AI reconciliation: from matching to investigation

What AI reconciliation means

AI reconciliation is the application of machine learning and, more recently, LLM-based agents to the reconciliation lifecycle: matching records across systems, predicting and classifying breaks, and — the newest and highest-value step — investigating unexplained breaks to a documented root cause. The goal is not to remove the control; it is to remove the archaeology around the control.

Reconciliation is a natural first territory for AI in investment operations because its economics are brutal and measurable. Matching the 95%+ of records that agree is a solved, deterministic problem; the residual breaks are where expert hours disappear — an analyst pulling transaction history, prices, corporate actions and counterparty data to explain a difference before anyone can decide what to do about it.

Two generations of AI in reconciliation

The first generation, in production across the industry for years, is machine-learning matching: models that learn matching rules from historical data, propose rule candidates, tolerate fuzzy fields, and flag anomalous patterns. This generation compresses setup time and raises straight-through match rates, and it is table stakes in modern reconciliation platforms.

The second generation is agentic investigation. Instead of a model scoring record pairs, an agent receives one break as a typed task and works it the way an analyst would: reading the break detail, querying match history, checking price movements over a trailing window, testing root-cause hypotheses, and returning a conclusion with the evidence attached. The difference in kind: generation one decides whether records match; generation two explains why they do not.

Anatomy of an agentic break investigation

A well-designed investigation agent is narrow, tool-driven and structured end to end:

• —Input — one break, typed: identifiers, amounts, sources, ageing, account context.
• —Tools — declared, read-oriented capabilities: break detail, historical match patterns, pricing history, reference data. The agent can only call what its descriptor permits.
• —Reasoning — hypothesis testing across known root-cause classes (for example timing differences, price-source divergence, missing transactions, static-data mismatches, corporate-action effects, booking errors).
• —Output — a schema-validated result: root-cause class, confidence score, supporting evidence with links to source records, and a recommended action (auto-correct, manual review, or escalate).
• —Trace — every model and tool call recorded, so the investigation can be replayed step by step in an audit.

What changes for the operations team

The operating model shifts from investigate-then-decide to review-then-decide. Breaks arrive pre-worked: the analyst opens an exception that already carries a probable cause, the evidence trail, and a proposed resolution, and spends judgment where it matters — confirming, overriding, escalating. Investigation effort that took tens of minutes per break compresses toward minutes of review, and the queue drains in priority order rather than first-in-first-out.

Quality control changes character too. Manual investigation quality is invisible — it lives in an analyst's head and a free-text comment. Agentic investigation is measurable: each agent version is scored against a curated dataset of historical breaks with known causes, and accuracy becomes a baseline that releases must not regress below. Combined with human review of every proposed action, this typically gives operations leaders more verifiable control than the manual baseline, not less.

Market context and how to evaluate claims

Reconciliation is among the loudest categories in agentic AI marketing. Clearwater Analytics' November 2025 announcement reported a 90% reduction in manual reconciliation effort among the outcomes of its GenAI agent rollout; specialist vendors and platform incumbents alike market AI-assisted matching and investigation. These figures are vendor-reported and unaudited — which does not make them false, but does make them non-transferable: break-mix, data quality and existing match rates dominate the achievable delta at any given firm.

A sturdier evaluation runs on your own data and audit standards:

• —Accuracy on your breaks — root-cause accuracy on a sample of your historical breaks with known causes, not a vendor demo set.
• —Evidence quality — does every conclusion link to verifiable source records?
• —Auditability — can compliance replay a full investigation months later, including the agent version that ran?
• —Workflow fit — do typed outputs route into your existing exception workflow, or does the tool assume you adopt a new one?
• —Economics — cost per investigated break against your measured manual minutes.