Permits Rose 5%. Starts Fell 12.4%. Your AI Needs to Know Both Can Be True.

Description: A practical guide to building a six-Oppy market desk that separates permits from starts, applications from closings, and public inventory from broker-visible inventory before a human approves the story.

Author: Anna with Oppy
Date: September 9, 2026
Category: Agentic AI
Keywords: Oppy, AI agents, real estate market data, housing permits, housing starts, mortgage applications, MLS data, agentic AI, human approval, A2P 10DLC, TCPA compliance

July issued housing operators three honest numbers that look like a typo.

Building permits rose 5.0% from June to a seasonally adjusted annual rate of 1.443 million. Housing starts fell 12.4% to 1.239 million. Completions fell 9.1% to 1.212 million, according to the joint release from the U.S. Census Bureau and the U.S. Department of Housing and Urban Development.

All three can be true because they measure different stages of construction.

An AI assistant can quote every number correctly and still write a false market story. It only has to compress three stages into one sentence such as, “Housing supply improved 5% in July.” Fluent. Clean. Wrong.

The fix is not a longer prompt. It is a small operating system for evidence.

This guide shows how to configure six narrowly scoped Oppies that ingest approved market releases, turn each fact into a structured signal card, block incompatible comparisons, add authorized local context, draft a brief, and wait for a named human to approve the exact version. It is a blueprint built from Oppy’s publicly documented role, knowledge, permission, tool, communication, and human-attention capabilities. It is not a claim that a prebuilt “Market Desk” template currently exists.14 15

The job is not to make the market sound simple. The job is to make every sentence traceable.

The market does not run on one clock

The construction split is only the opening example.

The National Association of REALTORS reported 4.06 million existing-home sales at a seasonally adjusted annual rate in July, down 1.7% from June and up 0.7% from a year earlier.2 Its separate pending-home-sales index, which tracks signed contracts rather than completed transactions, fell 2.3% from June and 2.2% from a year earlier.3

The Mortgage Bankers Association then reported that purchase applications rose 2.0% during the week ending August 28, while the unadjusted purchase index remained 0.2% below the same week a year earlier.4 Freddie Mac reported a 6.71% average 30-year fixed rate for a defined borrower and loan profile.5 The Federal Housing Finance Agency reported that house prices rose 2.1% year over year in the second quarter.6

These are not rival answers. They are answers to different questions.

A credible market brief must preserve six clocks:

1. Release date: When the publisher issued the number.
2. Observation period: When the measured activity occurred.
3. Transaction stage: Inquiry, application, contract, closing, permit, start, or completion.
4. Geography: National, state, metro, county, ZIP code, service area, or property.
5. Method: Sample, census, index, seasonality, denominator, and revision policy.
6. Revision status: Preliminary, revised, final, proprietary nowcast, or modeled estimate.

If one clock is missing, the number is not ready for client communication.

Why this matters now

Agentic software can perform more operational work inside ordinary business systems. OpenAI’s September 3 announcement says its newest computer-use model can fill forms, update CRM records, and organize calendars.9 Meta’s September 8 launch described a consumer agent that can act across connected applications, while independent reporting from Reuters also described internal testing concerns involving stalling, silent errors, repeated logins, and a reported data-exposure incident.10

The useful conclusion is not “turn on autonomy.” It is separate evidence handling, judgment, and release authority.

A HousingWire podcast released within the last 24 hours described its conversation with Total Expert CEO Joe Welu as being:

“about AI and how companies are using it to turn relationships into appreciating assets.”

HousingWire editorial description, September 8, 2026.13

That is an attractive objective. Relationships appreciate when the information supporting them survives scrutiny. They depreciate rather quickly when this week’s polished update quietly mixes last month’s contracts, yesterday’s rates, a national price index, and a portal inventory count.

Mike DelPrete put the implementation problem neatly in his analysis of coming-soon inventory:

“The difference is the ground game, not the product.”7

The same is true here. A clever model is useful. A disciplined handoff is the product.

Build the signal card before you build the prompt

Every approved fact should become an immutable signal card before any Oppy drafts prose.

signal_card_id: census-nrc-2026-07-permits
publisher: U.S. Census Bureau and HUD
source_url: https://www.census.gov/construction/nrc/current/index.html
release_date: 2026-08-18
observation_period: 2026-07
geography: United States
transaction_stage: permit
metric: privately-owned housing units authorized
value: 1.443 million
unit: seasonally adjusted annual rate
comparison: 5.0% above revised June rate
revision_status: preliminary release, prior period revised
methodology_note: see source definitions and sampling notes
permitted_use: national construction-stage context
prohibited_inference: completed supply rose 5.0%
human_owner: market_editor

This format is intentionally tedious. Tedious is underrated when the alternative is confident fiction.

