Security, Privacy & Data Handling
Applied AI Partners builds software for insurance agencies, which means we handle client records, policy documents, recorded calls, and connected mailboxes. This page documents exactly how that data is protected, who can reach it, and what we will never do with it — in enough detail for a carrier vendor-security review.
Last updated: October 25, 2026 • Questions: info@appliedaijax.com
Talk with our team ↗Infrastructure & Hosting
Every Applied AI product runs on managed, independently audited U.S. infrastructure. We do not operate our own data centers or physical servers.
Where data lives
Application code runs on Vercel. Customer records and credentials live in a Supabase-managed PostgreSQL database. Documents and call or meeting audio are stored in Amazon S3. All three run in United States regions, and customer data is not replicated outside the U.S.
Encryption
All traffic is served over HTTPS with TLS 1.2 or higher; there is no unencrypted path into the application. Data at rest is encrypted with AES-256 at the storage layer. Stored audio uses server-side KMS-managed encryption.
Credential encryption
Mailbox and calendar refresh tokens get a second, application-level layer beyond disk encryption: AES-256-GCM authenticated encryption under a key we hold outside the database. A stolen database copy alone does not yield a usable mailbox credential, and any tampering is detected on decrypt.
Passwords
Authentication runs on Supabase Auth using the PKCE flow. Passwords are salted and hashed by the auth provider — Applied AI never stores, sees, or logs a user password, and staff cannot retrieve one.
Access Control & Tenant Isolation
One agency's data must never be reachable by another. We enforce that in the database, not just in application code.
Row-level security on every tenant table
Each tenant-scoped table enforces PostgreSQL row-level security policies. An authenticated query physically cannot return rows belonging to another agency, even if application code contained a bug. Tables default to deny; access is granted by explicit policy only.
Two-factor authentication
TOTP-based 2FA with any standard authenticator app, built on Supabase's native MFA. As of August 2026 it is mandatory for every user — there is no agency-level or user-level setting that turns it off. Members who have not enrolled are gated at sign-in until they do, and the check is enforced server-side, before any request reaches agency data. It fails closed: if the assurance level cannot be read, the request is refused rather than allowed.
Role-based permissions
Members hold explicit roles within their agency. Administrative endpoints re-verify the caller's role and active status on every request and on every HTTP method independently — an authenticated session is not by itself authorization.
Immediate revocation
When an agency removes a member, that member's access is cut at the next request across the web app, the mobile app, and the API. Revoked members cannot read agency data through any surface, including previously issued sessions.
Microsoft 365 & Outlook Access
Producer IQ connects to Microsoft 365 through the Microsoft identity platform using OAuth 2.0. Agents sign in to Microsoft directly — Producer IQ never sees, receives, or stores a Microsoft password.
Delegated permissions only
Producer IQ uses delegated Microsoft permissions, bounded by the connecting user's grant and mailbox permissions. Optional shared-mailbox access also requires the relevant delegated grant and the agency's enabled sharing controls. Producer IQ requests no application-level or tenant-wide permissions such as Mail.Read.All, Directory.Read.All, Files.Read.All, or Sites.Read.All. Agency roles determine which authorized teammates can view connected or shared mailboxes in the product; the connection screen explains that access.
Exact permissions requested
| Microsoft Graph scope | Why Producer IQ needs it |
|---|---|
offline_access | Keeps the connection alive so agents don't re-authenticate every hour. Issues the refresh token, which we store encrypted. |
openid / email | Identifies which mailbox was connected. Reads the address only — no directory or org-wide lookup. |
Mail.ReadWrite | Syncs inbound and sent mail onto the matching client's timeline, and powers the in-app inbox (read, flag, move, archive). |
Mail.Send | Sends email from the agent's own address, so clients see the agent — not a no-reply relay. |
User.Read | Reads the signed-in user's own basic profile to confirm the connected identity. |
Calendars.ReadWrite | Two-way calendar sync between Outlook and the Producer IQ team calendar. |
Mail.ReadWrite.Shared / Mail.Send.Shared | Optional delegated permissions for shared mailbox features when the agency enables them and the connecting user has the required Microsoft mailbox permissions. |
What we do with mailbox data
Synced mail can be matched to client records and displayed within the agency's workspace. Enabled features can send or organize messages, synchronize calendar events, and use the authorized records needed to prepare a requested AI draft. Agency roles and sharing controls determine access. Providers may process the minimum information needed to deliver the enabled feature, subject to applicable data-use restrictions, including Google's Limited Use requirements for Google API data.
