When a large company buys an AI tool, procurement and counsel run diligence before a contract is drafted: what data the tool touches, what the model is built on, who the subcontractors are, what happens when versions change. Regulators are moving the same direction; Colorado's new automated-decision law requires developers of covered AI to document intended use, limitations, and monitoring. And as William Galkin, an attorney focused on AI agreements, puts it: with AI tools, "the most important work often begins before the first draft is exchanged."
A solo agent doesn't have a procurement department. But when the AI in question answers your phone, in your name, to your clients, you are the procurement department. This guide compresses that diligence into six questions you can ask any AI voice vendor in one email, plus what a good answer looks like.
The six questions
1. Does it declare itself as an AI?
Not every state requires this yet, but the direction is set: Utah's AI Policy Act (SB 226, 2025) already requires disclosing generative AI when a person clearly asks, and requires regulated-occupation providers to disclose it proactively in higher-risk interactions. Beyond the law, there is the trust math: callers are primed for phone suspicion for good reason. The FBI's Internet Crime Complaint Center logged $275 million in real-estate-related fraud losses in 2025 (fraud generally, not AI voice specifically), and voice phishing accounted for over 60% of the phishing-related incident-response engagements Cisco Talos worked in Q1 2025. An AI that opens by saying what it is removes the question. A vendor whose assistant cannot be configured to disclose has made a choice about whose risk that is.
2. What data does it take, and what happens to it?
Be specific about what the vendor receives. For a voice assistant that means: are calls recorded? Transcribed? Retained how long? Used to train models? Treated as the vendor's asset or yours? Policies vary widely, and some say nothing at all about recording or retention. Silence in a privacy policy is an answer.
The industry itself is drawing this line now, and it has a name. Bright MLS, one of the country's largest listing services, announced rules taking effect in summer 2026 that "explicitly prohibit uploading Bright data into any AI tool that lacks a zero-training guarantee," free ChatGPT tiers included. Notice who carries that rule: the member, meaning the agent. If your MLS adopts the same pattern, and the large MLSs are moving together on this, you will need your AI vendor's zero-training guarantee in writing. Ask for it now; a vendor who already operates that way will hand it over without a meeting.
3. Who reviews what the AI says for fair-housing compliance?
The question that is unique to real estate, and the one most vendors hope you skip. Read any vendor's AI-disclosure page with one question in mind: who does it say is responsible for reviewing the AI's output? Ask directly, get the answer in writing, and read the full treatment in our companion guide, The Realtor's Guide to AI and Fair Housing Compliance.
4. Does it use your cloned voice or a distinct assistant voice?
Cloning your voice and likeness is now pitched to real estate as a routine feature. Avatar-video platforms court agents on conference stages with tools that turn a short recording of you into endless content, and voice platforms market cloned-voice phone agents directly to realtors. The word "clone" rarely appears on the pricing page; the capability is the product. Two separate risks follow. Legally: the FCC ruled in February 2024 that AI-generated voices fall under the TCPA's artificial-voice rules for calls the AI initiates, so if your assistant ever places outbound calls, such as callbacks or reminders, consent and identification requirements can apply, and a voice presenting as you personally raises the stakes of getting that wrong. Practically: a caller who later learns "you" were software feels deceived in a way no fine print repairs. A distinct, disclosed persona avoids the impersonation question entirely. It is a design answer, not a legal cure-all.
5. Who else is in the pipeline?
Voice AI runs on a supply chain: a telephony carrier, a model provider, hosting, sometimes annotation or QA vendors. Third-party and supply-chain risk is a standing concern of security diligence, called out in NIST's AI Risk Management Framework, and every link in the chain is a place your callers' words may be processed or retained, depending on architecture and contract. A trustworthy vendor will name the chain and say what each party can see. "We can't share that" from a company handling your client conversations is a finding, not an inconvenience.
6. Does it screen its own generated copy for fair-housing language?
If the tool writes anything in your name, listing descriptions, ads, or follow-up texts, either it screens its own output for fair-housing language or that screening lands on your desk at 11pm. Agents in real-estate tech communities have flagged AI-drafted listing copy slipping in language a compliance review would catch. Ask where the screen lives.
Turning answers into terms
Diligence only matters if it changes the deal; that is the heart of Galkin's pre-contract framework, and it applies at every scale. For an agent, the vendor's written answers are your evidence, so keep the thread: no recordings, no training, disclosed AI, a named subcontractor chain. Then get the ones that matter into the actual paperwork, whether that is the order form, the terms, or a data-processing addendum, because a contract's fine print can override a helpful email. A vendor who answers cheerfully by email but won't reflect it in the document is telling you which one they intend to honor. And a refusal to answer in writing at all is a risk signal you are allowed to act on.
The regulatory direction backs you up. Colorado's SB 26-189 (effective January 2027, explicitly covering residential real estate) requires developers of covered AI to document intended use, limitations, training-data categories, and monitoring, and to notify deployers of material updates. A vendor who cannot answer your six questions today will struggle to answer a regulator's tomorrow.
Which raises the version question: AI products change under the hood, and an assistant that handled fair-housing questions carefully in the demo may behave differently after a model update. Ask how the vendor tests updates against their compliance constraints before rollout, and whether you get notified when behavior materially changes.
How we answer our own questions
Rocalyn Edge publishes its answers: AI disclosure on every call, no recordings, no training on your conversations, a distinct configured persona rather than voice cloning, and fair-housing constraints in the product with compliance grading behind it. The details live on the trust page, and the live demo will disclose itself in the first breath. Hold us to the same six questions; hold everyone to them.