Your AI Should Not Ask for the Same Document Twice: Build a Reconciliation Oppy
By Anna with Oppy
A practical operating guide for brokerages, mortgage teams, title offices, insurers, property managers, transaction coordinators, and other service businesses.
A customer sends the same document twice for a reason that is usually invisible to them. One system received it. Another system never learned that it existed. A third system still triggered a reminder.
Adding a general-purpose AI assistant can accelerate that confusion. It can draft the second request faster, personalize it beautifully, and send it at scale. This is not progress. It is clerical slapstick with excellent grammar.
Richard A. Dixon described the larger property-technology problem neatly in a September 9 commentary:
“Technology has digitised the duplication rather than removed it.” 1
The better first AI employee is not an autonomous closer. It is a Reconciliation Oppy that reads only approved sources, determines where records disagree, attaches the evidence, and routes the next step to an accountable person.
Its first job is not to act. Its first job is to prove what happened.
The idea in 30 seconds
A Reconciliation Oppy watches one narrow handoff across two or three approved systems. It creates a compact ledger showing whether the handoff is matched, missing, conflicting, duplicated, stale, or uncertain. It then prepares an exception packet for a human reviewer.
It does not decide whether a borrower qualifies, a policy is adequate, a contract means what someone hopes it means, a patient needs treatment, or a legal filing is sufficient. It reconciles administrative evidence. Authorized people make professional decisions.
| The Reconciliation Oppy may | The Reconciliation Oppy may not |
|---|---|
| Read records the assigned user is allowed to access | Search every connected system because access happens to be technically possible |
| Compare case IDs, document IDs, timestamps, and approved status fields | Infer that two similar people, properties, or matters are the same case |
| Flag missing, conflicting, duplicate, or stale state | Resolve a material conflict by choosing the most convenient answer |
| Draft an internal exception packet | Send, file, bind, approve, deny, schedule, or update a system without the defined gate |
| Recommend one narrow next administrative action | Make a regulated or professional determination |
This is an operating design that a team can configure with Oppy. It is not a claim that a prebuilt Reconciliation Oppy template exists, and it is not a promise of a particular return.
Why reconciliation is suddenly the useful AI problem
Dixon’s commentary describes a housing transaction still fragmented among agents, lenders, lawyers, surveyors, moving companies, and connected digital tools. 1 Connection did not create agreement.
A September 9 RISMedia commentary by Kendra O’Brien of FBS describes teams that have led roughly 100 multiple listing service conversions since 2013. One conversion found 67 spelling variations for a frequently used county in one open-text field. 2 That is not an exotic model problem. It is what ordinary operations look like after years of human entry, migrations, integrations, and local habits.
“Accumulated data and rules must move from one home to another, and each step of that process is important.” 2
Current AI product work is also moving from answer generation toward access-aware action. On September 9, Amazon Web Services announced two application programming interfaces for its managed knowledge base: one checks whether a specific user may access a given ingested document, and another returns the document’s access-control list for audit. 3
The adjacent documentation supplies the more important lesson. AWS says its Confluence Data Center connector does not support document-level access-control lists. Authenticated knowledge-base users can see all content crawled through that connector. 4
The point is not that one connector is good or bad. The point is that searchability and permission are different facts. An AI employee needs evidence for both.
Build one Reconciliation Oppy in five moves
1. Pick one expensive handoff
Do not begin with “fix our operations.” Pick one event that customers or staff can recognize.
Good starting points include:
| Handoff | Observable failure |
|---|---|
| Document receipt | A customer is asked for evidence already present in an approved repository |
| Consent capture | Outreach is prepared although the consent record is missing, stale, or attached to another campaign |
| Appointment confirmation | Calendar status and the service system disagree |
| Transaction task | A task is marked complete without the required closure evidence |
| Referral | The receiving team never acknowledges the case, but the sending team assumes ownership transferred |
| Record update | A newer value exists in email while the designated source of record remains unchanged |
The handoff must have a stable case unit, two or three bounded sources, a recoverable failure mode, and one accountable human owner. If the team cannot name those four things, it is not ready to write the prompt.
2. Define the case key and source hierarchy
The case key is the identity of the matter being reconciled. It might be a transaction ID, loan file ID, policy service request, property-management ticket, patient referral ID, or legal matter number.
Never ask the Oppy to join cases using a name alone. Names change, repeat, and arrive misspelled. Use a durable identifier. If no single identifier crosses systems, document an approved composite rule, such as transaction ID plus property address plus customer ID. Route collisions to review.
Then name the source of record for each fact. Email may prove that a document arrived. It may not control the transaction status. A document repository may hold the file. It may not prove that an authorized person reviewed it. The newest timestamp is not automatically the highest authority.