What we never do
- Never train AI models on mail content
- Never sell, rent, or broker mail data
- Never disclose mail for another company's advertising or marketing
- Never grant unrestricted, application-level access to a Microsoft tenant
- Never authorize access beyond the provider grant and agency sharing permissions
AI & Your Data
AI is the core of what our products do, so the data question deserves a direct answer rather than a footnote.
No training on customer data — ever
We do not use customer content to train AI models or authorize providers to do so. Our approval requirements include commercial terms and account settings that prohibit training on the content and derived outputs, including for AssemblyAI and ElevenLabs where used. Accounts must be excluded from model-improvement or data-sharing programs that permit that use. We do not enroll customer content in training-permitting feedback or evaluation programs. Limited operational retention under the applicable provider terms is separate from model training.
Least-context by default
AI features send only the specific records needed to answer the request — not the agency's book, and not a bulk export. A summarization request sends that call. A client brief sends that client's records.
AI inherits your permissions
The in-app assistant runs its lookups server-side, scoped to the requesting user's agency. It cannot surface a record the user could not already open directly, and it cannot cross an agency boundary.
A human approves outbound AI
AI-drafted client communication — including renewal outreach — lands in a review queue. Nothing generated by AI sends to a client without a person approving it. There is no autonomous send path.
Data Sharing & Subprocessors
Applied AI does not sell, rent, or trade customer data or disclose client content for third-party marketing. The list includes providers selected by Applied AI and systems clients connect under their own contracts. Client-connected systems such as Microsoft, Google, PandaDoc, and QuickBooks receive only the information authorized for the enabled integration. Applicable law or an agreement may also require a limited disclosure. Contact us for the list and processing restrictions applicable to your service or custom engagement.
| Provider | Role | Data it touches | Region |
|---|---|---|---|
| Vercel | Application hosting and edge network | Application traffic and operational metadata needed to host and protect the service | United States |
| Supabase | PostgreSQL database, authentication, and feature-specific file storage | Customer records, encrypted credentials, and files stored by the enabled feature | United States |
| Amazon Web Services | Object storage for documents and audio; optional Polly text-to-speech | Uploaded documents, call and meeting audio, and agency-authored greeting text for enabled speech synthesis | United States |
| Twilio | Voice and SMS carrier services | Call audio, phone numbers, message bodies | United States |
| AssemblyAI | Speech-to-text transcription | Call and meeting audio, processed and returned | United States |
| Recall.ai | Debrief IQ notetaker for Google Meet calls | Meeting link, call audio, and participant names from Google Meet calls the notetaker joins; Recall.ai's copy of the recording is deleted once Debrief IQ saves it | United States |
| Anthropic | AI summarization, drafting, and search | Only the record text needed for the request | United States |
| OpenAI | Embeddings for document and content search | Policy document text and search queries — only the text needed to build and query the search index | United States |
| Resend | Transactional email delivery | Recipient address and message body | United States |
| Stripe | Subscription billing | Billing contact and payment method; we never see full card numbers | United States |
| ElevenLabs | Text-to-speech for voicemail greetings | Agency-authored greeting text — no policyholder data | United States |
| Expo / Apple / Google push delivery | Native-app notifications when enabled | Device push identifiers and the notification payload needed for delivery | See the applicable delivery provider's terms |
| PandaDoc | E-signature for the finalized-policy workflow — only for agencies that connect their own PandaDoc account | Documents an agency explicitly sends for signature, plus the record fields merged into them, delivered into that agency's own PandaDoc workspace under its own PandaDoc contract | United States |
| Microsoft / Google | Mail and calendar sync | Only if an agency connects an account, and only under the scopes it grants | United States |
| Intuit QuickBooks Online | Client-connected accounting integration, where enabled | Approved expense, receipt, reimbursement, and payout records the client directs into its own QuickBooks company | Client's provider terms apply |
Retention, Deletion & Recordkeeping
Insurance has longer recordkeeping obligations than typical SaaS. Our defaults are built for E&O exposure, not for storage savings.
