Research-backed calculator guide

Document Classification Cost per Document: Budget the Routing Step Before OCR

Estimate document classification cost per document by separating page or document meters, retries, review, and downstream extraction with a worked routing example.

classification cost per incoming document
= (classification charges
 + retry charges
 + review labor
 + downstream route charges
 + fixed-cost allocation)
 ÷ incoming documents

For a page-metered classifier:

classification charges
= documents × pages per document × classifier rate per page

For a document-metered classifier:

classification charges
= documents × classifier rate per document

Do not mix those units. A three-page packet billed at $3 per 1,000 pages is not equivalent to a $0.01-per-document classifier unless the workload, included pages, and feature behavior are the same. A rate card is a dated planning input, not a vendor quote. On 2026-11-02, Azure’s U.S. pricing page listed custom classification at $3 per 1,000 pages and custom extraction at $30 per 1,000 pages; it also says prices are estimates and can vary by agreement, date, and currency.^[2]

A classification-only number can still understate the decision. If 8% of documents go to review, reviewer minutes may exceed the API charge. If the classifier sends 70% of packets to extraction, that route may dominate the budget. Price the routing step as a small decision system, not as an isolated API call.

What document classification includes

In an IDP pipeline, classification answers questions such as “is this an invoice, a contract, or an identity document?” Splitting answers a related question: “where does one document end and the next begin inside this packet?” A routing step can do one or both. It may also return an other, unknown, or unprocessable outcome.

Those distinctions affect cost. A single upload may contain 20 pages and four underlying documents. A page-based classifier can charge for every analyzed page. A document-based processor may charge for each classified item, possibly with an included page limit. A splitter may be charged before classification, after OCR, or as part of a specialized processor. Read the provider’s definition of a count before converting it into cost per document.

Google’s Document AI rate card separates custom splitter, custom classifier, OCR, layout parsing, extraction, and specialized processors. It also shows some specialized processors priced per classified document and notes that one count can represent up to 10 pages for those processors.^[1] That is why “per document” still needs a page-limit assumption beside it.

Azure’s rate card separates Read, prebuilt models, custom classification, custom extraction, query fields, add-ons, and batch variants. Its FAQ says Document Intelligence is billed by pages analyzed and that all pages are analyzed unless a page range is specified.^[2] AWS Textract uses a different structure: its APIs are feature-specific, and some workflows combine classification and extraction rather than exposing a generic standalone classifier price. Its Analyze Lending API, for example, can classify, split, and extract mortgage-package documents under a page-based workflow.^[3]

The first budgeting task is therefore to create a meter map:

• Input unit: file, packet, logical document, page, image, or processor count.
• Classification coverage: the share of inputs that actually uses the classifier.
• Retry rate: the share of pages or documents sent again after timeout, parse failure, or low confidence.
• Route shares: extraction, straight-through acceptance, other, quarantine, or human review.
• Quality work: reviewer minutes, re-keying, correction, or a second model pass.
• Fixed costs: model hosting, minimum commitments, queue workers, storage, and integration allocation.

A practical cost-per-document formula

Let:

• D = incoming documents per month;
• P = average pages per incoming document;
• R = classifier rate per page, in dollars per page;
• Q = classifier rate per document, in dollars per document;
• x = fraction of inputs classified;
• t = retry fraction on classified work;
• h = fraction of documents reviewed by a person;
• m = review minutes per reviewed document;
• w = loaded reviewer cost per hour;
• E = downstream extraction and route charges per month; and
• F = monthly fixed-cost allocation.

A page-metered estimate is:

classifier = D × x × P × R × (1 + t)
review     = D × h × (m / 60) × w
monthly total = classifier + review + E + F
cost per incoming document = monthly total / D

A document-metered estimate replaces the classifier line with:

classifier = D × x × Q × (1 + t)

When a provider publishes a price per 1,000 pages, convert it first. Azure’s observed custom-classification rate of $3 per 1,000 pages is $0.003 per page under the displayed U.S. pay-as-you-go list price.^[2] That conversion is only valid for the stated feature, region, tier, and date. It is not a universal IDP classification benchmark.

Review should be based on measured routing behavior, not a generic accuracy claim. Microsoft’s OCR guidance describes confidence values from 0 to 1 and gives thresholding as a way to send higher-confidence results straight through while routing lower-confidence results to human review. It also warns that scan quality, resolution, contrast, rotation, and text attributes affect performance.^[4] In cost terms, the threshold changes h, the reviewed share, and sometimes m, the minutes per exception.

Worked cost example: 50,000 mixed documents per month

Assume a scanning operation receives 50,000 incoming documents per month, averaging 3 pages each. It uses a page-metered custom classifier at the Azure list rate observed on 2026-11-02: $3 per 1,000 pages, or $0.003 per page.^[2] Assume 15% of classified pages are retried, 8% of incoming documents need human review, review takes 2 minutes, and loaded reviewer cost is $32 per hour. These are planning assumptions, not vendor measurements.

The classifier handles 150,000 pages. The base classifier charge is 150,000 × $0.003 = $450. The retry allowance is $450 × 15% = $67.50, so classification plus retries are $517.50, or $0.01035 per incoming document.

