Security & Privacy for Agent AI: How Client Data Is Protected

When you give an AI agent access to your CRM, your email, and your transaction files, you're trusting it with the most sensitive data your business holds. Here's how we handle it.

The threat model

What the agent sees on a typical day:

• Names, emails, phone numbers for every contact in your CRM
• Lead-source attribution (which website they came from, what they searched for)
• Behavioral data (which listings they viewed, which emails they opened, which texts they replied to)
• Transaction data (offer details, contract terms, closing dates)
• Personal notes you've added to contacts ("prefers email, available after 6pm, kids in elementary school")
• Communications (drafts and sends) — emails to clients, SMS, follow-up messages

A data breach is a serious incident. A solo agent who loses client trust from a leak doesn't recover. We treat that as the baseline.

What we DON'T do with your data

Three things we will not do, full stop:

No training on your data. Our model training pipeline is firewalled from customer data. Your contacts, your CRM activity, your transaction notes — none of it is used to train any AI model. This is contractually required with our model providers (we use frontier models from OpenAI, Anthropic, and Google — each has a zero-retention policy for our account, meaning your inputs are not retained beyond the request and never used for training).

No selling or sharing. Your data is not sold to data brokers, not shared with marketing partners, not aggregated with other customers' data. The agent is a tool you use, not a product built from your data.

No retention beyond what you need. Cached contact data is deleted within 60 days of last access. CRM credentials are revoked within 30 days of cancellation. Audit logs are retained for 1 year (for legal/compliance purposes), then purged.

Encryption

All data is encrypted in transit and at rest.

In transit. Every connection from your browser, from the agent, and from your CRM uses TLS 1.3. There are no plaintext connections anywhere in the pipeline. We use HSTS to force TLS on every request.

At rest. Customer data is encrypted at rest using AES-256-GCM. Database backups are also encrypted (AES-256-GCM with a separate key from the primary database). Encryption keys are managed in AWS KMS with annual rotation.

At the model layer. When the agent sends a request to the model provider, the request is encrypted in transit (TLS). The provider's contract specifies that requests are not logged, not retained, and not used for training. The provider's data deletion happens within 30 days of the request, even for ephemeral logs.

Access controls

Who can see your data:

• You — full access to everything the agent sees, sees, or writes on your behalf.
• Your authorized users — if you have team members (TC, ISA, partner agent), they get scoped access via your dashboard.
• Our engineering team — access is logged, requires an explicit ticket, and is granted for the minimum time needed to resolve the issue. We audit every access event quarterly.

We don't have:

• Standing access to customer data
• Shared service accounts
• Off-the-shelf admin panels with broad permissions

Access is role-based: a customer's primary agent gets full access; an engineer helping with a support ticket gets scoped, time-limited access.

BAA availability

For brokerages that need a Business Associate Agreement (some states require this for real estate workflows where health-adjacent information is involved — accessibility requests, age-restricted housing queries, etc.), we sign BAAs.

The BAA covers:

• Data we process on your behalf
• Our subcontractors (model providers, hosting)
• Breach notification timelines (we notify within 72 hours of confirmed breach)
• Your right to audit our security controls annually

To request a BAA, contact us before the deploy. Standard turnaround is 3 business days.

Data residency

All customer data is stored in US-East (AWS us-east-1, AWS us-east-2). Model providers process requests in their US data centers (OpenAI, Anthropic, and Google all have US-only processing available for our account).

If you're a Canadian agent or an EU-resident agent working with US clients, let us know — we can configure data residency for your account to keep everything in a specific region.

Compliance frameworks

We are not a HIPAA-covered entity (we're not a healthcare provider), but we follow HIPAA-grade practices because the data sensitivity is similar:

• SOC 2 Type II — audited annually
• Encryption standards — meet FIPS 140-2 requirements
• Access logging — every access event is logged with user, time, action, resource
• Backup and disaster recovery — daily encrypted backups, 30-day retention, RPO 24 hours, RTO 4 hours
• Incident response plan — tested quarterly, with a 72-hour breach notification SLA

We don't have a SOC 2 audit report published publicly (it's available on request for enterprise customers). If your brokerage requires one for vendor approval, contact us.

What you control

You have visibility into and control over what the agent sees:

• Per-workflow data scopes. If you only want the agent to read contacts from one tag in your CRM, you can scope it that way. The agent doesn't see your full CRM unless you give it access.
• Per-touch approval. Every client-facing message the agent drafts can require your approval before sending (default for high-touch workflows; auto-send only for routine confirmations).
• Data export. You can request a full export of all data the agent has stored at any time. JSON, CSV, or your CRM's native import format.
• Deletion on demand. You can delete specific contacts, specific touches, or your entire account at any time. Deletion completes within 30 days.
• Audit log access. You can see every action the agent took on your behalf, with timestamps and reasoning. Exportable.

The honest limits

Three things we can't promise:

1. Model providers have their own security practices. We trust OpenAI, Anthropic, and Google to handle requests per their contracts with us. If one of them has a breach, we're impacted. We mitigate by using providers with strong SOC 2 practices, but we don't control their infrastructure.

2. Your CRM's security is your responsibility. If your CRM password is weak or shared, the agent's integration can be compromised regardless of our security. We require OAuth (not password-sharing) and recommend SSO where available.

3. No system is 100% breach-proof. We do our best to be defensible against every realistic threat. We can't promise perfect security. What we can promise is: rapid detection, rapid response, and full transparency if something goes wrong.

What's next

If security is a blocker for your brokerage decision, book a 30-minute consult and we'll walk through the controls + provide the SOC 2 audit summary on request.

If you want to read more, the CRM integrations post covers the OAuth scopes and data flows in more detail.