Azure Document Intelligence

# Fix list: Azure Document Intelligence

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/azure-document-intelligence, the October 2026 research run, assessed 8 October 2026. Grade B, 66 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on Azure Document Intelligence: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Payments & pricing, 20 out of 100, up to 10 more on the total

Why it scored 20: No x402, MPP or L402 (0). Per-1,000-page prices are public through the Azure Retail Prices API without a login, while the pricing page shows "$-" until a script fills it (20). F0 gives 500 pages a month, but it sits on an Azure account, and the free account page asks for a phone number and a credit or debit card (0). A person signs up in a browser (0).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-payments):

The published rubric, also on the [x402 page](https://www.anchorterminal.com/x402/).

- 40, a machine payment protocol (x402, MPP or L402) on the tool's own endpoints. 10 to 30 when it covers only some endpoints or only goes through a third party, and the note says which.
- 20, per-call or per-unit pricing published without a login. 10 for public plan-only pricing, 0 for "contact sales" or prices behind a login.
- 20, a free tier or trial that doesn't need a card.
- 20, autonomous onboarding, meaning an agent can get access without a person signing up in a browser (keyless use, x402, a programmatic key API).

Payment platforms and agent wallets rarely charge for their own API over a machine protocol, so the first line has steps for them, and the highest one that applies counts. 40 when x402, MPP or L402 runs on all their own endpoints, 30 when it runs on part of their own API, 25 when their merchants can accept one, 20 for running a facilitator, 15 for paying as a buyer, and 0 when the only protocol is their own. Merchant acceptance sits above a facilitator because the platform's own customers can charge agents through it, while a facilitator settles for sellers who wire up the protocol themselves. The counter-argument (a facilitator does more for the protocol as a whole) has a point. Each note says which step applied.

Open-source software you run yourself is scored on its hosted or paid option if it has one. A free, self-hosted package with nothing to buy gets 20, 20 and 20 for the last three lines, and 0 to 40 for the first only if it ships a payment protocol.

## 2. Agent ergonomics, 64 out of 100, up to 5.9 more on the total

Why it scored 64: API reading. `pages` limits what is analysed, `outputContentFormat=markdown` gives one Markdown string and add-ons are off unless asked for. The result still carries every word and line with polygons, with no field selection, and the limits page allows a response of up to 500 MB (14 of 25). List calls page with `nextLink`. No filter on models or operations, and an analysis result comes back whole (12 of 20). Errors are JSON with a code, a message and an inner code such as `InvalidContentLength` or `ModelNotReady` that says what to change (16 of 20). No idempotency key, so a repeated POST starts and bills a second analysis. Results stay for 24 hours, so polling again is safe (10 of 20). A call needs only a model ID, `api-version` and a URL or bytes, and SDKs exist for Python, JavaScript, .NET and Java. Every analysis takes a POST and at least one GET (12 of 15).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-ergonomics):

- 0 to 25, context cost. For MCP, the number and size of the tool definitions (25 for ten or fewer compact tools, 15 for 11 to 30, 5 for more than 30, plus up to 10 back for toolsets, dynamic loading or read-only subsets). For APIs, whether responses can be sized (field selection, limits, summaries).
- 20, pagination, filtering and output-size controls.
- 20, actionable, documented error responses, codes and messages an agent can recover from.
- 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations.
- 15, sensible defaults, few required parameters, and official SDKs in at least two languages.

Models are read for tool use, structured output, prompt caching, context length, batch and SDKs. Frameworks for how much code and how many defaults a tool-calling agent with MCP needs.

## 3. Maintenance & community, 40 out of 100, up to 5.3 more on the total

Why it scored 40: Read as a closed service with official SDKs. The newest release is Java 1.0.11 on 6 October 2026, a dependency update. The service's own newest dated change is the March 2026 tax model update, more than 180 days back, and the other three SDKs last shipped between December 2024 and May 2025. We give half for recency, a departure from the checklist, which would give 30 on the Java date alone (15 of 30). Two releases in the last 90 days, Java 1.0.10 on 18 August and 1.0.11 on 6 October (0 of 20). A what's new page with one entry in 2026, Microsoft Q&A and the SDK issue trackers on GitHub, which we did not read (8 of 15). Official SDKs in four languages target the current GA API, but Python, JavaScript and .NET have had no release for 16 to 22 months (12 of 15). The Python package declares Python 3.8 to 3.12 and is built in the azure-sdk pipelines, whose results we did not check (5 of 10).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-maintenance):

- 0 to 30, time since the last release, or the last published model or API change for a closed service. 30 within 30 days, 20 within 90, 10 within 180, 0 older.
- 20, at least three releases or dated changelog entries in the last 90 days.
- 0 to 25, responsiveness. Issues and pull requests answered on GitHub (the open issues and how recent the replies are). For closed services, a public changelog and a support or community channel that answers, 0 to 15.
- 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models).
- 10, package health, current dependencies and CI.