A useful card must include the source, release date, observation period, geography, stage, unit, denominator, adjustment method, revision status, methodology, permitted use, prohibited inference, and named owner. The prose comes later.

The six-Oppy Market Desk

Oppy publicly describes configurable AI employees with access to communication channels and over 60 business tools. Its product materials also document uploaded knowledge, Public and Admin permissions, admin allowlists, and a needs_attention flow that can hold outbound work for human response.14 15

Use those primitives to create six roles with incompatible powers.

This structure prevents the same agent from sourcing a claim, judging its compatibility, polishing its language, and sending it. Efficiency fans may consider that excessive. Auditors call it Tuesday.

The handoff sequence

Approved source
  -> Source packet
  -> Signal card
  -> Compatibility decision
  -> Authorized local context
  -> Draft with claim-level links
  -> Human approval token
  -> Exact release

A handoff is complete only when the receiving role accepts a named object, such as source_packet_id, signal_card_id, or draft_hash. A chat message saying “done” is not a control.

Copy this prompt into the Compatibility Referee Oppy

The Compatibility Referee is the article’s central operator. It needs rules, not vibes.

ROLE
You are the Compatibility Referee for a residential real estate market desk.
You do not research, edit source values, draft persuasive copy, or publish.

INPUTS
You receive immutable signal cards and one proposed sentence or comparison.

REQUIRED TESTS
1. Compare transaction stage.
2. Compare observation period and release date.
3. Compare geography.
4. Compare unit and denominator.
5. Compare population and inventory universe.
6. Compare seasonal adjustment and revision status.
7. Test whether the sentence changes correlation into causation.
8. Test whether a national measure is being presented as local evidence.
9. Test whether a benchmark rate is being presented as an individual quote.
10. Test whether portal-visible inventory is being presented as all inventory.

DECISION
GREEN: Fields are compatible and the sentence preserves the source meaning.
YELLOW: The evidence may appear together only with a visible qualifier.
RED: The comparison or inference is prohibited. Send needs_attention to the
named Market Editor. Do not draft a replacement claim unless the editor asks.

OUTPUT
Return decision, matched fields, mismatches, required qualifier,
prohibited inference, source links, and affected signal_card_ids.

Three red-team tests

The third test is not theoretical. Northwest MLS says its First Look option gives access to its 30,000-plus member brokers while allowing a seller to withhold Internet Data Exchange distribution for up to 21 days.8 In that setting, “visible to a broker” and “visible on a public portal” are different inventory universes.

A worked brief that does not invent a market

Assume the desk receives five approved inputs:

• Census construction data for July.1
• NAR existing and pending sales for July.2 3
• MBA mortgage applications for the week ending August 28.4
• Freddie Mac’s September 3 benchmark rate.5
• An authorized local inventory export with a documented license and observation date.

The system should not calculate a synthetic “market health score.” There is no defensible denominator for an average of permits, contracts, applications, rates, and local listings.

It should produce something closer to this:

National construction moved in different directions in July. Permits rose 5.0% from June, while starts fell 12.4% and completions fell 9.1%. These figures represent separate construction stages and do not establish a change in locally available inventory. Existing-home sales and pending contracts also measure different points in the resale process. The local panel below uses an authorized brokerage dataset and is not directly comparable with national series. A lender must answer questions about a specific borrower’s rate or qualification.

That paragraph is less thrilling than “the market just flipped.” It is also usable by a serious operator.

Add local context without laundering it

The Local Context Oppy should add a separate panel. It should never overwrite a national card.

For every local metric, require:

RISMedia recently reported concerns about browser extensions and confidential MLS information. WAV Group principal Victor Lund summarized the risk this way:

“The exposure is invisible because it looks completely like normal behavior.”11

The operational answer is not to forbid every useful tool. It is to use approved sources, least-privilege access, explicit data rights, and immutable provenance. No Oppy should receive restricted MLS or client data merely because a user can see it on a screen.

Put the human on the release, not in the footnotes

The Briefing Editor can draft. The Release Controller can distribute. Neither should have final editorial authority.

The human Market Editor should see one review screen or board containing:

• The draft and its draft_hash.
• Every claim-linked signal card.
• Green, yellow, and red compatibility results.
• Required qualifiers.
• Missing or stale fields.
• Local data license and distribution status.
• Proposed audience and channel.
• Open needs_attention items.
• A named correction owner and stop-send control.

BCG’s recommendation for decision agents is admirably narrow:

“Companies should choose one cross-functional, high-friction decision process to test.”12

A weekly market brief for one geography is enough. Do not begin with automated market predictions for every lead in the CRM. Ambition is not a test plan.

