Email AI that never leaves your perimeter.

Local-LLM inference inside the customer data boundary. Built for the regulated enterprises that keep saying no to cloud AI.

1M+
Emails processed
97%
Phishing detection
<1s
Per-email classify
60s
Scan cycle
NextEmail.ai Architecture Data flow: email enters from customer tenant, classified by local LLM inside perimeter, folder moves written back via API. Email body never crosses the boundary. CUSTOMER PERIMETER — zero egress Microsoft 365 or Gmail Customer email tenant No MX rerouting NextEmail.ai Inference Engine Fine-tuned open-weight LLM Runs on customer hardware No external API dependency + post-LLM heuristic safety nets + custom fine-tuning on private dataset Folder move / Label add Graph API · Gmail API write-back HEADERS + BODY VERDICT SYNC BACK Closed loop — nothing leaves the customer network TYPICAL CLOUD-AI PATTERN Email Cloud AI API × Email body crosses the perimeter to a third party. Regulated buyers say no.
How it works

Standard OAuth. Local inference. API write-back.

NextEmail.ai connects via standard OAuth with read/write scopes — no admin elevation, no MX record changes. Email arrives normally through the customer's existing provider.

The inference engine runs a fine-tuned open-weight LLM trained on a private threat dataset, executing entirely on customer hardware. Each email is classified in-process: headers and body in, structured verdict out. A suite of post-LLM heuristic safety nets catches the edge cases pure-LLM approaches miss — brand impersonation, attachment scams, sender spoofing, authentication failures.

The email body never leaves the customer's network. No cloud AI endpoint. Model weights, inference, and classification output all reside inside the customer data boundary.

Deployment

Three modes, one guarantee: your data stays put.

On-prem appliance

Full inference stack on customer site. GPU-accelerated classification inside the customer LAN.

Bare-metal · air-gapped

Customer VPC

Same stack provisioned in the customer's cloud tenant. Customer-controlled network, customer-owned compute.

AWS · Azure · GCP

Managed on-prem

On-prem appliance with encrypted health telemetry (never email content) for managed operations.

Managed ops · data resident
Trust surface

What we touch. What we don't.

What NextEmail.ai reads

  • Email metadata (sender, recipient, timestamp, subject)
  • Email body & attachments — in-memory classification, never persisted
  • Folder hierarchy and label structure
  • Authentication headers (SPF, DKIM, DMARC)

What NextEmail.ai never does

  • Send email content outside the customer perimeter
  • Require admin elevation beyond standard OAuth
  • Delete email — folder moves and label adds only
  • Access drafts, contacts, or calendar
  • Train on customer data or emit content telemetry
Compliance

Architecturally defensible, not checkbox-defensible.

HIPAA PHI in email never crosses the BAA boundary. Classification runs inside the covered entity.
SOC 2 Type II No third-party data processor in the inference path. Customer retains the full audit trail.
SOX Material communications stay inside the issuer's information barrier. No external transit.
GDPR Art. 28/44 No cross-border data transfer for inference. Data subject content stays in-jurisdiction.

Curious how this would land at your scale?

15 minutes. I'll walk the architecture and answer the hard questions.

Let's talk →

Phil Chiu · Co-Founder · NextEmail.ai