AI Reconciliation Exception Management: A Handoff Card

AI reconciliation changes the first-pass investigation, not the firm’s accountability for an open break. A useful operating question is therefore not “did the agent clear it?” but “who accepted the remaining risk, what protection is in place, and what evidence permits closure?”

That question is timely for investment operations. On September 1, 2026, Ridgeline announced that Cabot Wealth Management had gone live with its AI-assisted managed-service offering after its sole reconciliation specialist left. The announcement describes daily operations being handled as a service; it does not make a provider’s status update sufficient evidence that an investment firm should close an exception. Ridgeline’s product description likewise frames reconciliation as a workflow supported by AI.

This guide provides a reusable handoff card, a three-party RACI, a state model, and a copyable prompt for a firm that keeps the final acceptance decision with an identified human owner. It complements a data provenance contract for wealth-management AI: provenance establishes whether an input is fit for use; this card establishes how an open operational exception is transferred, contained, and closed.

Why a matched record is not enough

Position, cash, transaction, and corporate-action breaks have different operational consequences. A tentative match may be useful for triage, yet the firm still needs to know whether the break affects a client report, a trade restriction, a valuation, a cash movement, or a downstream ledger. A managed-service provider can investigate the record; it cannot leave the firm’s ownership ambiguous.

Existing reconciliation managed-services pages emphasize processing exceptions, workflow visibility, and operational scale. For example, Broadridge describes reconciliation and exception processing across asset classes, while IVP describes managed reconciliation for positions, activity, and cash balances. The buyer-side gap is the acceptance test: a compact record that says what is still unresolved, who owns each next action, and what must be attached before the state can change.

Treat the handoff as a controlled state change

Do not let a language model set a break to resolved, create an accounting entry, release a trade, or alter client-facing data. Its role is to organize supplied evidence, classify the break, and draft the handoff. A person with the documented authority changes the operational state.

Use five states:

State	Meaning	Required human action
OPEN	A deterministic control or reviewer found a break.	Assign a firm owner and decide whether a temporary protection is needed.
CONTAINED	A documented hold, disclosure block, or other local safeguard limits use of the affected data.	Approve the safeguard and its review time.
INVESTIGATING	Provider or firm is collecting evidence and proposing a cause.	Review the scope, owner, and target time.
ACCEPTED_CLOSED	The firm accepted a documented resolution and retained supporting evidence.	Named approver signs the closure.
REOPENED	New evidence invalidates the earlier conclusion or the break recurs.	Reassign the owner and preserve the earlier closure record.

ACCEPTED_CLOSED is deliberately stronger than “matched” or “provider closed.” If the firm owner, evidence reference, or approval time is missing, the record remains open. This separation mirrors a broader systems rule: an asynchronous success signal is not necessarily a completed financial outcome. See Financial Systems: Ledger, Idempotency & Reconciliation for the engineering controls behind that distinction.

The reconciliation exception handoff card

Create one immutable card per exception and link it to the underlying records rather than copying sensitive transaction data into a general-purpose prompt.

Field	What to record	Closure gate
Exception identity	Stable exception ID, business date, account or portfolio reference, and source-record links	ID and record links present
Break classification	Data, timing, corporate action, suspected unauthorized change, or workflow issue	Classifier gives a reason and confidence; human confirms material cases
Impact	Affected process, potential client or accounting impact, and restricted downstream uses	Firm owner accepts the stated scope
Temporary protection	Hold, review queue, report block, or other approved local control	Owner and next-review time present
Responsibility split	Firm owner, service-provider owner, escalation contact, and decision authority	Every action has one accountable owner
Evidence	Source snapshots, matching logic or rule result, investigation notes, and approval link	Evidence references are retrievable
Resolution	Proposed cause, correction reference if any, validation result, and reopen trigger	Human approval with timestamp

The card is not a substitute for the firm’s supervisory procedures, vendor agreement, books and records, or compliance review. It is the small operating record that lets those systems connect to a daily exception without relying on memory or a chat transcript.

