13.9 Million Homes, 380,000 Starters: Build the AI Move-Up Relay

13.9 million homes. Only 380,000 starter-size homes. That is the uncomfortable ratio in Realtor.com’s October 5 housing-succession research. The firm projects that Baby Boomer and Silent Generation households will vacate a net 13.9 million owner-occupied homes from 2026 through 2036. Of those, it estimates 0.38 million with zero to two bedrooms, 9.9 million with three to four, and 3.6 million with five or more. HousingWire reported the finding the same day.

The 380,000 represent about 2.7% of the projected total. That is a calculation from Realtor.com’s rounded estimates, not a measured share of future listings. A released home means a net reduction in homes occupied by those older owners. It may never appear for sale. A small home is not necessarily an affordable one. The national projection cannot tell you which address, buyer, or market will move next.

So do not build an AI agent that turns a demographic forecast into a lead list. Build one that finds the move-up relay already visible in your authorized files: a household looking for a larger home that has actually said it may sell its smaller one. That possible second listing matters to a first-time buyer, but only if a real person confirms the intent and the timing.

The assignment: Match a current, permissioned buyer need to a current listing. If a separate, consented seller file indicates a smaller home may become available after a move, prepare a two-file brief for human review. Never treat the national forecast as inventory.

Three agents, one carefully limited job

This is an original proposed configuration, not a prebuilt Oppy template, MLS integration, or performance claim. Oppy documents knowledge search, notes, checklists, conversation status, follow-ups, administrator permissions, and a needs_attention human hold. Verify the functions and approved data connections in your account before configuring anything.

1. The Local Mix agent reads an approved, dated listing export for a defined market. It counts active homes by zero-to-two, three-to-four, and five-plus bedrooms. It labels the period, geography, source, listing status, and whether a record is a listing or a completed sale. Realtor.com’s decade-long projection stays in a separate context field. No invented local forecast, no person-level age targeting.

2. The File Fit agent reads only consented buyer requirements and authorized listing or seller records. It explains why a specific active home fits or misses the buyer’s stated constraints. It may flag a possible second file only when the owner has explicitly indicated an intention to sell and the organization may use that information for this purpose. It cannot infer an owner’s plans from age, address, or a national dataset. No second file? Say so.

3. The Handoff agent drafts an internal brief for an assigned human. It names the buyer-side question, the separately confirmed seller-side question, the unresolved dates, and the service owner for each next step. A mortgage professional handles financing questions. Title and insurance staff assess their own work. A transaction coordinator checks deadlines. The agent does not approve a loan, clear title, bind coverage, promise availability, or send an outreach message. The person approving an external step remains accountable.

This is not an argument for a giant new system. HomeServices’ August addition of mortgage servicing to its OnePoint network illustrates how a residential journey can span brokerage, mortgage, title, insurance, and the post-close relationship. That reported service expansion does not establish an AI implementation. It does make the handoff problem concrete.

The prompt that prevents a forecast from becoming a fictitious listing

Give the File Fit agent an approved source shelf and this instruction. Replace every bracket with your organization’s own permitted data and role names.

ROLE: Move-Up Relay file analyst for [TEAM] in [DEFINED MARKET].

AUTHORIZED INPUTS ONLY:
- [Dated, permissioned active-listing export and its field definitions]
- [Buyer's volunteered requirements and approved contact record]
- [Seller's explicit, recorded intent, if a separate file exists]
- [Approved transaction-status and service-owner records]

Realtor.com's 2026-2036 national projection is BACKGROUND, never a
property-level signal or a list of sellers.

For each buyer case:
1. State the requested price range, bedrooms, location, and timing only
   if the buyer supplied them. Cite the record and its timestamp.
2. Compare each proposed active listing with those stated constraints.
   Mark FIT, MISMATCH, or UNKNOWN, with a source for each material fact.
3. If a distinct, authorized seller file records intent to sell a
   smaller home, label it POSSIBLE SECOND FILE. Cite the exact recorded
   intent. If none exists, write: "No supported second file."
4. Identify the unanswered question for the assigned human owner.
   Do not assume that either transaction will occur.
5. Prepare a draft internal handoff with both file IDs, source links,
   stale-data flags, human owners, and next review date.

Do not infer age, family status, protected traits, eligibility,
coverage, title clearance, value, or future price. Do not send a
message, alter a record, book a showing, or make a customer promise.
If sources conflict or an owner is unconfirmed, set HUMAN REVIEW
REQUIRED and stop.

The useful output is a case card, not a cheerful paragraph: buyer need | current listing fit | supported second file or none | evidence + dates | missing fact | assigned human | proposed next step. A reviewer should be able to open every cited record. Oppy’s documented needs_attention pattern is one possible human hold, subject to how your account and permissions are configured. Oppy Functions Guide

Keep unlike supply signals in separate drawers

A second, less-discussed supply measure makes the point. The Census Bureau and HUD’s September 17 release reported August 2026 housing completions at a 1.128 million seasonally adjusted annual rate, down 27.1% year over year, with a published margin of error of ±8.9 percentage points for the change. That is an annualized rate, not 1.128 million homes completed during August. It is also not the number of Realtor.com’s projected older-owner homes offered for sale.

A useful agent stores the two series with different labels, dates, geographies, and definitions. It does not add them together or infer that national completions predict a particular brokerage’s inventory. The question for a local team is narrower: Which active, permitted files are waiting on a buyer, a seller, financing, title, or insurance, and who owns the next decision?

Test a relay before declaring a revolution

Run a small pilot with one market, one permissioned listing export, and perhaps 20 consented cases. Have staff review every case card before any action. Log four things: source-link errors, false fits, supposed second files lacking explicit seller intent, and time from brief to a documented human disposition. Count a move-up chain only when both file owners confirm the relationship. Do not report case cards as closings.

In the October 5 Chrisman Commentary episode, Mortgage Education Institute’s Andrew Conner said of interactive professional training, “People actually learn when they're involved.” He was discussing mortgage education, not this workflow. Borrow the discipline, not an endorsement: ask reviewers to challenge a few cards each week and teach the agent where the evidence or language failed.

If the relay works, the win is not that AI predicted a “silver tsunami.” It is that a buyer did not receive the wrong house, a potential second file was not invented, and the right person could see the next decision before another follow-up disappeared into the inbox.

References