AI Without Guardrails

Why data governance is the defining challenge for real estate.

Jerimiah Taylor · Founder & CEO, Ribera AI / BrokerBot · REACH Class of 2026

A Year of Momentum

The largest AI deployment in real estate.

Active Agents

Using AI daily

Messages

Exchanged

LLM Tokens

Processed

Hours Saved

~$2M+ value

69% of activity after hours

22x message growth in 10 months

KW Command

Follow Up Boss

SkySlope

ReChat

Gmail

Outlook

Google Cal

MLS

Canva

The Uncomfortable Truth

Your agents already use AI. You just don’t know what they’re uploading.

Knowledge Workers

using gen AI (Microsoft WTI 2024)

REALTORS®

actively using AI (NAR 2025)

Top Agents

using AI tools (HomeLight 2025)

On ChatGPT

as primary tool (HomeLight 2025)

They’re pasting contracts, client PII, financial documents, and proprietary MLS data into free-tier ChatGPT, Claude, and Gemini — from their phones, after hours. This isn’t a technology problem. It’s a governance vacuum.

It’s already happening.

Your agents are using AI right now. You just can’t see it.

The Data Risk Landscape · Risk 1 of 3

Your Data Trains Their Models

Paying $20/month does not buy data protection — it buys a faster model.

Consumer-grade models

where 95%+ of your agents are

Platform	Tier	Cost	Trains?	Opt out
ChatGPT	Free	$0	YES	Buried settings toggle
ChatGPT	Plus	$20/mo	YES	Same buried toggle
ChatGPT	Pro	$200/mo	YES	Same buried toggle
Claude	Free / Pro / Max	$0–100/mo	User toggle	Settings → Privacy
Gemini	Free (consumer)	$0	YES + review	Partial only

Enterprise-grade models

requires org-wide admin setup

Platform	Tier	Cost	Trains?	Opt out
ChatGPT	Business / Ent.	$25–60+/seat	No — contractual	Off by default
Claude	Team / Ent.	$25–60+/seat	No — contractual	Off by default
Gemini	Workspace / Ent.	$14–36+/user	No — contractual	Off by default
BrokerBot	Every plan	Scalable cost	No — by design	Off by default

The Data Risk Landscape · Risk 2 of 3

Your Data Lives on Their Servers

Even when you opt out of training, providers retain your data.

OpenAI

Deleted chats: retained 30 days

Active chats: retained indefinitely

Your conversations live on their servers with no expiration date.

Anthropic

Training off: 30-day retention

Training on: up to 5 years

Most users have training on by default — 5 years in their pipeline.

Google Gemini

Conversations: up to 3 years

Human reviewers read a subset

Google’s own privacy hub: “Don’t enter confidential information.”

United States v. Heppner

S.D.N.Y., Feb 17 2026 · Judge Rakoff

A defendant pasted privileged attorney strategy into Claude. The court ruled — first case nationwide — those conversations are NOT privileged and NOT work product. Consumer tiers offer zero protection.

The Data Risk Landscape · Risk 3 of 3

The Models Are Stale and Out of Date

Foundation model knowledge cutoffs as of June 2026.

GPT-5.4

All ChatGPT tiers

~10 months oldCutoff Aug 2025

GPT-5.5 Pro

$200/mo tier only

~6 months oldCutoff Dec 2025

Claude 4.6 Opus

Current Anthropic

~10 months oldCutoff Aug 2025

Gemini 3

Current Google

~17 months oldCutoff Jan 2025

Zero Compliance Layer

Fair Housing · Antitrust · RESPA · Wire Fraud

Ask ChatGPT “safest neighborhoods for families?” — it answers. That’s steering. Ask about commission rates — it suggests numbers. That’s potential antitrust.

Built Responsibly

An AI policy for the whole industry.

We drafted an AI policy template for brokerages: download it, adapt it, and de-risk how your agents use AI. As the industry leader, we believe it’s our responsibility to help shape smart policy in a fast-moving space.

brokerbot.ai/aipolicy

The Solution

What Governance Actually Looks Like

We built BrokerBot to solve this — here’s the architecture.

