Reliability & Guardrails · 9 min read
Privacy Policy
The Omni model: one managed AI Employee owns one recurring workflow; specialist employees add capacity around the same business context; your team keeps the judgment calls.
Last quarter, a regional dental practice we work with in Ohio deployed a 24/7 booking agent on their website. Two weeks in, a patient emailed asking where her conversation transcript was stored, whether it was being used to train the underlying model, and who at the practice could read it. The practice owner forwarded the email to us with one line: "I have no idea how to answer this."
This is the moment most service businesses hit when they add an AI agent. The technology works. The handoffs are clean. But the privacy policy on the website still reads like it was written in 2019, because it was. And now there's a new category of customer interaction—automated, transcript-generating, often voice-based—that the document doesn't cover.
This article walks through what actually needs to change in your privacy policy when you deploy an AI agent, and how to do it without a six-week legal review cycle.
What Changes When You Add an AI Agent
A privacy policy written for a standard service business assumes three things: a human collects information, that human stores it in a CRM or practice management system, and access is limited to employees with login credentials. An AI agent breaks all three assumptions at once.
First, the AI agent collects information without a human present. The customer types or speaks into an interface, and the system captures structured data (name, phone, service needed) plus unstructured data (the actual conversation, including incidental details like "I'm calling about my mother's kitchen remodel").
Second, the data flows through a processing layer your vendor controls. If you're using a hosted LLM or a transcription service, that vendor's servers see the content. Your CRM does too. Now there are two or three places customer data lives instead of one.
Third, conversation transcripts are a new data category entirely. They're not a lead form submission. They're not an email. They're a free-form record of what a customer said, which may include health information, financial details, family circumstances, or other sensitive context the customer didn't intend to disclose.
The NIST AI Risk Management Framework (NIST.AI.100-1) flags this category explicitly under the "Manage" function: organizations deploying AI systems need clear policies around data provenance, retention, and access. It's not enough to extend your existing CRM policy. The interaction model has changed.
The Sections Your Privacy Policy Needs Now
When we run a privacy review for a client before deployment, we walk through five sections that need to be either added or rewritten. None of this is theoretical—these are the questions customers and their lawyers actually ask.
- AI-specific disclosure language. Your policy needs a paragraph explaining that automated systems interact with customers, what those systems do (answer questions, collect information, route to staff), and that customers can request a human instead. The Federal Trade Commission's 2023 guidance on AI and consumer protection recommends explicit disclosure of automated decision-making, and several states have moved to codify this.
- Data categories for AI interactions. List what the AI collects beyond your existing CRM fields. This usually includes voice recordings (if you use voice agents), transcripts, timestamps, conversation duration, and any metadata the platform captures (browser, device, referring page). Don't bury this. Customers have a right to know.
- Vendor and sub-processor list. Name the AI vendors, transcription providers, and any analytics tools that touch customer conversations. If your AI provider uses customer conversations to train shared models, that needs to be in the policy—and you may need an opt-out or a zero-retention configuration. The IAPP's 2024 privacy and AI governance report found that vendor transparency was the single most common compliance gap in mid-market AI deployments.
- Retention rules for transcripts. Decide how long conversation logs are stored. Most regulators and platform providers now offer 30-day or zero-retention configurations. Pick one, document it, and configure your system to enforce it automatically.
- Human review and access controls. Specify who at your business can read transcripts, under what circumstances, and how access is logged. If your AI provider reviews conversations for quality purposes, that needs to be disclosed too.
This is also where state-level rules start to bite. California's CCPA, Virginia's VCDPA, and similar laws in Colorado, Connecticut, and others all grant consumers rights around automated decision-making. If your AI agent makes any determination that affects the customer—scheduling priority, qualification for a discount, routing to a senior technician—that's a decision, and the customer may have the right to opt out of it.
A Workflow Example: Home Services Company Adding a Voice Agent
Here's how this plays out in practice. We recently helped a regional HVAC company deploy a voice agent for after-hours calls. The agent answers, gathers the customer's issue and address, checks technician availability, and either books the appointment or takes a message for morning follow-up.
The workflow looked like this:
- Step 1: Disclosure language on the call. The agent opens with: "Hi, this is the automated assistant for [Company]. This call may be recorded for quality and training purposes. To speak with our team directly, press 0 at any time." This is configured in the voice agent's first prompt.
- Step 2: Updated website privacy policy. The company's existing policy was rewritten to add a new section titled "Automated Calls and Voice Interactions." It covers what's recorded, how long it's kept (30 days), who can listen (the owner's account manager and the AI provider's quality team), and how customers can request deletion.
- Step 3: Vendor configuration. We configured the AI provider with zero data retention for transcripts and disabled training-on-customer-data at the account level. This took about ten minutes and was documented in the vendor's admin console.
- Step 4: Internal access controls. Only two employees at the HVAC company have transcript access—the owner and the operations manager. The AI provider's quality team has access for flagged conversations only, and the customer is notified if a transcript is reviewed.