A simple source hierarchy looks like this:
CASE IDENTITY: Transaction platform
DOCUMENT PRESENCE: Approved document repository
REVIEW STATUS: Workflow system
COMMUNICATION CONSENT: Consent ledger
APPOINTMENT TIME: Scheduling system
This small list prevents the Oppy from treating every connected source as equally authoritative.
3. Create the handoff ledger
The output is not a chat transcript. It is a row that another person can inspect.
| Ledger field | Required content |
|---|---|
case_key |
Durable case ID and any approved matching attributes |
handoff_type |
Exact event being checked |
source_of_record |
System that controls the relevant status |
evidence_pointer |
Document ID, message ID, record URL, or other approved reference |
observed_at |
Time the Oppy read the record |
source_updated_at |
Time the source last changed it |
permission_basis |
Role, rule, or authorization that allowed retrieval |
reconciliation_status |
matched, missing, conflicting, duplicate, stale, or needs_human_review |
suggested_next_action |
One narrow, reversible proposal or no_action |
action_owner |
Person or team responsible for resolution |
approval_gate |
State required before any external action |
idempotency_key |
Unique key that prevents retries from duplicating a task, message, or update |
closure_evidence |
Proof that the handoff actually finished |
The idempotency_key deserves more attention than its name receives. It tells a connected system, “If you see this exact intended action again, do not create another copy.” Without it, a retry can become two reminders, two tasks, or two updates. Computers are admirably persistent. Customers use a different adjective.
4. Give the Oppy a stop-first prompt
The first prompt should optimize for a correct stop, not maximum activity.
ROLE
You are the Reconciliation Oppy for [one named handoff].
SCOPE
Read only these approved sources: [systems, objects, and fields].
Use [case key or approved composite matching rule].
Treat [named system] as the source of record for [named status].
PERMISSION RULE
Before using evidence, confirm that the assigned role may access the source.
If permission cannot be confirmed, set status to needs_human_review.
Do not quote, summarize, or expose inaccessible content.
RECONCILIATION RULES
Return exactly one status:
matched, missing, conflicting, duplicate, stale, or needs_human_review.
Never resolve identity collisions or material conflicts by inference.
Never treat a recent timestamp as authority unless the source hierarchy permits it.
OUTPUT
Produce the handoff ledger fields and an exception packet.
Every conclusion must include an evidence pointer and source timestamp.
ACTION BOUNDARY
You may read, compare, classify, and draft an internal next action.
You may not contact a customer, change a source system, schedule, submit, file,
approve, deny, bind, quote, recommend, or make a professional determination.
STOP CONDITIONS
Stop when the case key is ambiguous, a required source is unavailable,
permission is unconfirmed, records materially conflict, evidence is stale,
or closure evidence is missing.
Assign the exception to [owner role] with [service-level deadline].
That prompt is intentionally tedious. Tedious rules are often cheaper than exciting incidents.
5. Make the exception packet the default product
When sources disagree, the Oppy should not produce a confident paragraph. It should produce a compact case for review.
CASE: [case_key]
HANDOFF: [handoff_type]
STATUS: [reconciliation_status]
SOURCE OF RECORD: [system, record ID, updated time]
CONFLICT: [one factual sentence]
EVIDENCE: [approved links or IDs]
PERMISSION CHECK: [passed, failed, or unavailable]
SAFE NEXT ACTION: [draft only or no action]
REQUIRED OWNER: [role]
DEADLINE: [time]
STOP REASON: [why the Oppy did not act]
This format transfers context without transferring judgment. A title team can receive proof that a declared document is present, while an authorized reviewer determines its legal significance. An insurance team can receive a packet showing that a declaration page arrived, while a licensed person determines whether coverage is adequate. A mortgage team can reconcile whether the proper credit-report evidence is attached and which model appears in the file, while the lender retains credit and underwriting decisions. Fannie Mae’s September 9 lender letter on broad VantageScore 4.0 availability is a current example of why model consistency, ordered reports, delivery data, and file evidence need controlled workflow implementation. 5
The pattern travels, but the authority does not
The administrative shape is reusable across industries. The professional boundary is not.