Models are read for deprecation notice periods and model churn rather than release counts.

## 4. Security & auth, 71 out of 100, up to 5.1 more on the total

Why it scored 71: Microsoft Entra ID bearer tokens with Azure RBAC and managed identities, or two resource keys. The OpenAPI document and the docs give the key only as a header (30). RBAC can limit a caller, but a resource key reaches every data-plane call, among them deleting custom models, and nothing asks for confirmation. Analysis itself changes nothing (14 of 20). The service returns text from untrusted documents, and no guidance on injected instructions was found in the pages read (0 of 15). The Foundry Tools diagnostic logging page describes Audit and RequestResponse logs through Azure Monitor once the owner adds a diagnostic setting. We did not confirm the fields for Document Intelligence (12 of 15). MSRC coordinated disclosure and bounty policies and a CSAF feed. The microsoft.com security.txt passed its Expires date on 23 September 2026, and we did not confirm Document Intelligence on an audit scope list (15 of 20).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-security):

- 0 to 30, the credential model. 30 for OAuth 2.1 with scopes, or scoped and revocable keys with rotation. 20 for plain revocable API keys. 10 for one all-powerful key. 10 off when a secret can travel in a URL query string as a documented option.
- 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions.
- 0 to 15, prompt-injection posture where the tool returns untrusted content (documented mitigations or guidance). A tool that returns no untrusted content gets 10.
- 0 to 15, audit logs or per-call visibility for the operator.
- 0 to 20, a security programme. security.txt or a disclosure policy, a bug bounty, SOC 2 or ISO 27001, advisories handled in public.

Models are read for retention, whether API data trains models (and whether that's off by default), zero-retention options and certifications. Frameworks for telemetry defaults, approval hooks, guardrails and sandboxing.

## 5. Reliability, 82 out of 100, up to 3.6 more on the total

Why it scored 82: Hosted reading. Azure status page with a history of post-incident reviews (20). The history lists three incidents in the last 90 days. One, on 29 September 2026 from 10:03 to 15:58 UTC, names Cognitive Services with intermittent failures and latency in Sweden Central, and Document Intelligence resources are Cognitive Services resources. None names Document Intelligence, and the page lists only broad incidents, so we count the record as minor (20 of 30). Limits published with numbers, 15 analyse calls, 50 result reads, 5 model management calls and 10 list calls a second on S0 (15). The limits page asks for retry logic, exponential backoff on 429 and at least 2 seconds between polls, and the 202 response carries `Retry-After`. No idempotency key exists for a repeated analyse call (12 of 15). The pricing page links Microsoft's Online Services SLA, a download we did not open, so half (5 of 10). API version 2024-11-30 is GA (10).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reliability):

Hosted APIs, MCP servers, models and platforms.

- 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own).
- 0 to 30, the incident record for the last 90 days on that page. 30 for a clean record or trivial incidents only, 20 for minor incidents only, 10 for one major outage (an hour or more of a core API down, or errors across the board), 0 for several. 5 when there's no history we could read, and the note says so.
- 15, rate limits documented with numbers.
- 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved.
- 10, an SLA published for any paid tier.
- 10, the surface agents use is generally available, not beta or preview.

Local packages, SDKs, frameworks and stdio MCP servers.

- 20, installs from an official package with supported runtimes stated.
- 25, a public CI and test suite, passing on the default branch.
- 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered).
- 15, semver discipline and breaking changes called out in a changelog.
- 15, version 1.0 or later, or declared stable.

Protocols are read from their reference implementations, the public facilitators or servers, spec stability and test vectors.

## 6. Schema & documentation, 86 out of 100, up to 2.3 more on the total

Why it scored 86: A public OpenAPI 2.0 document for 2024-11-30 in Azure/azure-rest-api-specs, generated from TypeSpec, with 27 operations (25). learn.microsoft.com/llms.txt returns 404, but every Learn page we asked for with `Accept: text/markdown` came back as Markdown, so half (5 of 10). Operation descriptions are one line each and restate the operation name. The overview says which model fits which document and links a page on choosing between this service and Content Understanding (14 of 20). 25 enums, 94 length, pattern and range constraints and a pattern on `pages` and `modelId`. `queryFields` is a free list of strings and the either-or between `urlSource` and `base64Source` is stated only in prose (13 of 15). An example on every operation, and an error page with nine top-level codes and 30 inner codes. 429 is not in that table (14 of 15). Dated `api-version` values, a what's new page and an SDK release history page (15).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-schema):

APIs and MCP servers.