Assign responsibility before an exception arrives

The provider can be responsible for investigation without being accountable for the firm’s risk acceptance. Make the distinction explicit.

Activity	AI workflow	Managed-service provider	Investment firm
Detect and group supplied breaks	Assists	Responsible	Accountable for control design
Gather supplied evidence and draft classification	Assists	Responsible	Consulted
Apply a data-use restriction or operational hold	No authority	Recommends	Accountable and approves
Correct a record or release a workflow	No authority	Performs only under agreed authority	Accountable for approval and verification
Accept closure or reopen	No authority	Consulted	Accountable; named approver acts

If a contract or local procedure gives the provider a different role, change the table and make the boundary testable. Do not infer authority from the fact that a service is outsourced. The same discipline helps keep agent access scoped in financial AI analysis controls.

A copyable AI prompt for the handoff

Use this with approved, minimized input data. Give the model no credentials that can change records and no instruction path that can override the schema.

You are preparing a reconciliation exception handoff card.

You may summarize only the supplied records. You may not mark an item resolved,
change accounting data, release a transaction, or invent evidence.

For each exception, return valid JSON with these fields:
- exception_id, business_date, affected_record_refs
- classification: one of DATA | TIMING | CORPORATE_ACTION |
  SUSPECTED_UNAUTHORIZED_CHANGE | WORKFLOW_FAILURE | UNKNOWN
- classification_reason and confidence: HIGH | MEDIUM | LOW
- possible_impact and restricted_downstream_uses
- temporary_protection_needed: YES | NO | UNKNOWN; proposed_protection
- firm_owner, provider_owner, escalation_contact, decision_authority
- evidence_refs, missing_evidence, proposed_next_action, target_review_time
- proposed_state: OPEN | CONTAINED | INVESTIGATING | REOPENED
- closure_eligibility: NOT_ELIGIBLE | NEEDS_FIRM_APPROVAL
- reopen_conditions

Rules:
1. If owner, evidence, or decision authority is absent, use UNKNOWN and add it to missing_evidence.
2. Never return ACCEPTED_CLOSED. Only a named firm approver can set that state outside this prompt.
3. If impact or source freshness is uncertain, recommend a review or restriction; do not assume there is no impact.
4. Cite supplied record references for every factual statement. Separate facts from hypotheses.

The closure_eligibility field forces a productive limit: the AI may show that a card is ready for review, but it may not complete the review. For adversarial tests of agent behavior rather than day-to-day operations, use the separate AI red-team test harness for banking agents.

Run an eight-question morning escalation

At the daily control meeting, ask:

1. Which exceptions could affect a client-facing report, trade, valuation, cash movement, or accounting output?
2. Does each open card have one firm owner and one provider owner?
3. What protection is active until the next review?
4. Which source record or timestamp is missing, stale, or contradictory?
5. Which classifications are only hypotheses rather than evidenced conclusions?
6. Is the provider’s proposed fix within its agreed authority?
7. What evidence will let the firm approver accept closure?
8. What event automatically reopens the card?

Track the count and aging of open cards, the share with a named firm owner, time from detection to containment, closure cards returned for missing evidence, and reopen rate. These are operational signals, not performance promises; compare them to the firm’s own risk appetite and procedures.

Make the evidence reviewable later

Preserve the source snapshot or identifier, the rule or comparison that raised the break, the AI input/output version, the provider’s investigation, the safeguard decision, and the human approval. Version the prompt and the allowed classification list so a later reviewer can tell what the system was permitted to do at the time.

This is especially important when AI is used in a provider-operated workflow: the artifact should make it possible to distinguish detection, investigation, correction, and acceptance. A clean dashboard alone does not establish any of those facts.

Sources

• Ridgeline: Cabot Wealth Management becomes first customer to leverage Intelligent Outcomes
• Ridgeline Intelligence: AI workflows and reconciliation
• Broadridge: Reconciliation managed services
• Indus Valley Partners: Reconciliation as a service for investment funds