Call recordings
Recordings, transcripts, summaries, and call records can have different retention schedules. The applicable service settings, insurance recordkeeping requirements, signed agreements, and valid legal holds determine the period. Contact us for the current schedule and deletion process for your records; a subscription ending does not necessarily erase them immediately.
Legal hold
When a dispute, complaint, or regulatory inquiry arises, an agency can place a one-click legal hold that exempts the affected records from scheduled deletion. Every disposal event is written to a tamper-evident log.
Customer-controlled deletion
Available export, deletion, and preservation controls vary by record type. Hiding or archiving an item may retain its underlying record and recording. Request return or deletion at info@appliedaijax.com. We verify authority, apply any signed agreement and legal requirements, and explain exceptions and timing. Backups, delivered copies, and separately retained device files have their own retention or deletion processes.
Sensitive-data handling
Uploaded documents are stored under UUID-only filenames that carry no client information. Log output is sanitized so personal data, policy numbers, and credentials are not written to logs. Consent and quiet-hours rules are enforced on outbound calling and messaging.
Compliance Posture
Where we stand today, stated plainly — including what we have not yet completed.
Independently audited infrastructure
The providers that host and transport customer data — Vercel, Supabase, Amazon Web Services, Twilio, and Stripe — each maintain SOC 2 Type II certification, and their current reports are available directly from them.
Applied AI's own SOC 2
Applied AI Partners has not yet completed a SOC 2 Type II audit of its own; it is on our roadmap. We say so directly rather than implying otherwise, and we will complete a carrier's own vendor security questionnaire on request in the meantime.
HIPAA-defensive design
Health-adjacent lines mean protected information can appear in agency records. We apply HIPAA-defensive controls throughout — PII redaction before logging, UUID-only filenames, encryption at rest and in transit, and audit logging of sensitive actions. Where a covered engagement requires a Business Associate Agreement, we execute one as part of contracting.
Regulatory alignment
Outbound calling and messaging enforce TCPA quiet hours and consent state. Call recording follows the consent rules of the applicable jurisdiction. SMS runs under a registered 10DLC brand rather than unregistered traffic.
Resilience & Incident Response
Backups and recovery
The production database runs managed automated backups through our database provider. Object storage is redundantly stored across multiple availability zones by the storage provider. Hosting runs on a globally distributed edge network with automatic failover.
Secure development
Every release runs an automated security suite that verifies an authorization check exists on every API route and every HTTP method independently, that row-level security covers every database table, and that server-side administrative queries filter by agency. Secrets live in environment configuration and are never committed to source control or exposed to the browser.
If something goes wrong
On confirming an incident affecting customer data, we contain it, determine scope, and notify affected agencies directly with what we know, what was affected, and what we are doing — without waiting for a full post-mortem to make first contact.
Reporting a vulnerability
Security researchers and agency IT teams can report a suspected vulnerability to info@appliedaijax.com. We investigate every report and will not pursue action against good-faith research.
Security Questions
Where is Applied AI customer data stored?
All customer data is stored in the United States. Our database (Supabase/PostgreSQL), object storage (AWS S3), and application hosting (Vercel) all run in U.S. regions. We do not replicate customer data outside the United States.
Does Applied AI use customer data to train AI models?
No. We do not use customer content to train AI models or authorize our providers to do so. Approval of an AI, transcription, or voice provider requires applicable commercial terms and account settings that prohibit training on the content and derived outputs. This requirement includes AssemblyAI and ElevenLabs where used. Accounts must be excluded from model-improvement or data-sharing programs that permit that use. We do not opt customer content into feedback or evaluation programs that permit training. Provider processing may include limited retention for security or abuse monitoring under the applicable terms; no training does not mean zero retention. Contact us for the provider information applicable to your service.
What Microsoft 365 permissions does Producer IQ request?
Producer IQ requests delegated permissions: offline_access, openid, email, Mail.ReadWrite, Mail.Send, User.Read, and Calendars.ReadWrite. Optional shared-mailbox features can also request Mail.ReadWrite.Shared and Mail.Send.Shared. Access is bounded by the connecting user's Microsoft grant, mailbox permissions, and the application's agency sharing controls. We do not request application-level or tenant-wide permissions such as Mail.Read.All, Directory.Read.All, Files.Read.All, or Sites.Read.All. We do not request MailboxSettings.ReadWrite.
