Self-Hosted Voice AI
Your callers' data never leaves your infrastructure.
On a managed platform, every second of caller audio is processed on someone else's servers, in another country. We build voice agents that run on hardware you control — your cloud region, your premises, or fully air-gapped.
Should you buy this?
Three reasons to own the stack. Only three.
The law requires it — data-residency rules under Saudi PDPL, UAE PDPL, DIFC or HIPAA: caller audio and transcripts can't be processed abroad
Volume changed the math — past ~25,000 minutes a month, owning beats renting. At 60,000 min/month it saves around $5,500 a month
A client contract forbids third-party processing — their compliance obligations are now yours
If none of those is you, don't buy this. Below ~15,000 minutes a month, self-hosting costs more. Start on the managed track, prove the use case, and we migrate you when volume justifies it — your conversation design and integrations carry over. Managed track →
| Monthly minutes | Managed | Self-hosted | Note |
|---|---|---|---|
| 5,000 | ~$1,700 | ~$2,400 | Managed wins |
| 15,000 | ~$3,200 | ~$2,900 | Roughly even |
| 30,000 | ~$5,700 | ~$3,500 | Saves ~$2,200/mo |
| 60,000 | ~$10,200 | ~$4,700 | Saves ~$5,500/mo |
| 150,000 | ~$24,700 | ~$8,300 | Saves ~$16,400/mo |
Usage + support + infrastructure, at a $0.15/min blended managed cost.
Two options
Choose by what your rules actually require.
Everything is identical — same agents, same integrations, same ownership, same handover. One question decides your tier: may an AI provider process your data at all, once identifiers are removed?
SHIELDED
from $10,000
one-time build fee · 6–8 weeks
AED 36,700 · £7,900
Your infrastructure, your data. Personal identifiers stripped before anything reaches an AI model.
Right for
regulated businesses that may use cloud AI providers provided personal data is protected, and high-volume operators who want the economics without running models themselves.
ISOLATED
from $15,000
one-time build fee · 8–12 weeks
AED 55,050 · £11,850
Nothing leaves. Not audio, not text, not metadata. Ever.
Right for
absolute data-residency requirements, air-gapped environments, or any regulator or contract permitting no external processing.
Air-gapped deployment
Shielded
Isolated
Compute requirement
Shielded
Standard compute
Isolated
GPU capacity
Build time
Shielded
6–8 weeks
Isolated
8–12 weeks
See exactly what this means for Shielded, component by component, below ↓
Both tiers include the full integration suite — see all 9 modules
Included in both tiers
- CRM integration
- Knowledge base (RAG)
- Calendar / booking sync
- Payments
- Outbound campaigns & follow-ups
- Messaging channels (WhatsApp / SMS / webchat)
- Supervisor tooling & QA
- Audit logging
- Full source + infrastructure-as-code handover
Support scope and response times are confirmed at scoping, alongside your infrastructure and hardware sizing.
Shielded — What runs where
Exactly what crosses your perimeter.
| Shielded | Isolated | |
|---|---|---|
| Language model runs | Provider API | Your hardware |
| What leaves your perimeter | Call audio · de-ID text · response text | Nothing |
| Call handling | Your cloud | Your cloud |
| Conversation logic | Your cloud | Your cloud |
| Logs, recordings, transcripts | Your cloud | Your cloud |
| CRM / booking / business data | Your systems | Your systems |
| PII de-identification layer | Included | Not needed — nothing leaves |
| Third-party outages affect you | Yes | No |
| Model choice | Any frontier provider | Open-weight models |
If you're unsure, you're probably Shielded. We'll say so at scoping rather than sell you the larger build. Isolated exists for requirements that are genuinely absolute.
What the PII shield does on Shielded+
Names, dates of birth, national ID and passport numbers · phone numbers, emails, addresses · card numbers, IBANs, account numbers · medical record and patient identifiers · your own patterns — all tokenised before text leaves your perimeter, and restored locally so the output stays useful.
You also get: a redaction log for every call · detection rules tuned to your vocabulary · a measured accuracy report on your own real calls before go-live.
Deployment
Your account. Your infrastructure. Your bill.
Deployed into environments you control
- Client cloud
- On-premises
- Air-gapped
- Deployed into your cloud or hardware, in whatever region you require. We host nothing
- We work through a scoped service account — deploy and monitor the voice stack only, never your wider environment
- Revocable anytime, with no effect on the running system
- End the engagement and everything keeps running — you hold the account, the infrastructure and the source
You'll need: a cloud account or hardware in your region · someone who can approve infrastructure provisioning · (Shielded) your own provider accounts · infrastructure budget, sized and quoted at scoping.
Proof
We've shipped this, not just scoped it.
- A fully local voice pipeline — speech recognition, language model and voice, all on the client's own hardware, for a client requiring zero cloud dependency
- An on-premise AI document review system for forensic reports, running entirely inside the client's environment — Read the case study →
- 7+ years building production AI across healthcare, call centres and sales
What we won't claim
Compliance certification.
We build the technical capability that makes data residency achievable — where data flows, where it's stored, what's logged — and hand your compliance specialist the architecture documentation they need. Certifying it against HIPAA, PDPL or GDPR is their work, not ours. Be wary of any vendor who tells you otherwise.
Why it's safe to say yes
Nothing here asks you to commit blind.
Technical discovery comes first, and stands alone — architecture, data-flow map, infrastructure and GPU sizing, deployment plan and cost model. Scoped and priced separately, credited in full against the build. The documents are yours whether or not you proceed
Hear it before you provision anything — we build a working prototype of your hardest call flow in our own sandbox, so you judge the conversation before your team touches infrastructure
Success is a number in the contract — a containment target agreed at discovery and tested against your real calls
Milestone-based delivery — payments tied to deliverables you can inspect, not to a calendar
Pilot on a defined slice of live traffic before full rollout, tuned daily, with exit criteria agreed in advance
Revoke our access anytime — the agent keeps running. You hold the account, the infrastructure and the source
Your security review, procurement process and compliance approvals set the real timeline. We plan around them and tell you at scoping which milestones depend on your side.
FAQ
Questions
How is this priced?+
A fixed build fee plus a flat monthly retainer for engineering. You pay your cloud provider — and, on Shielded, your AI providers — directly, at cost. We never mark up usage.
Which languages work self-hosted?+
Whatever open-weight models genuinely support, not what a roadmap promises. We assess your language at scoping and demonstrate real output before you commit.
What hardware do we need?+
Sized at scoping against your concurrency and latency targets, then quoted before you commit. Isolated needs GPU capacity; Shielded runs on standard compute.
Can we start managed and move later?+
Yes — the right sequence for most businesses. Conversation design and integrations carry over. Managed track →
What if you disappear?+
You hold the account, the infrastructure and the full source. Revoke our service account and nothing stops. That's the architecture, not a promise.
Own the stack. Own the data. Own the economics.
Bring your volume, your jurisdiction and your constraints. We'll tell you which tier fits — including when the answer is “neither, yet.”