Review labor is:

50,000 documents × 8% × (2 minutes ÷ 60) × $32/hour
= $4,266.67 per month

That is $0.08533 per incoming document, even though it is not on the classifier invoice. At these assumptions, review is much larger than the classification API charge. A different threshold may lower review volume but increase false accepts; test the trade-off on representative files.

Now add the routes selected by classification. Suppose 70% of the pages go to custom extraction at the observed Azure rate of $30 per 1,000 pages, 20% go to a prebuilt path at $10 per 1,000 pages, and 10% are other or quarantined without extraction. The route calculation is:

custom extraction = 150,000 × 70% × $0.030 = $3,150
prebuilt route    = 150,000 × 20% × $0.010 = $300
other/quarantine  = 150,000 × 10% × $0.000 = $0 assumed

The modeled monthly total is therefore $517.50 + $4,266.67 + $3,150 + $300 = $8,234.17, before storage, orchestration, taxes, or fixed platform costs. Dividing by 50,000 gives $0.1647 per incoming document. The classification API itself is only about 6.3% of this modeled total; the route and review assumptions drive the rest.

The other assumption deserves care. Google’s pricing page says a procurement document classified as “other” is not billed for that specialized processor.^[1] That does not mean every provider waives every downstream charge for an unknown document. In this example, $0 is an explicit modeling assumption for avoided extraction, not a general rule. If quarantined files incur storage, resubmission, or manual sorting, add those costs.

How to compare page-metered and document-metered offers

Normalize both offers to the same input cohort. For a page-metered quote, multiply documents by measured average pages. For a document-metered quote, record the included page limit and the overage rule. Then add the same review and route assumptions to both.

A short-document workload can favor a document meter, while a long-packet workload can make a page meter more predictable. Mixed packets create another issue: a “document” may mean a file, a split component, or a processor count. Ask the vendor to price a sample month containing one-page forms, 10-page contracts, and multi-document packets rather than accepting a single blended example.

Also separate classification cost from classification-enabled extraction cost. If a classifier is cheap but routes nearly every page to an expensive extractor, its effective workflow cost may be higher. Conversely, a more expensive classifier can be rational if it avoids extraction on irrelevant or unsupported material. Report both numbers:

1. cost per classified item; and
2. cost per accepted or routed item after review and downstream processing.

The model pricing comparison is useful when a classifier or fallback uses a model with token pricing. The RAG cost calculator helps when classification is part of a retrieval-ingestion workflow and the route creates embedding or storage work. If the estimate later needs to be checked against provider exports, use the AI API invoice reconciliation calculator.

Assumptions, exclusions, and limitations

The worked example uses public U.S. list-price observations checked on 2026-11-02. Azure says displayed prices are estimates rather than quotes and may vary with agreement, purchase date, region, currency, and offer.^[2] Google also notes that prices can depend on consumption model and processor eligibility.^[1] Recheck the rate card before making a commitment.

This model assumes average pages per document, fixed route shares, and a measured or estimated retry rate. Real costs can vary with file format, page ranges, scan quality, language, handwriting, table density, concurrency, batch mode, data transfer, storage, model hosting, minimum commitments, taxes, and support plans. It excludes the cost of collecting or labeling training data, building the classifier, integration engineering, security review, and the value of misclassification errors.

It also does not assume that a provider’s confidence score is calibrated to your business loss. A false “invoice” route and a false other route can have different consequences. Use a pilot with representative documents, record the confusion matrix by class, and measure reviewer minutes by class. Then update h, m, and route shares rather than applying a generic accuracy percentage.

FAQ

What is the cheapest way to classify documents?

There is no universal cheapest method. A deterministic rule or local model may have low marginal API cost but higher engineering and infrastructure cost. A managed classifier may be faster to deploy but bill per page, document, processor count, or workflow step. Compare cost per accepted route after review, not only the classifier’s list price.

Is document classification billed per page or per document?

Both models exist. Azure’s displayed custom-classification rate is per 1,000 pages, while Google’s page includes specialized processors with per-classified-document prices and separate custom-classifier and splitter meters.^[1][2] Read the unit definition, included pages, page-range behavior, overage, and whether splitting is a separate operation before converting it to a per-document estimate.

How do I calculate document splitting cost?

Count the unit the splitter actually processes. If a splitter analyzes every page, use documents × pages per document × splitter rate per page. If it charges per packet or classified item, use the document meter. Add retries for boundary failures and the labor needed to correct bad splits. For mixed packets, track both the number of uploaded files and the number of logical documents created.

Should human review be included in classification cost?

Yes, when the goal is a useful operating cost. Keep it as a separate line so the API rate remains visible, then report the fully loaded cost per incoming document. Confidence thresholds, scan quality, and class difficulty can change the reviewed share and review minutes.^[4]

What does an “other” classification do to the budget?

It depends on the provider and workflow. It may avoid an extraction charge, or it may trigger quarantine, manual sorting, a fallback model, or a resubmission. Model the observed outcome separately. Do not assume that “other” is free merely because one specialized Google processor documents no charge for that outcome.^[1]

How often should document classification rates be rechecked?

Recheck whenever the provider changes a processor, region, batch mode, commitment tier, or page-count definition, and before a budget is approved. The figures in this guide are dated observations from 2026-11-02, not evergreen quotes. Reconcile a pilot month to the provider usage export so the estimate reflects actual pages, retries, route shares, and review work.