Can an agency or carrier revoke Producer IQ's mailbox access?
Yes, at any time and from either side. An agent can disconnect the mailbox inside Producer IQ, which deletes the stored credential. A Microsoft 365 administrator can revoke the grant tenant-wide from the Microsoft Entra admin center, and an individual user can revoke it at account.microsoft.com. Revocation is immediate and requires no action from Applied AI.
Does Applied AI sell or share customer data?
We do not sell, rent, or trade customer data or disclose client content for another company's marketing. Limited disclosures support the service through its providers, recipients and connected systems you authorize, or legal obligations. Providers selected by Applied AI and systems connected under a client's own contract have different roles; both are described in the provider information below.
How is data isolated between agencies?
Every tenant-scoped table in Producer IQ enforces PostgreSQL row-level security, so an authenticated request can only ever return rows belonging to the requesting user's agency. Isolation is enforced by the database itself, not by application code that could be bypassed. Server-side administrative queries are additionally filtered by agency in code and reviewed against that requirement.
Does Applied AI require two-factor authentication?
Yes — as of August 2026 it is mandatory, not optional. Every user enrols a TOTP authenticator at sign-in, and there is no user-level or agency-level opt-out to turn it off. The check runs server-side before any request reaches agency data, and it fails closed. The mandatory requirement is live today on every web app — Fine Print IQ, Debrief IQ, Producer IQ, and the client portal; the native mobile apps complete the rollout as in-app enrolment ships. The Producer IQ app already asks for the authenticator code before it opens a connected mailbox, and the server refuses mailbox access from the app without it.
Is Applied AI SOC 2 certified?
Applied AI Partners has not yet completed a SOC 2 Type II audit; it is on our roadmap. Our core infrastructure providers — Vercel, Supabase, AWS, Twilio, and Stripe — are SOC 2 Type II certified today, and we are glad to complete a carrier's own vendor security questionnaire in the meantime.
Documentation for Your Review
These documents exist today and are provided on request — most within one business day. We would rather hand a reviewer the evidence than ask them to take our word for it.
Completed security questionnaire
A pre-filled CAIQ Lite, structured on the Cloud Security Alliance's Cloud Controls Matrix v4, covering audit and assurance, governance, identity and access, data security, application security, cryptography, logging, incident response, continuity, supply chain, and AI-specific controls. It states our gaps as plainly as our controls, including the absence of a SOC 2. If your review uses its own questionnaire instead, we will complete that one.
Data-flow documentation
A written map of every trust boundary — what data crosses each one, and where it comes to rest rather than merely passing through. It identifies the specific paths that can carry health-adjacent information and the controls applied to each.
Information security policy set
Twelve documents covering information security, access control, data classification, cryptography and key management, secure development and change management, incident response, business continuity, vendor management, logging and monitoring, personnel security, and risk assessment — including a maintained risk register with treatment decisions and owners.
Data Processing Agreement
A data processing agreement is available for execution as part of contracting. It is written for United States insurance distribution — anchored to state Insurance Data Security Act obligations for third-party service providers and to CCPA service-provider terms — and names our subprocessors, breach-notification timeline, and deletion commitments. Where a covered engagement requires it, we execute a Business Associate Agreement alongside.
Executed agreements with our own AI vendors
We hold executed data processing agreements with the AI providers in our data path, not merely a reliance on their published policies. Our organization-level settings with those providers are configured so that customer content is not shared for model training, and we verify those settings rather than assume them. We will identify the operative documents on request.
Internal audit findings
We run recurring adversarial security audits across every repository, with written findings tracked to closure rather than noted and forgotten. Summaries are available under NDA. A finding is not closed until the capability has been re-tested — not when a console reports success.
Reviewing Applied AI as a Vendor?
We work with carrier compliance and agency IT teams regularly, and we will complete your vendor security questionnaire, walk your team through the Microsoft 365 permission model, or provide additional documentation on request. Ask for whatever your review requires. Farm Bureau Insurance Company of Idaho completed its IT and security review in September 2026 and added us to its list of approved AI companies.
Applied AI Partners LLC • Jacksonville, Florida • (904) 830-9969