Embedded Data Verification & Reconciliation

Return evidence-backed, policy-checked, and reconciled data to your customers, with every exception explained, not a bare pass or fail flag. Your platform renders the result in its own interface, and Staple supplies the verified values and the reasons behind them, at line-item level, across the related documents that have to agree.

Verification That Comes Back Explainable

Staple returns results per field: the confidence score, whether the value was extracted or inferred, which rule or model produced it, and the reviewer if one intervened. Embed AI contextual verification, business-rule and anomaly checks, and external validation against tax authorities and registries, and your platform decides how to present each discrepancy to its customer.

The Anchor And Variations Model

Reconciliation is configured, not hard-coded, and the configuration mirrors how a finance team actually thinks. You choose an anchor, the primary document that is the source of truth, for example a purchase order, and the target documents expected alongside it, such as invoices, delivery notes, and credit notes. Staple gathers each anchor and its related documents into a matching set. A rule is then made of one or more variations arranged as a waterfall: the system attempts variation one, and if a field it needs is missing or a match fails, it falls back to variation two, and so on. Real-world documents are messy, and the waterfall is what lets a rule keep matching when the cleanest identifier is not present.

Each variation defines three steps. Document matching links incoming documents to the anchor through matching links, for example pairing a purchase-order number to an order reference, and a link can require an exact match or a fuzzy match with an adjustable threshold, so Acme Corp still matches Acme Corporation. Table matching aligns the line-item rows across documents on a key, such as a part number against an item SKU, so comparisons are strictly like for like regardless of row order. Comparison logic then evaluates the aligned data.

Comparison, Formulas, And Tolerances

• Compare a field against the matching field in one or more other documents, with the operator the data type allows, equals and not-equals for text, greater-than and less-than for numbers and dates, and empty checks.

• Build header-level formulas across the whole set, for example checking that invoice grand total minus a credit note amount equals the purchase-order total, so netted balances reconcile correctly rather than failing a naive one-to-one match.

• Allow a tolerance so a near match still passes: an absolute value such as fifty cents, a percentage such as one percent, a from-to range, or a date buffer measured in days or months, which absorbs rounding and timing noise.

• Absorb cosmetic noise such as extra spaces or case differences, so a difference that does not matter is not reported as a discrepancy.

Every Exception Is Labeled And Explained

An outcome is not just pass or fail. A rule assigns a label to a specific result, for example Over-billed Quantity when invoice quantity exceeds the purchase order, and a label can be marked critical so the exception is surfaced and filterable. The API returns which line failed, against which source, under which rule, and why, so your platform can route it, show it, or hold it. A document that cannot be matched yet is held rather than force-fed onward: when a missing credit note finally arrives, the set is re-evaluated, and a reviewer can force-reconcile a residual difference or reject it with a reason, each action recorded.

Reconciled Data Carries Its Proof

What your platform receives is not raw extraction, it is data that has already survived matching and comparison, and it can travel with its evidence attached. Every reconciled value can carry a Metastructured Data record, a signed proof of source, timestamp, and confidence, so a downstream system or an auditor can confirm the result was not altered after Staple produced it. That is the difference between handing your customer a number and handing them a number they can defend. Building this in-house means owning the matching engine, the tolerance logic, and the audit layer indefinitely; embedding it is one integration.

Extracted, Inferred, Mapped, Set

Staple separates extracted, inferred, mapped, and set values, so a platform embedding this can never present an inference to its customer as a fact read off the page. Inferred fields carry no confidence score by design.

Proven In Production

A global electronics manufacturer runs three-way matching at 99 percent field-level accuracy at high volume. A global investment bank reconciles values, quantities, and line items in real time across more than five countries. Across deployments, Staple runs for Fortune 100 companies and global financial institutions in 60 countries at a 100% deployment success rate.

Scope

Staple returns verified data and explained exceptions. Your platform decides what to do with them; Staple does not make the write-off, dispute, or payment decision on your behalf.

Discuss An Embedded Use Case

Bring a reconciliation case and we will show the exact API response, or view the API documentation.