The 39-Minute Mirage: Audit the Leads Your AI Dashboard Leaves Out

By Anna with Oppy

Your AI pilot can report a faster median response while leaving more people unanswered. Both numbers can be correct. The missing denominator is the problem.

In a 2024 secret-shopping study, Mike DelPrete found that 47% of online property inquiries received no response. The mean response time was 8 hours 17 minutes. The median was 39 minutes. Researchers tested more than 25 brokerages, two or three times each, through brokerage websites during weekday business hours, asking about median-priced properties.

This is a historical sample, not a 2026 industry estimate. The mean-to-median gap signals a slow tail, not its cause. More importantly, a response-time statistic cannot tell you how many inquiries never received a reply.

Build an audit that starts with every incoming inquiry, not just conversations the assistant touched. Call the proposed role a Denominator Oppy.

New agents need better receipts

In an October 13 Inman contributor essay, Kyle Crawford offered a useful test for brokerage technology:

“Ask for a live login, and watch a basic task actually happen, not a screenshot in a deck.”

Apply that standard to the dashboard, too. Can it show an inquiry that received nothing?

BCG's September 30 Applied AI Index, surveying more than 1,300 CxOs and senior leaders across 20-plus sectors, found that 44% of its most AI-mature companies reported realizing agentic-AI value, versus 2% of laggards. That self-reported, cross-sector comparison is not a real estate return-on-investment estimate. Your own measurement still needs to survive inspection.

On October 7, Microsoft announced policy-driven execution containers for AI agents, with access boundaries outside the agent's control. The transferable principle is simple: define the job's permissions in the system, not merely in its instructions. This announcement does not establish an Oppy integration.

Define the denominator before the pilot

For a two-week pilot, choose one inbound source and a fixed 24-elapsed-hour observation window per inquiry. This is an experimental definition, not a recommended service standard or legal deadline.

Keep inquiries younger than 24 hours in a separate pending cohort. For older inquiries, inspect only what happened within their first 24 hours. Otherwise, Monday's inquiry gets more time to succeed than Friday's.

Keep staffed-hours and after-hours cohorts separate. Retain duplicates and spam in the raw ledger, then label any exclusion with a reason. Never quietly delete inconvenient rows.

Audit rule: The assistant's activity log is evidence of its activity. It is not evidence that every incoming inquiry reached it.

A brokerage might audit showing questions. A mortgage office might audit requests for a callback. A title team might audit status inquiries. Use the same source before and during the pilot; do not call a quieter week an AI success.

Give one Oppy a narrow, read-only job

Oppy's setup guide documents goals, personality, a teach section, prompts, imported webpages and reviewing unsent drafts. Use those supported concepts to explain the audit role and approved definitions.

The ledger and audit role below are a proposed configuration, not a shipped Oppy analytics product, schema or connector. Provide sanitized, approved records through an input method your account supports. If the required records cannot be supplied, perform that portion manually. A prompt cannot create an integration.

For each inquiry, preserve these fields:

inquiry_id: stable, non-sensitive identifier
source: approved inbound source
received_at_utc: original receipt timestamp
observation_end_utc: received_at_utc plus 24 elapsed hours
first_ack_at_utc: acknowledgement, if any
first_substantive_reply_at_utc: useful reply, if any
reply_evidence: redacted text or approved record reference
staffed_hours_cohort: staffed / after_hours / unknown
outreach_hold_reason: recorded reason, or none / unknown
exclusion_reason: recorded reason, or none

A substantive reply answers the request or asks a necessary question that moves it forward. “We received your message” is an acknowledgement, not a resolution. Record both. Preserve corrections rather than overwriting the original timestamps.

Copy this audit prompt

Replace the bracketed values with approved policies. Configure actual tool access separately.

Role: Denominator Oppy for [TEAM NAME].
Purpose: audit inbound coverage, not send messages or evaluate customers.

Inputs: approved, sanitized inquiry ledger and reply evidence.
Source: [ONE INBOUND SOURCE].
Observation: first 24 elapsed hours after each inquiry's UTC receipt.
Snapshot time: [EXACT UTC TIMESTAMP].

For every inquiry ID, return one label:
PENDING_WINDOW
SUBSTANTIVE_REPLY_WITHIN_WINDOW
ACKNOWLEDGEMENT_ONLY
NO_REPLY_EVIDENCE_IN_WINDOW
MISSING_OR_CONFLICTING_RECORDS

Rules:
1. Preserve every input ID. Flag duplicates; do not silently remove them.
2. Do not infer a reply, timestamp, consent or exclusion.
3. Cite the record supporting each label. Missing evidence is not proof that no reply occurred.
4. Preserve recorded outreach holds separately. Do not label a lawful hold an AI failure.
5. Flag contradictory times and missing source records for human review.
6. Do not contact anyone, change CRM records, access loan documents or alter permissions.
7. Produce labels and exceptions. Have approved deterministic tools calculate counts and elapsed times; do not invent missing calculations.

Return the labeled records, exceptions and unanswered audit questions.

Compare the exported inquiry count with the audited ID count. A human resolves exclusions and missing evidence before anyone announces results.

Report coverage beside speed

Use a verified spreadsheet or approved calculation tool, not an unsupported AI estimate, to report:

A missing record stays unresolved. It does not become a success because the chart needs a cleaner ending.

An outreach hold is not a defect

This audit must never trigger a rescue text merely because a row looks silent. The FCC treats AI-generated human-sounding voices as artificial or prerecorded voices under the TCPA; advertising or telemarketing calls require prior express written consent, subject to applicable exemptions.

Twilio's A2P 10DLC guidance describes provider registration and verifiable opt-in requirements. Registration is not proof of statutory consent. Keep revocations and do-not-contact controls in the actual sending systems, with qualified counsel reviewing applicable rules.

The pilot's first win may be discovering that your old reporting omitted people. Fix that before making the assistant faster. You cannot improve a queue your dashboard refuses to admit exists.

References

  1. Mike DelPrete, “Secret Shopping: 47% of Online Property Inquiries Are Ignored”, August 13, 2024.
  2. Kyle Crawford, Inman, “Your brokerage's tech is a promise. Make them keep it”, October 13, 2026, contributor essay.
  3. BCG, “The Formula for Agentic AI Value”, September 30, 2026.
  4. Microsoft, “Microsoft Execution Containers: Policy-driven containment for AI agents”, October 7, 2026.
  5. Oppy, “Anatomy of an Oppy - A.I Virtual Assistant”, August 20, 2024.
  6. Federal Communications Commission, FCC 24-17, released February 8, 2024.
  7. Twilio, “A2P 10DLC registration application quickstart”, accessed October 14, 2026.