the Observer’s next-step directives into the speech model as context for the following turn.

On a live call there is no window to review a response before the caller hears it. So the correction happens before the fact rather than after it. An agent that has started to drift is pulled back toward the verified next step while the conversation is still running.

This is what extends the control loop from text into spoken conversation, where the timing is unforgiving and a wrong answer cannot be recalled.

Which use cases does Newo Zero-Hallucination Architecture make better?

The architecture matters most where a wrong answer costs money, breaks trust, or sends a caller to the wrong place, which is to say the calls a business cannot afford to hand to an unsupervised model.

Six examples, and the specific control that carries the weight in each.

IVR transfer and call routing for customer service

The failure it prevents

A caller is sent to the wrong queue, has to repeat everything, and the routing error is only discovered by the person who picks up.

How the architecture helps

The Observer holds the reason for the call, the details the caller confirmed, and the routing rules your business defined, then produces the next step against those rules rather than against the model’s guess. The Supervisor executes the transfer only when the agent has actually committed to it, and the context travels with the call so the person receiving it starts informed.

Relevant to: contact centers, B2B telephony, multi-location service operations.

Insurance verification for dental and orthodontics

The failure it prevents

An agent tells a caller their plan is accepted, or quotes what their visit will cost, based on a plausible guess. That answer sets an expectation the practice then has to walk back at the front desk.

How the architecture helps

Coverage questions are answered from the practice’s approved information and the systems the practice uses, not from the model’s general knowledge of insurance. Field tracking keeps carrier, plan and member details in a confirmed state before anything is checked or recorded, and a question outside the approved scope routes to a person rather than being answered.

Relevant to: orthodontic groups, dental groups, multi-location practices.

After-hours consultation booking

The failure it prevents

An evening caller is offered a slot that does not exist, or is booked against the wrong provider, and nobody notices until the morning.

How the architecture helps

Availability is read from the calendar and practice software rather than inferred, and the booking is a gated tool call rather than a sentence the agent produced. If the caller’s details are only detected rather than gathered, the agent confirms them before writing anything.

Relevant to: orthodontics and dental, healthcare, home services, any appointment-driven business.

Reservations and guest policy questions for restaurants

The failure it prevents

An agent invents a large-party policy, a corkage rule, or a private-dining minimum. The guest arrives holding the restaurant to something it never said.

How the architecture helps

Policy answers come from what each restaurant approved, which matters when one Voice AI Agent serves many locations with different rules. The reservation itself is written through a gated action into the platform the restaurant already uses.

Relevant to: restaurant technology platforms, hospitality groups.

Service area, pricing bands and dispatch for home services

The failure it prevents

The agent promises a same-day visit outside the coverage area, or quotes a number the business does not honor, and a dispatcher has to make the apology call.

How the architecture helps

Coverage and pricing sit in approved business information rather than in the model, so an address outside the area produces the answer your business defined rather than an accommodating one. Job details move to a confirmed state before dispatch is triggered.

Relevant to: HVAC, plumbing, cleaning, and other dispatch-driven trades.

Multi-location operations where the rules differ by site

The failure it prevents

An agent applies one location’s hours, policies or services to a caller asking about another, which is the most common way a rollout that worked at one site fails at ten.

How the architecture helps

The working context is assembled per conversation, so the agent answers from the right location’s information. A proven configuration can be extended to more locations without each site relying on the model to remember which rules apply.

Relevant to: franchise groups, dental and orthodontic groups, multi-unit restaurants, regional service brands.

What this architecture does not do

Stating the limits is part of the design, and any vendor who will not state them is worth a second look.

It does not make errors impossible. It reduces the conditions that produce them and catches classes of failure before they reach a caller.
It does not invent knowledge of your business. The agent works from the information, policies and procedures you provide, so the quality of that source material sets the ceiling.
It does not correct a caller who is misheard. Speech recognition still has to hear the word. What the architecture does is keep an uncertain value from being treated as confirmed.
It does not expand its own responsibility. New tasks, new tools and material changes in behavior are configured and approved by your team.
It does not remove the need for a person. Escalation paths exist because some conversations should reach a human, and the quality score is there to route them.

FAQ

Frequently asked questions

How do you stop Voice AI from hallucinating?

You orchestrate it. Put a control layer around the language model that verifies each turn against approved business information, tracks which details the caller actually confirmed, and routes every real action through gated tool calls. Newo does this with Newo Zero-Hallucination Architecture, which runs two verification systems in parallel with every conversation.

Can AI hallucination be eliminated completely?

No, and any vendor claiming otherwise is overstating. Hallucination can be reduced substantially by grounding responses in approved information and by designing the agent to say it does not know and hand off, rather than to guess. Newo Zero-Hallucination Architecture is the name of that control design, not a claim of perfection.

What is Voice AI orchestration?

Voice AI orchestration is the layer that decides what a Voice AI Agent should say next, what it is permitted to do, and when a person should take over. Without it, a Voice AI product is a pipeline where whatever the model produces reaches the caller unchecked.

Is a Newo Voice AI Agent an AI Employee?

That is the idea behind it. An AI Employee is a Voice AI Agent given a defined job, the information it needs to do that job, the systems it is allowed to act in, and the point at which it hands work back to a person. The orchestration layer is what makes the job description enforceable rather than aspirational.

What happens when the agent does not know the answer?

It says so and moves the conversation to a person, rather than producing a confident guess. The Observer scores conversation quality as it runs, and a degrading conversation triggers a handoff, a callback offer, or another recovery path your business defined.

Does this work on live phone calls or only on chat?

Both, and voice is the harder case. On voice-to-voice conversations the next-step directives are injected into the speech model mid-call, so the agent is corrected before it speaks rather than after.

How do I test whether it actually works?

Call it and try to break it. Interrupt it, change your mind mid-sentence, ask something outside its job, and watch whether it guesses or hands off. Newo’s position is that the honest test is your own hardest calls, not a scripted demo.

To know it is to experience it.

Reading about a control loop is not the same as hearing one work. Call a Newo Voice AI Agent, interrupt it, change your mind, and ask it something it should refuse to answer.

Experience Newo Back to the platform

Experience the accuracy and speed for yourself.

HIPAA-compliant. SOC 2 Type 1 (Type 2 in progress). GDPR. 256-bit encryption.