What exactly is controlled?

The risk of an error
in a selected original extraction.

An invoice counts as an error if one or more required critical fields are wrong or missing relative to the source. The example covers supplier identity, invoice number, PO reference, currency, VAT total and gross total.

The target population is original extractions eligible to skip capture checking, before audits or corrections. A guarantee needs a statistical procedure that accounts for policy selection and meets its sampling assumptions. The observed error rate alone is insufficient.

The limit is a fraction of invoices with the defined error. It is not a bound on monetary loss, a promise about each invoice or an approval to pay.

Scope & conditions ↗︎

What the demo shows

A review workflow you can inspect.

The documents, scores and policy outcomes are authored. No invoice-specific statistical bound or measured time saving is computed by this walkthrough.

1,000

In-scope invoices

The prepared operating example uses control-cleared invoices.

415

Capture checks retained

350 retained checks plus 65 audits. Audits are already included in this total.

585

Capture checks avoided

Compared with checking every extraction in this example. This is not a measured time or cost saving.

The chosen tolerance is 1%. The prepared 4 errors among 650 eligible original extractions concern the population before audits; they do not estimate the error rate among the final 585 unchecked invoices.

Which numbers are real, prepared or still missing?
8 interactive practice invoices
Fictional documents and extractions. Your review choices exercise the workflow; they do not fit a scorer or produce a statistical bound.
2,400 pairs / 1,600 grouped situations
An explicitly loaded, authored warm-up cohort. The alternative insufficient-evidence branch has 280 pairs / 180 groups. Neither is a recommended evidence requirement.
1,000-invoice operating example
Separate prepared counts for workload and errors. These are not independent measured test results or a replay of the practice labels.
Upper bound & confidence setting
Authored policy fields, not computed statistics. The three budget profiles are illustrations, not a measured risk–workload curve.
Not yet established here
Invoice scorer performance, a validated statistical procedure for this population, a matched baseline comparison, post-audit error rates and measured net time or cost savings.

Your own evidence

Establish the limit.
Compare the work it requires.

  1. 01 / Define

    Agree the error and population.

    Choose critical fields and tolerances, source references, invoice types and operating scope. Identify the manual task and baseline to compare.

  2. 02 / Prepare

    Review original outputs.

    Keep source documents, uncorrected extractions and versions. Account for supplier/template dependence and sampling history. Include correct cases and errors.

  3. 03 / Assess

    Calibrate and test independently.

    Assess scorer suitability, separate development, calibration and evaluation, and use a procedure that supports the selected population. Compare current controls, the score alone and the policy on held-out invoices.

  4. 04 / Decide

    Measure operational usefulness.

    Inspect retained checks, audits and net work removed. Include scoring, monitoring and integration costs before deciding whether to run a monitored pilot.

See evaluation inputs and deliverables ↗︎

Beyond a confidence flag

Does the score help distinguish
the mistakes you care about?

A score ranks extractions. The calibrated policy establishes a supported risk limit. Both usefulness and statistical validity need to be assessed for the invoice population.

Who supplies the score—and what must it prove?

The proposed EyeTrustAI role

Prepare or assess a scorer for your source documents, extracted fields and error definition, then evaluate the risk-control policy. A scorer ranks original extractions by risk; it is not itself an error guarantee, and raw extractor confidence is not automatically a suitable substitute.

In this walkthrough

invoice-fidelity@0.3-demo and its scores are authored fixtures. No fidelity model runs here. The interface does not establish a production-ready invoice scorer or its price.

Before deployment

Test whether the score distinguishes incorrect extractions—including confidently wrong ones—from correct ones. Compare the deployment’s existing controls, the score alone and the calibrated policy on a held-out evaluation. A score that cannot separate cases may offer too little review reduction, even if a conservative policy can control risk.

Scorer choice, hosting, per-invoice cost and throughput belong in the scoped evaluation. A suitable scorer may need adaptation or development; that work and its cost must be agreed.

Use the evidence you already have

Existing reviews may help.
The original history matters.

Source invoices paired with original extractions, explicit outcomes and sampling history can be assessed for reuse. Corrected final exports alone do not show whether the original extraction was right.

Can we reuse reviewed invoices or extend the scope?

Reuse evidence when its history is intact

We can assess existing reviewed examples if they retain the source invoice, original extraction and versions, explicit correct/error outcomes and sampling history. Corrected final records alone do not show whether the original was correct; uncorrected records are not automatically clean. Check missing cases, supplier/template dependence and whether the data still represents current work.

