● Voice & chat AI for regulated teams

Safe, sovereign AI for customer conversations.

Bespoke voice and chat agents that run on your infrastructure, answer only from sources you approve, and never keep your data.

  • Zero data retention
  • HIPAA-ready
  • Your keys, your cloud

Safe by default, on every channel

Scripted demo

Try a conversation

A scripted preview of Box One. Real deployments answer only from the systems and policies you approve, with every exchange logged.

Why sensitive organizations choose Box One

Request access
Your answers are written down01
On holdis where your callers wait for answers you already wrote down
PoliciesSchedulingBillingClaims
01

Your answers are already written down

Policies, program rules, scheduling, billing, claims. Your team answers the same questions by hand, all day, while callers wait. Most AI tools can’t be trusted with those conversations, because they guess, keep your data, and can’t prove what they said.

Box One is built for exactly those conversations.

Safe by design02
Citedevery answer, every time
HR handbook · p.12#people · Mar 3
02

Safe by design, not by promise

Box One answers only from sources you approve and cites every answer. It verifies identity before disclosing anything personal, never gives medical, legal, or financial advice, and hands off to a person when the source doesn't support an answer.

Before launch we red-team it on your real conversations: prompt injection, data-extraction attempts, and off-script questions. It doesn't go live until it passes.

Refusal behavior gets evaluated as rigorously as accuracy. A confident wrong answer is worse than no answer.

It runs inside your walls03
Yourscode, prompts, evals, docs
your cloudzero retentionno platform fee
03

It runs inside your walls

Deployed into your own cloud accounts, respecting the permissions your systems already enforce. Model calls run under zero-retention, no-training terms, and the model sits behind an interface you own, so a better one swaps in without a rewrite.

At handoff the repository is yours. Stay with us for ongoing engineering, or take the keys.

Safe · Sovereign · Secure

Built for conversations that can’t go wrong.

Healthcare, finance, insurance, and public agencies can’t afford an AI that guesses, leaks, or forgets what it said. Every Box One deployment ships with these safeguards. None of them are add-ons.

01 · Grounded

No guessing

Answers come only from sources you approve, and every answer cites where it came from. When a source doesn’t support an answer, it says so instead of inventing one.

02 · Guardrails

Stays in its lane

No medical, legal, or financial advice. Topics you rule out stay ruled out. Off-script, risky, or emotional conversations go straight to a person, with context passed along.

03 · Identity

Verifies before it discloses

Callers and chatters are verified before anything personal is shared. PHI and PII are redacted from logs and transcripts by default.

04 · Red-teamed

Attacked before launch

We test every deployment against prompt injection, jailbreaks, and data-extraction attempts on your real conversation types. It doesn’t go live until it passes, and the tests re-run whenever the model changes.

05 · Zero data retention

Nothing kept, nothing trained

Every third-party model call runs under the provider’s zero-data-retention, no-training terms, disclosed to you before any data is sent. Your data never trains a model, ours or anyone’s.

06 · HIPAA-ready

Built for protected health information

PHI stays in your environment. Encrypted in transit and at rest, role-based access, HIPAA-eligible model endpoints only, and the administrative and technical safeguards your compliance team will ask for, documented.

07 · Sovereign

Your infrastructure, your keys, your model

Deploy on your own servers or in your own cloud account, in the region you choose. You hold the encryption keys and the model keys. If your policy rules out external APIs, Box One runs on open-weight models hosted entirely inside your network.

08 · Secure comms

Secure lines to your data and your customers

Connections to your records, scheduling, and billing systems run over private, encrypted channels with least-privilege access. Calls and chats with customers are encrypted in transit, and nothing is stored outside your environment.

09 · Audit trail

Every word on the record

Every call, chat, source, tool call, and handoff is logged in your environment, searchable, and exportable for audits and complaints. You can always show exactly what was said and why.

Box OneTypical AI voice & chat platforms
Data retentionZero retention, no training, disclosed per call pathVaries by vendor and plan
SovereigntyYour servers or cloud account, your region, your keysThe vendor’s multi-tenant cloud and keys
Connections to your dataPrivate, encrypted, least-privilege, inside your networkYour data synced out to the vendor
Protected health informationStays in your environment, HIPAA-eligible endpoints onlySent to the vendor’s cloud
Safety testingRed-teamed on your conversations before launch and on every model changeGeneric, vendor-wide
Audit trailEvery word, source, and handoff, stored by youWhatever the vendor’s dashboard keeps
How you pay for usageThe model provider’s rate, billed to you directly. No markup.Per seat, per resolution, or per minute
When you leaveYou keep the code, prompts, and evaluationsYou start over

Cost per use: no seat licenses, no per-minute or per-resolution fees, no markup on tokens. We benchmark models on your workload and pick the cheapest one that passes your safety and accuracy tests. Discovery ends with a written estimate of cost per call and per chat, next to what you pay today.

Safe
Yours
Early access

Bring Box One inside your company.

Tell us which conversations to take on first: your phone line or your website chat. We build it around your systems and your compliance requirements, deploy it on infrastructure you control, and the code, prompts, and evaluations are yours to keep.

Sovereign. Zero data retention. HIPAA-ready. We reply within two business days.

FAQs

FAQs

The questions we hear most, answered plainly. All questions

Request access →
You do. Box deploys into your own cloud, and the code, prompts, evaluation suites, and documentation are yours from day one. There is no proprietary runtime you are locked into.
All the major ones: Anthropic, OpenAI, Google, Meta's Llama family, Mistral, plus AWS Bedrock, Azure AI, and open-weight models self-hosted in your environment. We hold no reseller or referral agreement with any of them, so the recommendation comes from benchmarks on your workload.
No to both. Box never uses your data to train models for anyone, and every third-party model call runs under the provider's zero-data-retention, no-training terms, disclosed to you before any data is sent. Run fully on-prem and no data leaves your environment at all.
Yes. Each channel stands on its own. Most teams start where the volume is and add the other later. The second channel reuses the same approved knowledge, guardrails, and audit trail, so nothing is rebuilt.
Three layers. It answers only from sources you approve and cites them. Guardrails block topics you rule out, like medical, legal, or financial advice, and route anything risky to a person. And before launch we red-team it with prompt injection, jailbreaks, and data-extraction attempts on your own conversation types. It goes live only when it passes, and the tests re-run on every model change.
Box One is built to run HIPAA-regulated workloads. PHI stays in your environment, data is encrypted in transit and at rest, access is role-based, every exchange is logged, and model calls go only to HIPAA-eligible endpoints under zero-retention terms. HIPAA compliance is ultimately your organization’s program, and we document the safeguards your compliance team needs to sign off.
Yes. We deploy on your servers or in your own cloud account. For environments that can't call external APIs, voice and chat run on open-weight models hosted inside your network.
There is no platform layer taking a cut. You pay the model provider directly at their rate, or nothing per call on your own hardware, with no seat licenses and no per-resolution or per-minute fees. Because the model is swappable, we move you to a cheaper one whenever it passes your evaluations.