If the brief leaves the building, add channel controls

The market desk itself should use aggregate evidence and avoid client personal information. Distribution introduces a separate control layer.

For email, CRM notes, or website publishing, the Release Controller should validate the approved version, audience, and correction owner. For text or voice outreach, it must also check consent provenance, Do Not Call and suppression status, sender and campaign identity, opt-out handling, jurisdiction, and the planned use of artificial or prerecorded voice.

The Federal Communications Commission treats AI-generated human voices as artificial or prerecorded voices under the Telephone Consumer Protection Act. Covered telemarketing calls require the applicable prior express written consent.16 The broader federal revocation rules are already operative. A limited waiver delays only the cross-topic global-revocation consequence until January 31, 2027, not the entire opt-out regime.17

Application-to-person messaging over a U.S. 10-digit long code also requires appropriate A2P 10DLC registration under carrier rules. Registration helps identify the sender and use case. It does not create legal consent.18

The Release Controller should hold outbound work when any required field is missing. Counsel or a qualified compliance owner should decide the actual rules for the channel and jurisdictions involved.

What the Market Desk should refuse

A useful refusal policy is short.

The system should not forecast prices, estimate an individual’s mortgage rate or affordability, value a property, recommend a neighborhood, rank schools, interpret insurance availability, give legal advice, or turn aggregate data into a client-specific conclusion.

It should also refuse to:

• Average incompatible sources into one score.
• Replace a release date with an observation date.
• Convert a weekly change into a yearly trend.
• Generalize a local policy nationwide.
• Present a portal count as the complete market.
• Remove a qualifier to improve the headline.
• Send while needs_attention remains open.

Refusal is not failure. It is the moment the system proves it knows the job.

Run a four-week internal pilot

Start with one geography, one weekly internal brief, six roles, and one named Market Editor.

During week one, build the signal-card schema and source allowlist. During week two, test the three red-team sentences above plus five examples from your own previous market updates. During week three, add one authorized local dataset. During week four, let frontline staff use the internal brief and record where they still need a lender, insurer, appraiser, broker, or counsel.

Measure the system with operational evidence:

Do not lead with word count, draft volume, or token price. A fast market brief that needs a public correction is not productivity. It is expedited archaeology.

The opportunity

The newest housing data is not waiting for a better storyteller. It is waiting for a better referee.

Permits can rise while starts fall. Applications can rise week over week while remaining below last year. Prices can increase nationally while individual metros decline. A broker can see inventory that a portal user cannot. Every sentence can contain accurate numbers and still mislead.

A six-Oppy Market Desk makes the distinctions operational. One role preserves the source. One maps the metric. One tests compatibility. One adds authorized local context. One drafts. One releases only the human-approved version.

The result is not an artificial market oracle. It is more useful: a briefing system that knows when two true facts do not belong in the same conclusion.

Editorial and compliance note: This article provides an operational design pattern, not legal, lending, insurance, valuation, investment, or fair-housing advice. Confirm current Oppy features, permissions, integrations, data rights, and applicable laws before implementation.

References

1. U.S. Census Bureau and HUD, “Monthly New Residential Construction, July 2026”
2. National Association of REALTORS, “NAR Existing-Home Sales Report Shows 1.7% Decrease in July”
3. National Association of REALTORS, “NAR Pending Home Sales Report Shows 2.3% Decrease in July”
4. Mortgage Bankers Association, “Mortgage Applications Increase in Latest MBA Weekly Survey”
5. Freddie Mac, “Mortgage Rates Average 6.71%”
6. Federal Housing Finance Agency, “U.S. House Prices Rise 2.1 Percent Year over Year”
7. Mike DelPrete, “Unique Inventory and the Coming Soon Scoreboard”
8. Northwest MLS, “Northwest MLS Expands Listing Options”
9. OpenAI, “GPT-6 Astra: A New Generation of Intelligence”
10. Reuters, “Meta Launches AI Agent That Can Access Other Apps”
11. RISMedia, “How AI Tools May Be Exposing MLS Data and Your Brokerage to Risk”
12. Boston Consulting Group, “Decision Agents Bring AI into the Boardroom”
13. HousingWire, “Joe Welu on How AI Is Changing Customer Relationships”
14. Oppy, “Virtual Assistants in a Click”
15. Oppy, “Oppy Functions: 52 Tools with Examples and Best Practices”
16. Federal Communications Commission, “Declaratory Ruling on AI-Generated Voices”
17. Federal Communications Commission, “FCC Extends Effective Date of Limited TCPA Consent Revocation Rule”
18. Twilio, “Programmable Messaging and A2P 10DLC”