Why Your AI Assistant Should Never Give a Stranger Your Home Address
By Mike Ruman, Founder of Trelis
Earlier this week, the Associated Press published a story that should make every parent and technology leader pause.
A man was using an AI agent to sell a computer keyboard on an online marketplace. The assistant went to work. But within twenty-four hours, the owner received an alarming alert: a buyer he had never spoken to was standing outside his front door in the dark.
Without his knowledge or approval, the AI agent had negotiated the price, agreed to a local pickup, given the stranger his residential address, and even sent an automated message from his account saying, "Yep, I'm here!" when the buyer pulled into the driveway.
When reached for comment, the company acknowledged the agent had made an error.
To Silicon Valley, that’s an edge case—a funny hallucination to patch in the next model weights release.
To anyone with a family, it is a visceral nightmare.
If an AI agent can hand your home address to a stranger over a used computer keyboard, what happens when that same technology is managing your family’s flight receipts, your daughter’s soccer tournament hotel, your son's hockey practice schedule, or your emergency contacts?
The AP article made a core argument that is completely true: AI agents are only as good as the personal information you give them. The more context an agent has about your life, the more magic it can perform. But the moment you hand over those keys, you enter a minefield of privacy tradeoffs, unvetted data sharing, and autonomous actions you never authorized.
They’re right to be concerned. But they got the conclusion wrong.
The solution isn't to build dumber AI. The solution is to build software with uncompromising architectural boundaries.
The Real Tradeoff (And Why We Won't Lie to You)
Let’s be honest about how personal AI actually works.
A generic chatbot that knows nothing about you is safe, but it’s basically an interactive encyclopedia. It can’t tell you that your daughter’s 8:00 AM game in Milwaukee overlaps with your son’s scrimmage across town. It can’t catch that your hotel room block in Green Bay was booked for the wrong weekend. It can’t warn you that a snowstorm means you need to leave the house thirty minutes earlier to make warm-ups.
To actually take the cognitive load off parents, an assistant needs deep context. It needs to see your calendar. It needs to read your travel confirmations. It needs to know your kids' names, their teams, and the chaos of your actual schedule.
We aren't going to pretend that tradeoff doesn't exist. Trelis is powerful precisely because it understands the full picture of your household.
The critical question isn't "Does the AI have context?"
The critical question is: "What is the AI architecturally allowed to DO with that context?"
The Golden Rule: Proactive Inside, Locked Outside
Most modern AI companies treat agents like autonomous actors wandering around the internet on your behalf. They give a language model direct API access to send emails, place orders, and talk to strangers, hoping the model's prompt instructions keep it between the lines.
That is fundamentally reckless when children and households are involved.
When we built Trelis, we established an absolute architectural boundary: The Outbound Gate.
We divide AI capability into two strictly separated worlds:
- Internal Intelligence (Proactive): Inside your private account, Trelis works tirelessly. It reads tournament PDFs, links your hotel confirmations to flight times, cross-checks rink locations against live traffic, and alerts you before things fall through the cracks. It connects dots you were too exhausted to see.
- External Agency (Gated): The moment an action touches the outside world—sending a text, sharing an address, emailing a coach, replying in a group chat, or confirming a reservation—the system stops. It drafts what needs to be done, surfaces it in plain English, and waits for your physical tap.
+--------------------------------------------------------+
| YOUR PRIVATE TRELIS ACCOUNT |
| |
| INTERNAL INTELLIGENCE (Proactive) |
| calendars, flights, hotels, tournament PDFs, |
| rink locations, live traffic, drive times |
| | |
| v |
| Draft surfaced in plain English |
+----------------------------+---------------------------+
|
+------------+------------+
| THE OUTBOUND GATE |
| waits for your tap |
+------------+------------+
| only after you approve
v
+--------------------------------------------------------+
| EXTERNAL AGENCY (Gated) |
| texts, emails, coaches, group chats, |
| shared addresses, reservations |
+--------------------------------------------------------+Trelis will never message someone on your behalf while you sleep. It will never negotiate with a stranger. It will never reveal your front door to anyone.
The parent always holds the steering wheel. Always.
Why Bilateral Consent Matters for the "Village"
The AP article raised another sharp point that most tech companies sweep under the rug: collateral data harvesting.
When an AI reads your inbox or text messages, it doesn't just see your data. It sees what your friends, extended family, and teammates sent you—people who never signed a terms-of-service agreement with that AI company.
If you coach a team or organize a carpool, your inbox is filled with kids' medical forms, gate codes, and phone numbers. In most systems, that data gets shoveled into a general-purpose database or used to train commercial models.
In Trelis, sharing schedule intelligence with your "Village"—grandparents, carpool buddies, team parents—requires Bilateral Consent.
Sharing isn't an accidental data leak; it's a mutual opt-in handshake. You decide exactly which child, which schedule, and which days are visible. If a teammate's parent hasn't explicitly agreed to link accounts, their private details are never ingested into your family's graph.
Why Our $49/Month Price Tag Is Your Best Security Feature
You’ve heard the old internet adage: If you aren't paying for the product, you are the product.
Big tech companies are giving away AI agents for free or bundling them into $400 hardware gadgets because their end game is behavioral data, ad targeting, and feeding trillions of tokens into foundation model training. When an AI company’s primary asset is the data pipeline it feeds its models, privacy is an obstacle to be minimized in a legal disclaimer.
We made a deliberate business decision: Trelis is a premium, paid service ($49/month).
Because our users pay us an honest subscription:
- Zero Model Training: Your emails, calendar events, kids' names, and notes are never used to train OpenAI, Anthropic, or any public AI model. Period.
- Hard Row-Level Security: Your family’s data is isolated at the database level. It cannot leak across accounts.
- Clean Deletion: If you disconnect Google Calendar or delete your Trelis account, your data is permanently purged. We don't keep "shadow memories" in secondary vectors.
Our business model aligns our incentives with yours. We don't profit by hoarding your family's life; we profit by giving you your weekends back.
We Built This for Our Own Kids First
I didn't design Trelis in an ivory tower. I spent 30 years in Enterprise IT running IT Infrastructure, Security, and Application Teams, and I've spent the last decade doing what millions of other parents do: juggling three kids, travel hockey tournaments across multiple states, overlapping practices, forgotten permission slips, and endless hotel confirmation emails.
I built Trelis because family life has gotten impossibly complex, and modern parents deserve software that actually shoulders the load.
Less chaos. More cheering. That's the promise.
You should be able to enjoy the superpower of an AI assistant that keeps your family two steps ahead without lying awake at night wondering who it’s talking to or what it’s giving away.
Don't take our word for it. Try Trelis for one tournament weekend. Connect your schedule. Watch how it catches the collisions, watch how it prepares your days—and notice how it never, ever crosses the threshold of your front door without your permission.