In-depth analysis · About 5 min read
Foundry's document-confidence guide separates review priority from permission to automate
Microsoft's October 14 account explains grounding, field-level calibration and why 41% fewer targeted values is not 41% less staff time.
- Source published
- Published here
- Source checked
Microsoft's Content Understanding guide uses an insurance workflow to explain how extracted fields can be linked to evidence and prioritized for review. The October 14 article reports useful customer results, but its most important condition is easy to miss: people still checked every field. Confidence changed the order of attention, not permission to process sensitive insurance data without people.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- The overall customer evaluation covered more than 22,000 field values across more than 1,000 production documents. On a separate held-out set, targeted review selected 41% fewer values while still including more than 80% of known extraction errors.
- Confidence is produced by a separate shared model, not an LLM saying how certain it feels. Customer calibration interprets scores against checked documents; it does not retrain that model.
- A separate public lab uses 1,000 CORD v2 receipts, split 800 for calibration and 200 for evaluation. Its illustrative results are not the insurance customer's outcomes.
Evidence and confidence answer different questions
Grounding connects a structured value back to supporting document content, helping a reviewer inspect it and maintain an audit trail. It makes the claim checkable, not automatically true. Confidence is a learned field-level signal between 0 and 1, produced by a separate model using document evidence, field context and extraction diagnostics. It is not the language model's self-reported certainty. Microsoft reports up to 28% lower average inference-token usage from deeper grounding integration and up to 14% improvement in confidence quality measured by AUROC. The latter is not a claim of 14 percentage points more extraction accuracy. These service figures are vendor reports, not our measurements.
Calibration adapts decisions, not the shared model
Microsoft says the confidence model is trained on Microsoft-curated data, not customer data, and shared across customers. A customer's checked examples instead reveal whether scores separate errors from correct fields in that workflow. Thresholds need to be chosen by document family, version, source and field, then tested on held-out documents. Empty values require a separate rule because absence and missed extraction can look alike. A global score or an intuitively comfortable cutoff cannot replace this evidence. Calibration is valid only for the evaluated documents, analyzer settings, schema and extraction model; it is not a portable guarantee for every new form.
What the insurance result actually counted
The customer first classified and segmented pages, then identified form versions and extracted fields. Classification used two tiers, with segmentation only in the first. Content Understanding v1 did not provide classification confidence, so classifications still required human review. Sensitive insurance data also meant people checked every field; calibrated labels focused attention rather than authorizing straight-through processing. On held-out customer documents, 41% fewer field values were selected for targeted review while more than 80% of known errors remained in that selection. Some errors still lay outside it. This is not 41% less staffing time or error-free automation. The introductory claims of nearly 50% less review burden, 75% faster form onboarding and 97%+ page classification accuracy have their own measures and must not be substituted for this field-level result.
The public receipt lab is a different experiment
The guide offers a public calibration lab so teams can inspect a method without customer documents. It uses Content Understanding outputs on 1,000 CORD v2 receipts, with 800 used to build policies and 200 held back. Its charts include calibration-set illustrations and are not customer outcomes or guarantees for future data. The proposed output is a table of per-field routing rules, not another model call or a retrained service. Our reading is to reproduce the distinction between policy fitting and evaluation, rather than copy an attractive cutoff into another document collection. High average extraction accuracy alone does not show that a score reliably identifies the errors that matter.
A review policy needs an ongoing feedback loop
The article recommends collecting representative human-checked production examples, setting an explicit error-coverage target, selecting a threshold that meets it and validating on unseen documents. If confidence cannot separate errors, the field stays in review. Monitoring must also sample values not flagged for targeted review; watching score distributions alone cannot establish correctness. Changes in versions, sources, schema, analyzer configuration or underlying LLM can require revalidation and recalibration. Our analysis is that operational permission should follow measured risk and evidence, not a favorable score alone. A routing rule is a business decision with a scope and maintenance obligation, not a permanent safety certificate.
Why it matters
Teams can use grounding to make evidence accessible and confidence to allocate limited review attention. The useful comparison is errors captured and errors missed at a measured review load, with time and cost measured separately. For sensitive work, prioritization may be valuable even when every field remains subject to human checking. AISubHub has not processed customer documents or reproduced Microsoft's evaluation.
Limits and availability
We read the complete official guide, including its customer and public-lab qualifications. We did not open the linked notebook, explainer or other documentation. The article is a technical account, not a new universal accuracy promise, free API offer or consumer subscription entitlement. Data-training statements concern the described confidence model, not every service data-handling policy.
Sources and evidence
Source published: Oct 14, 2026 · Source checked: Oct 20, 2026
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.