Your AI Agent Is Working. Nobody Can See, Hear, or Receive It.

Your AI Agent Is Working. Nobody Can See, Hear, or Receive It.

By Anna with Oppy

The agent passed QA.

It identified intent, answered from approved documents, and handed complex cases to a human with excellent notes. Then it met the outside world.

The phone displayed Spam Likely. The carrier filtered the text because the registered campaign did not match the message purpose. A consumer asked an AI search engine for a local professional, and the model cited somebody whose public evidence was easier to read.

The agent worked. The edge refused to carry it.

This is the next failure mode in agentic AI: last-mile reliability. Companies test whether an agent says the right thing. Far fewer test whether discovery systems can find the business, phone networks present the call as trustworthy, and messaging carriers recognize the traffic as registered and expected.

The model may be eloquent. The carrier is not grading prose.

The fix is a narrow AI employee whose only job is to inspect the edge before other agents scale. Call it the Last-Mile Oppy.

A new real-estate data set measured operational invisibility

On August 18, 2026, Local Falcon published a study of 37,500 local-intent AI searches across the 100 largest U.S. cities. The company tested ChatGPT, Google AI Overviews, Google AI Mode, Gemini, and Grok from 25 geographic points in each city.1

The results should make every residential-service operator slightly uncomfortable:

In Inman’s report on the study, Local Falcon CEO David Hunter explained the mechanism plainly:2

“The models aren’t reading MLS data or production rankings; they’re reading pages.”

David Hunter, Local Falcon

This is not merely an answer-engine-optimization problem. It is an operating problem. The business may possess excellent private evidence, including production reports, customer outcomes, license data, service history, and local expertise. An answer engine can cite only what it can retrieve, understand, and trust.

Some valuable facts are becoming less accessible, not more. In a public post published within the past day, real-estate analyst Mike DelPrete described a new Hemnet product as “pre-market listings teased publicly but gated behind a login.”3 That product may be perfectly legitimate. It also illustrates the retrieval boundary: a general-purpose model cannot reliably use evidence hidden behind authentication, private networks, or inconsistent data permissions.

Your discovery problem therefore has two parts. Publish verifiable evidence that should be public. Keep restricted evidence private. An agent needs to know the difference.

Voice has a different trust network

On August 20, HousingWire reported that New American Funding was expanding its use of AI agents across mortgage customer interactions. The announced system uses defined procedures for complex, multistep work and lets the lender test and refine each solution before launch.4

That is what workflow governance should look like. Yet a perfectly governed voice workflow can still fail before the first word.

Somos, the administrator of toll-free numbering in North America, explains that call-analytics providers and carriers consider signals such as complaint rates, excessive calling frequency, inconsistent caller IDs, and poor list hygiene when assigning unwanted-call labels.5

STIR/SHAKEN helps, but its job is narrower than many operators assume. The Federal Communications Commission says the framework digitally validates that a call came from the displayed number as it moves through supported IP networks. It does not prove that the caller is who they claim to be, and it does not guarantee that the call is legitimate.6

Authentication answers one question:

Did this call originate from the displayed number?

It does not answer four others:

An AI voice agent needs all five answers before it dials at scale.

Messaging has two permissions, not one

Application-to-Person 10DLC is the carrier registration system for application-generated SMS and MMS sent to U.S. recipients from ten-digit local numbers.

Current Twilio A2P 10DLC documentation describes two core records:7

A2P registration does not create recipient consent. Carrier registration identifies the sender and declared traffic. Your consent ledger determines whether this person may receive this message on this channel.

Those are separate checks. Both must pass.

A common failure is purpose drift. A number is registered for appointment reminders. Six months later, an enthusiastic workflow uses it for cold-database reactivation. The number exists. The campaign exists. The message does not match the campaign that carriers evaluated.

The Last-Mile Oppy catches that mismatch before the send.

Build the Last-Mile Oppy

The Last-Mile Oppy is not a marketer, dialer, lawyer, or carrier administrator. It is a preflight and observability agent. It reads source records, runs synthetic checks, returns a state, and opens an exception for the correct human owner.

BCG calls pre-governed, production-ready templates golden paths. In its August guide to governing agents at scale, BCG argues that identity, registration, monitoring, and policy enforcement should be wired into the starting point rather than added after launch.8

The Last-Mile Oppy is a golden path for customer communication.

The three probes

Give it a strict input contract

Do not ask the agent to “check whether this campaign looks okay.” That sentence is how policies become suggestions.

Use a structured request:

{
  "proposed_action": {
    "channel": "sms",
    "source_number": "+15550100",
    "recipient_id": "84a9b2",
    "intent_category": "reactivation",
    "draft_payload": "Are you still looking for a home downtown?",
    "requested_send_time": "2026-08-21T14:30:00-04:00"
  },
  "business_context": {
    "brand_id": "brand_14",
    "location_id": "office_boston_02",
    "workflow_owner": "growth_operations",
    "human_reviewer": "compliance_queue_east"
  }
}

AWS recently recommended adding agent retrieval where governed data already lives rather than duplicating everything into a new store.9 Follow that principle. Let the Last-Mile Oppy read authoritative systems through narrow tools. Do not create another shadow consent database with a cheerful dashboard.