| Business | Safe first reconciliation | Decision that stays with an authorized person |
|---|---|---|
| Brokerage or transaction coordination | Confirm that a disclosure, inspection, or signature packet is attached to the correct transaction and awaiting the correct owner | Legal sufficiency, required disclosure advice, negotiation, fair-housing judgment |
| Mortgage | Reconcile document presence, report identity, workflow status, and human-review queue | Eligibility, pricing, credit, underwriting, suitability |
| Title | Match identity, property, commitment, and curative-document references across approved systems | Legal interpretation, insurability, clearance |
| Insurance | Reconcile quote requests, document receipt, policy references, and renewal tasks | Coverage advice, adequacy, binding, cancellation |
| Property management | Match maintenance request, authorization, vendor proof, access notice, and completion evidence | Habitability or legal determinations, protected-class decisions |
| Dental office | Match referral, image, consent, appointment, and chart receipt | Diagnosis, treatment plan, clinical judgment |
| Legal office | Match intake, engagement, evidence receipt, deadline, and filing status | Legal advice, privilege determinations, filing sufficiency |
| Auto service or finance | Match appointment, estimate, proof-of-insurance, and document receipt | Credit, coverage, safety, or repair authorization beyond configured rules |
A case key that works for a property ticket may be unsafe for a patient or legal matter. Each deployment needs its own access rules, retention policy, escalation path, and qualified review.
If the Oppy can message, add a separate communication gate
Reconciliation does not create permission to contact someone.
If a later phase allows a reminder by text or voice, require a separate ledger containing the actual sender, recipient, campaign, purpose, consent language, capture evidence, timestamp, and revocation state. For calls and texts covered by 47 C.F.R. § 64.1200(a)(1) through (3) and (c)(2), current federal rule text recognizes specified STOP-type words as reasonable revocations, requires other reasonably understood revocations to be honored, prohibits an exclusive revocation channel, and requires compliance within a reasonable time, no later than ten business days. 6 Other laws, carrier rules, contracts, or sector-specific requirements may demand faster action.
For United States application-to-person messaging over a 10-digit long code, carrier registration typically includes a brand and campaign describing sender identity, message purpose, and opt-in, opt-out, and help flows. 7 A2P 10DLC registration is not a legal safe harbor. It does not replace analysis of the Telephone Consumer Protection Act, do-not-call rules, state law, consent language, or the actual message flow.
The safe sequence is simple:
- Reconcile the case.
- Verify the communication purpose.
- Verify sender-specific consent and suppression state.
- Draft the message.
- Apply the required human or policy gate.
- Send once with an idempotency key.
- Record delivery, failure, reply, and revocation.
Qualified counsel should review live calling and messaging workflows, including state-specific requirements.
Run a 30-day read-first pilot
Start with one handoff, two or three sources, a small group of authorized users, and no external writes. Establish the baseline before the Oppy begins reading.
| Pilot metric | Definition |
|---|---|
| Duplicate-request rate | Eligible cases in which the customer receives another request for evidence already present in an approved source |
| Reconciliation age | Time from the source update to a matched ledger status, reported at median and 90th percentile |
| Exception rate by cause | Separate counts for missing, conflicting, duplicate, stale, permission failure, and identity ambiguity |
| Human release rate | Proposed actions approved, modified, or rejected by the assigned reviewer |
| Rework after closure | Cases reopened because the handoff was not actually complete |
| Permission failures | Attempts blocked because source or user authorization was missing or unclear |
| Communication exceptions | Blocked sends, missing consent, opt-outs, and campaign mismatches |
Review false matches and false closures case by case. A low exception rate is not automatically good. It can mean clean operations, or it can mean the Oppy is confidently ignoring the mess.
After the pilot, grant exactly one new permission only if the evidence supports it. A reasonable second phase might let the Oppy create an internal task after a reviewer approves the proposed action. Keep the action reversible, logged, narrow, and easy to disable.
What success looks like
The Reconciliation Oppy is working when it can answer seven questions without improvising:
| Question | Evidence required |
|---|---|
| Which case is this? | Stable case key |
| What handoff are we checking? | Named event type |
| Which source controls the status? | Source-of-record rule |
| What evidence did the Oppy use? | Openable pointer and timestamp |
| Was it allowed to use that evidence? | Permission basis |
| Who owns the exception? | Named role and deadline |
| What proves the handoff closed? | Closure evidence |
Do not begin by asking AI to run the customer journey. Begin by asking it to prove what happened at one handoff.
When it can show the case key, source, timestamp, permission basis, conflict, owner, and closure evidence, and when people can correct it, then it has earned the next permission.
That is how an agentic operating system becomes useful. Not another interface. Fewer repeated asks. Cleaner handoffs. Accountable work.
References
- Richard A. Dixon, “What Has 25 Years of PropTech Actually Achieved?”
- RISMedia, “Reflections on Moving: Roughly 100 MLS Conversions and One Carefully Packed Trash Can”
- Amazon Web Services, “Amazon Bedrock Managed Knowledge Base adds APIs and console support for debugging document-level access control”
- Amazon Web Services, “Confluence Data Center”
- Fannie Mae, “Lender Letter LL-2026-06: VantageScore 4.0 Broad Lender Availability”
- Electronic Code of Federal Regulations, “47 C.F.R. § 64.1200: Delivery restrictions”
- Twilio, “Programmable Messaging and A2P 10DLC”