Keep evaluation data separate from scorer fitting and policy selection. Reuse is a deliberate assessment, not automatic carry-over. This demo archives prior rounds without counting their labels as new evidence.

Start with the work actually being checked

This example covers six header fields on English, EUR invoices from known suppliers for one entity, up to €25,000. Credit notes, line-item coding and other languages are outside its illustrated scope. Adding German invoices, new entities or layouts requires assessing that population and its evidence; it is not a settings-only extension.

For structured e-invoices, first establish whether extraction checking is needed at all. Benefits depend on the actual manual task, not on calling every document an invoice.

Size the effort from your evidence

Agree the target risk, supplier/template mix, review quality and amount of reusable evidence before estimating new labels or duration. Smaller accounts may not have enough evidence or volume to justify a separate policy; pooling requires a defensible shared population. There is no universal label count or deployment time.

The operating value

Count the work
that actually disappears.

If an invoice must still be opened for coding, matching or approval, count only the capture-checking time removed. Compare against the process already in use, including retained checks and audits.

Do fewer checks actually save time?

First identify the task that disappears. If someone must open the same invoice for coding, matching or approval, only the extraction-checking time that is actually removed counts as saved.

Ongoing work

Retained capture checks, random audits, adjudication, monitoring and incident investigation. Compare the time each task takes; a difficult exception and a random audit may take different amounts of time.

Setup & reassessment

Pairing source documents with original extractions, reference labeling, scorer assessment, integration and each new evidence round. Track this effort separately and state the period over which it is recovered.

For this example:
1,000 × minutes per baseline capture check
− non-audit capture checks × minutes per retained check
− audits × minutes per audit
− additional recurring staff time
= net recurring staff minutes saved

Use the counts for the selected policy above. The audit checks are already included in retained total checks; do not subtract them twice. Measure task times and overhead rather than assuming a standard number of seconds per invoice.

Scoring, hosting and integration charges belong in a separate cost comparison. No staffing reduction, monetary saving or payback period is measured in this demo.

Plan the integration

Agree the interfaces,
fallback and owners.

Source access, original-output identity, review routing and data handling are part of the evaluation. The demo does not establish a live API, ERP integration or deployment arrangement.

What would integration involve?

Proposed integration · no live invoice API in this demo

Keep your extraction pipeline and accounting integrations. Assess original extractions to establish where capture checks can be reduced.

  1. PairSource + original extraction + versions
  2. AssessScorer + policy + declared scope
  3. RecordReview or capture-handoff receipt
Inputs, outputs and retry behaviour

Input identity

Use stable document and extraction IDs, immutable source and original fields, extractor/schema/scorer versions, entity and scope metadata, and required control flags. Preserve corrected working records separately.

Output contract to agree

A version-bound assessment and decision receipt should identify the source, original extraction, policy, reason and audit selection. Label events should record the reviewer, outcome, source reference and relevant versions. They do not approve, post or pay an invoice.

Retries and reprocessing

Retries of the same request must not create duplicate handoffs or redraw audit selection. Agree idempotency and deduplication across EyeTrustAI and your queue. A changed source or re-extraction creates a new version and assessment; never attach an old receipt to new fields.

Fallback, data location and operating requirements

Retain capture review on timeout, unavailable scoring, missing input, incompatible versions or a suspended policy. Existing business holds remain authoritative. Agree timeouts, retry limits, capacity and outage handling before connecting a production workflow.

Agree where scoring runs and what data moves. Specify permitted source access, hosting, retention/deletion, credentials and logs. This browser-only demonstration neither proves an on-premises deployment nor promises that a production service keeps all data on your system.

Updates need an explicit review. Model, scorer, schema or declared-scope changes require checking evidence compatibility before reuse. The same versions can also encounter different data; audit outcomes and population monitoring remain necessary.

Who does what?

EyeTrustAI

Assess the scorer and evidence. Evaluate error risk and checking effort, and define the integration plan.

Your product team

Own integration, extraction versions, product controls and monitoring. Connect the review workflow and implement fallback.

Your customers’ finance teams

Agree critical fields and acceptable risk. Provide qualified reviews and audits; retain matching, approval and payment authority.

Agree these roles, source access and owners for suspension, impact review and recovery. This is a proposed evaluation scope.

Evaluate your invoice workflow.

Bring one scope, your current controls and the evidence available.

Discuss an invoice evaluation ↗︎