- Step 5: Quarterly review. Every 90 days, we pull a sample of transcripts, review the policy alongside them, and flag anything new (a customer mentions a medical condition, a billing dispute comes up, etc.). If the policy needs another update, we make it before the next quarter's calls go out.
This isn't a legal opinion and it isn't a substitute for one. But it's the operational baseline we use for service business clients in regulated-adjacent industries. A McKinsey survey on AI adoption in late 2024 found that companies with documented governance workflows were roughly twice as likely to scale AI successfully as those without—largely because the governance work surfaced edge cases early.
Where Human Review Belongs in the Privacy Workflow
One pattern we see repeatedly: businesses set up the AI agent, update the privacy policy once, and then never look at it again. That's where problems compound. A transcript from January might reference a medical condition. A customer in March might ask for their data to be deleted. A new state law might take effect in July that changes disclosure requirements.
The fix is a human review point built into the workflow, not bolted on afterward. For most service businesses we work with, that means three checkpoints:
Pre-deployment review: Before the AI agent goes live, someone at the business (not just at the vendor) reads the privacy policy end-to-end and confirms it matches what the system actually does. This takes 20 minutes and catches most gaps.
Quarterly transcript sampling: Pull 10-20 random transcripts, read them, and ask: did the customer say something the policy doesn't cover? Did the agent collect data the policy doesn't mention? This is the review point that catches drift.
Annual full policy review: Once a year, treat the privacy policy like a contract renewal. Read it. Update vendor names. Adjust retention windows. Check state law changes. The Harvard Business Review has written about this kind of operational discipline in AI deployments—it's not glamorous, but it's what separates organizations that scale AI responsibly from those that get caught flat-footed.
None of this requires a full-time privacy officer. It requires a calendar reminder and 30 minutes a quarter.
Common Gaps We See in AI-Ready Privacy Policies
After running about forty of these reviews in the past year, we have a reliable list of what gets missed. If you're updating your policy yourself, check against these.
No mention of training data. The policy talks about CRM data but never addresses whether customer conversations are used to train the AI model. If your vendor offers a zero-retention configuration, turn it on and say so in the policy.
Voice and text treated as the same. They aren't. Voice recordings often fall under additional state consent rules (two-party consent states like California, Florida, Illinois, and others). If you deploy a voice agent, you may need to disclose recording at the start of the call, not just in the privacy policy.
No deletion request workflow. CCPA, VCDPA, and similar laws give customers the right to request deletion of their data. Most AI platforms support this, but you need to test the workflow. If a customer emails asking for their transcript to be deleted, can your team actually do it within the legal timeframe (usually 30-45 days)?
Confusing AI with the website chatbot you retired in 2021. If your policy still references a "live chat feature" but you've replaced it with an AI agent that does much more, the policy is misleading. Update the language.
Missing escalation path. The policy should tell customers how to reach a human with a privacy concern—not just a generic contact form. An email address monitored by a real person (the owner, the ops manager) is enough.
Frequently Asked Questions
Do I need a separate privacy policy for my AI agent?
Not necessarily. Most service businesses can fold AI-specific disclosures into their existing privacy policy as a new section. A separate policy only makes sense if your AI agent operates under a different brand or handles a distinct category of sensitive data (medical, legal, financial).
What about recording phone calls for AI processing?
This depends on your state. If you're in a one-party consent state, you can record as long as one party (your business) consents. If you're in a two-party consent state, you need to inform the customer and get acknowledgment at the start of the call. The AI agent's opening prompt should handle this. Check the International Association of Privacy Professionals (IAPP) state-by-state guide for current rules.
How long should I keep AI conversation transcripts?
Most service businesses we work with set retention at 30 days, then auto-delete. That's enough time to catch a quality issue or respond to a customer request, without holding onto data longer than necessary. If your industry has specific retention rules (healthcare, financial services), follow those instead.
Does the AI vendor see my customer data?
It depends on the configuration. Many enterprise-tier AI providers offer zero data retention, meaning transcripts pass through their servers but aren't stored or used for training. Lower-cost tiers often retain data by default. Check your vendor's data processing addendum and configure retention settings before going live.
What happens if my privacy policy isn't updated?
Nothing dramatic in the short term. But if a customer files a complaint with a state attorney general or the FTC, an outdated policy is treated as a misrepresentation—which carries more risk than having no policy at all. It's also a trust issue. Customers are increasingly aware of how their data is handled, and a policy that doesn't mention AI looks negligent.
Closing Thought
Updating a privacy policy for an AI agent isn't a legal project. It's an operational one. The work is mostly in the configuration—turning off training data retention, setting transcript deletion windows, defining who on your team can read conversations. The policy itself just has to describe what you've already built. If you do the implementation first and the legal language second, the document writes itself.
If your service business is deploying or scaling AI agents and the privacy policy hasn't kept up, that's a normal place to be. It's also a place that's fixable in an afternoon with the right checklist. We run these reviews as part of our deployment work and we're happy to walk through yours.
Book a free AI automation audit and we'll review your current privacy setup alongside your AI workflow, flag the gaps, and leave you with a concrete list of what to update before your next deployment.