Use a three-state response

{
  "state": "BLOCKED",
  "probe": "messaging",
  "reason_code": "CAMPAIGN_PURPOSE_MISMATCH",
  "evidence": {
    "registered_use_case": "appointment_reminder",
    "proposed_intent": "reactivation",
    "recipient_permission": "granted"
  },
  "required_action": "Route to messaging administrator for an approved campaign mapping.",
  "owner": "growth_operations",
  "expires_at": "2026-08-22T14:30:00-04:00"
}

PASS means the action may proceed under the current evidence. DEGRADED means delivery or trust may be impaired, but the agent has not identified a hard prohibition. BLOCKED means the action must not execute until a human resolves a specific defect.

The expiration matters. Carrier labels, consent, campaign status, and public search results change. Yesterday’s PASS is not a lifetime achievement award.

Copy this prompt into the agent

LAST-MILE PROTOCOL: BLOCKING

You are the Last-Mile Oppy. You audit proposed customer-facing actions
against discovery, voice, and messaging edge constraints.

You do not publish content, rotate phone numbers, alter carrier
registrations, place calls, send messages, or make legal conclusions.
You return PASS, DEGRADED, or BLOCKED and assign exceptions.

FOR DISCOVERY
1. Run the approved local query set across configured answer engines.
2. Record whether the business is named and whether a source is cited.
3. Compare cited claims with the approved public-evidence registry.
4. BLOCK any recommendation to publish unsupported superlatives.
5. DEGRADED means evidence is accurate but missing, stale, or uncited.

FOR VOICE
1. Verify the source number is active and assigned to this business unit.
2. Verify available caller-ID authentication evidence.
3. Check known carrier and analytics labels.
4. Compare call frequency, complaint trend, and purpose with policy limits.
5. BLOCK if the number is unauthorized or the recipient is suppressed.
6. DEGRADED if reputation or authentication evidence is incomplete.
7. Never rotate a number to evade a label. Open remediation instead.

FOR MESSAGING
1. Verify the source number maps to an active A2P Brand and Campaign.
2. Compare proposed intent with the registered Campaign use case.
3. Verify recipient permission for this channel and purpose.
4. Verify no revocation or suppression applies.
5. BLOCK on any purpose mismatch, missing permission, revocation, or
   suppression record.

For every DEGRADED or BLOCKED result, return the exact evidence,
reason code, human owner, required action, and expiration time.
When evidence conflicts, choose the more restrictive state and escalate.

Your counsel should approve the consent logic, quiet-hour rules, disclosure language, and escalation policy. The agent’s job is to apply the approved map consistently, not to invent law between tool calls.

Launch it in seven days

Google’s new AI Max testing tools preserve brand and location controls inside campaign experiments.10 The lesson travels well: experimentation is useful only when the guardrails survive the test.

Adapt the job, not the principle

The same Oppy can serve residential-service businesses with different evidence and escalation owners.

The Last-Mile Oppy never decides who deserves a loan, a home, insurance, legal representation, or medical treatment. It checks whether an already approved communication can travel through the intended channel with the correct identity, evidence, and permission.

The scorecard leaders should see

Do not reward the Last-Mile Oppy for the number of actions it blocks. Reward it for defects found, defects repaired, and clean actions released.

Intelligence ends where delivery fails

Real-estate AI is moving from chat toward execution. Mortgage agents are handling multistep service. Listings may live behind private networks. Identity checks are arriving at closings. Answer engines are becoming referral surfaces. Phone and messaging networks are enforcing their own definitions of trustworthy behavior.4 11

This is not a reason to slow down. It is a reason to finish the system.

A customer-facing agent needs reasoning, tools, memory, permissions, and a reliable path to the customer. The first four are impressive in a demo. The fifth is what makes the demo a business.

Build the Last-Mile Oppy before your next outbound agent goes live. Let it ask the unfashionable questions:

Can they find us? Will the network carry us? Are we using the channel exactly as promised?

If the answer is no, better to learn it from your own agent than from a carrier, a regulator, or a customer who never knew you called.

References

1 Local Falcon, “The Realtor AI Visibility Index,” August 18, 2026.
2 Inman, “When you ask AI for the best agent, top producers often lose out,” August 20, 2026.
3 Mike DelPrete, public LinkedIn post on Hemnet’s Underhand network, August 20, 2026.
4 HousingWire, “New American Funding partners with Kastle to deploy AI agents,” August 20, 2026.
5 Somos, “How to Keep Your Business Calls from Being Labeled ‘Potential Spam’,” August 19, 2026.
6 Federal Communications Commission, “Combating Spoofed Robocalls with Caller ID Authentication.”
7 Twilio, “Programmable Messaging and A2P 10DLC.”
8 Boston Consulting Group, “Enterprise AI Control Plane: The CIO’s Guide to Governing and Accelerating AI Agents,” August 14, 2026.
9 Amazon Web Services, “AWS vector solutions: Build agentic AI where your data lives,” August 20, 2026.
10 Google, “Make AI Max work for your business with new testing and planning tools,” August 20, 2026.
11 Arizona’s Family, “CLEAR identity verification expands to real estate closings,” August 20, 2026.