Skip to main content
FirstCoast.ai
All case studies
Case study: Family real estate holdings

Elder-First AI Assistant

A family company with two properties needed both its day-to-day operations and its principal looked after. We built a private hub whose primary user is a non-technical owner in his eighties, with an AI assistant designed around what it cannot do.

How the elder-first assistant is fenced inA family company's private hub, built around a non-technical owner in his eighties. He works in large type and plain language, in the four topics he actually thinks in: the properties, the household, scheduling, and getting help. The assistant behind it runs on small, fast models and will answer anything he asks: the fence is around what it can do, not around what it can discuss. Every action it can take is on one list of four. Present options, record notes, update the family calendar, and escalate to family. Nothing in that list touches money, accounts, or any external site, suspicious or otherwise, so those sit outside the fence entirely. Escalation is the one path out: a request that needs action reaches the family as approve and reject buttons in Telegram, and an approval triggers the action from the system of record, never from the model. Separately, during a house clean-out the owner photographs an item and AI vision identifies it, flagging anything worth an appraisal before it leaves the house. Deadlines and news monitoring run as deterministic scheduled workflows with no AI in the loop, so a missed API call can never mean a missed date, and a weekly plain-English digest reaches the family through Gmail, which is the visibility they needed without a phone call. Scam protection is the next module, and it works the same way: his inbox is watched, junk is tidied daily, unsubscribes happen only on his own one-tap approval, and anything that looks like a scam is diverted and flagged to the family rather than left sitting in front of him to answer. That watcher reads and escalates like every other part of the system, so the protection arrives without handing the assistant a single new power, and it waits on his mail being consolidated into one account first. All of it sits on an isolated database with row-level security, scoped credentials, and no delete grants anywhere, and the whole platform runs on under $20 a month in AI spend.THE OWNERFAMILY HUBlarge type, plain language,patient answersPropertiesHouseholdSchedulingGetting helpThe assistantsmall, fast modelsIt answers freely. Only acting is fenced.EVERY ACTION IT CAN TAKEPresent optionsRecord notesUpdate the family calendarEscalate to familyIT NEVER TOUCHESMoneyAccountsSuspicious sitesTHE DATAIsolated databaserow-level securityNO DELETE GRANTSRUN COSTUnder $20a month in AI spendTHE FAMILYTELEGRAMApprove or rejectthen the action runs fromthe system of recordNEVER FROM THE MODELTHE CLEAN-OUTPHOTOHe photographs an itemduring the house clean-outAI vision identifies itflags anything worthan appraisal before it leaves the houseON A SCHEDULEDeadlines and newsNO AI IN THE LOOPa missed API call cannot meana missed dateGMAILWeekly plain-English digestvisibility without a phone callSCAM PROTECTIONINBOXSuspicious mail divertedTHE FAMILY IS TOLDjunk tidied, scams held backnext, once the mail is consolidated

The challenge

  • A non-technical owner in his eighties needed real help with property matters, scheduling, and a house clean-out, without learning new software
  • The family managing the company needed visibility without turning every small question into a phone call
  • Anything touching money, accounts, or outside parties had to stay in human hands, and that was non-negotiable

What we built

A private family hub with a chat assistant designed elder-first: large type, plain language, patient answers, organized into the topics the owner actually thinks in: the properties, the household, scheduling, and getting help.

  • The tool list is the guardrail: the assistant's only capabilities are presenting options, recording notes, updating the family calendar, and escalating to family. Nothing it can call touches money, accounts, or external sites
  • The owner photographs household items during a clean-out; AI vision identifies each one and flags potentially valuable pieces for appraisal before they leave the house
  • Requests that need action go to the family as approve and reject buttons in Telegram; approval triggers the action from the system of record, never from the model

How it's built

A web hub on an isolated database with row-level security and deliberately scoped credentials, with no delete grants anywhere. Deterministic scheduled workflows handle deadlines and news monitoring with no AI in the loop, so a missed API call can never mean a missed date, and a weekly plain-English digest goes out through Gmail. The assistant runs on small, fast models with per-topic context, so cost tracks the active conversation, not the archive. The next module extends the same escalate-to-family pattern to his inbox: junk tidied daily, unsubscribes only on his one-tap approval, and anything that looks like a scam diverted and flagged to the family instead of left sitting in front of him.

Results

  • An owner in his eighties uses an AI assistant daily, unassisted, because it was designed for him
  • The clean-out photo flow has already flagged items worth a professional appraisal before they left the house
  • The entire platform runs on under $20 a month in AI spend, on infrastructure the family already had

Who in your business needs technology to meet them where they are?

Elder-first design is just honest design. We build assistants people actually use.