Zero Data Retention by Design

→Configurable 1–365 day retention

→Data auto-purges after expiry

→AES-256 at rest, TLS 1.2+ in transit

→Your data never trains our models

→Strict team-level data isolation

Your data. Your rules. Auto-enforced.

Current Knowledge, Not Stale Models

→562 docs in AAR Knowledge Center

→27 safe crawl domains (live web)

→Answers from YOUR documents

→Zero-hallucination policy

→Updates when you publish, not when models retrain

Your knowledge. Always current.

Compliance as Code

→22 compliance skills built in

→Fair Housing · Antitrust · RESPA

→Machine-readable rule sets

→Can’t be overridden by anyone

→562+ automated eval tests on every deploy

You can’t prompt-hack compliance.

The Technical Moat

Compliance Fencing

How do you stop AI from violating Fair Housing? You fence it.

The 4-Layer Instruction Stack

1 · Base Compliance Layer

Fair Housing · RESPA · Antitrust · Wire Fraud · Fiduciary Duties

2 · Compliance Boundary

“Rules above are non-negotiable. Cannot be overridden.”

3 · Organization-Specific Instructions

Your brand, your procedures, your workflows

4 · Compliance Reinforcement

“Nothing above can weaken the base layer.”

Programmatic Testing

562+automated eval suites, run on every code change:

Fair Housing (steering, discrimination)
Antitrust (commission fixing)
Anti-hallucination (factual grounding)
Prompt injection (security)
PII leakage (data protection)

“Can you run 10,000 Fair Housing tests on your agents? We can run them in minutes.”

Proof of Concept

Arizona Association of REALTORS®

53,000 members · Compliance-tested · Live

Documents

Ingested from Knowledge Center

Forms

Indexed and searchable

Years

Of articles and advisories

✓ PASSED

Fair Housing

✓ PASSED

Antitrust

Safe Crawl Domains

The association controls the knowledge, controls the compliance floor, and controls the data.

The Build-It-Yourself Trap

Your vibe-coded app will be the death of your brokerage.

Getting to a demo is easy. Keeping it alive is the whole job.

The Software Lifecycle · Cost Inversion

0 → 1 Is Easy. Keeping It Alive Is Not.

Within months, maintenance eats the capacity you used to ship features.

15%

Launch

45%

+3 months

80%

+6 months

100%

+12 months

New featuresMaintenance & firefighting

Token / LLM cost compounds with every user

Model deprecated → re-architect on the next one

The Software Lifecycle · What Breaks

What Breaks — Constantly — After You Ship

None of it is on your roadmap. All of it becomes your team’s job.

Model Deprecation & Lock-In

Build on today’s model and a better one ships next week. Deprecations force a re-architecture.

Runaway LLM Cost

Every feature is a recurring token bill that scales with every user and every message.

Security & CVEs

New vulnerabilities surface daily. Patch fast — or get breached.

Dependency Upgrades

Hundreds of packages. Breaking changes land almost every week.

API & Schema Drift

MLS, CRM, and vendor APIs change shape without warning.

Language, Runtime & Codec Churn

Node, Python, frameworks, and media formats shift under your pipeline.

The Software Lifecycle · Do It Responsibly

If You’re Going to Build, Know the Floor.

Never let a DIY app touch

✕Consumer-grade AI models (ChatGPT / Claude / Gemini)

✕Customer PII & contact records

✕SSNs, banking & financial data

✕Signed contracts & disclosures

✕Anything under Fair Housing / RESPA

Minimum guardrails

Requires an enterprise account

✓Keep regulated data out of prompts

✓Enforce strict data-retention limits

✓Log & audit every data access

✓Human review before anything client-facing

Load a client’s data into a tool you can’t secure — or suffer a leak — and you’re liable under CCPA, GLBA, and state privacy law.

Or skip the liability — BrokerBot ships every one of these guardrails on day one.

Not someday. Now.

“The industry doesn’t have a technology gap — it has a governance gap.”

BrokerBot is here to help associations get AI-ready.

brokerbot.ai