- 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool).
- 10, llms.txt or Markdown docs served for agents.
- 0 to 20, descriptions that say what a tool is for, when to use it and when not to, read from the tool definitions in the source or the API reference.
- 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs.
- 0 to 15, examples and documented error responses.
- 15, versioning and a public changelog.

Models are read from the API reference, the OpenAPI file, llms.txt, the structured-output and tool-use docs and the model cards. Frameworks from docs a model can follow, typed interfaces, examples and the API reference.

## 7. Transparency & trust, 78 out of 100, up to 1.9 more on the total

Made of editorial 69, provenance 86.

Why it scored 78: Closed service under Microsoft's Product Terms, with MIT client libraries (15 of 30). The data privacy page, dated 22 July 2026, says data is processed in the resource's region, held with results for 24 hours in storage shared by customers in that region and logically isolated, then deleted, with a delete call for earlier removal. It says nothing on whether documents are used to improve the service, the Product Terms' training clause covers generative AI services only, and we did not read the Data Protection Addendum (22 of 30). The overview gives a version table with dated retirements, v2.1 on 15 September 2027, announced three years ahead, and v3.0 on 30 March 2029 (20). Processing stays in the resource's region. The sub-processor list is on the Service Trust Portal, which we did not read (12 of 20).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-transparency):

- 0 to 30, source availability and licence clarity. 30 for open source under an OSI licence, 15 for closed with clear terms, 0 for unclear terms.
- 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors).
- 0 to 20, a deprecation policy or notices with dates.
- 0 to 20, telemetry disclosed with an opt-out (local software), or subprocessors and data locations disclosed (hosted).

The other half of Transparency and trust is the provenance score, computed from checked facts (below). The category score is the mean of the two.

Provenance checks not met in full (half of this category, computed from checked facts):

- Terms of service: read, states 4 of the 7 things a reader expects, and has 2 clauses that cost points (3.4 of 10)
- Privacy policy: read, states 8 of the 8 things a reader expects, and has 1 clause that costs points (8 of 10)
- security.txt: published but past its Expires date (5 of 10)

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- unchecked: the Online Services SLA document is a download we did not open, so the uptime commitment for this service is not confirmed.
- unchecked: the PyPI project page answered with a client challenge. The Python version and date come from the changelog in Azure/azure-sdk-for-python.
- unchecked: the Data Protection Addendum, the Service Trust Portal sub-processor list and audit scope lists were not read.
- unchecked: whether documents sent for analysis are used to improve the service. The data privacy page does not say.
- unchecked: the built-in role and the log fields specific to Document Intelligence. The role and log categories come from the general Foundry Tools pages.
- Maven Central's robots.txt disallows every path. One request for the Java package's metadata was made before that was acted on and was discarded. The Java dates here come from the changelog on GitHub.
- The product page says Document Intelligence is now part of Azure Content Understanding. The docs keep it as a separate API with no end date for v4.0. A later pass should check whether Content Understanding needs its own listing.
- `lastRelease` is the Java SDK's 6 October 2026 dependency update. The service's newest dated change is March 2026.

## Weaknesses

- Every analysis is asynchronous. A POST returns 202 with `Operation-Location`, and the caller polls for the result
- The result carries every word and line with polygons, with no parameter to leave them out. Only `pages` narrows it
- No idempotency key. Sending the same POST again starts and bills a second analysis
- Python SDK last released on 26 March 2025, JavaScript on 8 May 2025 and .NET on 16 December 2024. Only Java shipped in 2026, with dependency updates
- The free F0 tier reads only the first two pages of a request and needs an Azure account, which asks for a card and a phone number
- No guidance on instructions hidden in document text was found in the pages read

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- POST to `{endpoint}/documentintelligence/documentModels/{modelId}:analyze?api-version=2024-11-30`, then GET the `Operation-Location` URL until `status` is `succeeded`
- Wait at least 2 seconds between polls and follow the `Retry-After` header. Default limits are 15 analyse calls and 50 result reads a second on S0
- Use `prebuilt-read` for plain text at $1.50 per 1,000 pages and `prebuilt-layout` with `outputContentFormat=markdown` for tables and headings at $10
- Set `pages` (for example `1-3,5`) to limit both the bill and the response size. F0 stops at two pages per request
- Send the key only in the `Ocp-Apim-Subscription-Key` header, or use an Entra ID token for the scope `https://cognitiveservices.azure.com/.default`
- Treat extracted text as untrusted input. Results expire 24 hours after the job completes

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.