Claude AI Legal Review: Long Documents, PDF Limits and UK Practice

JSH LAW / LEGAL AI PRODUCT RESEARCH

England and Wales perspective · Sources reviewed 23 October 2026 · Documentary assessment

The product and the legal-work question

Anthropic develops Claude as a general-purpose AI service, with individual and organisational plans. The relevant question for legal users is how the available reading, drafting and tool functions fit a bounded task under appropriate controls.

The current file guidance describes supported document formats and different processing paths. The service can help structure material, but the existence of a document capability does not establish a complete or legally reliable interpretation of every file.

Separate three jobs: extracting what the source says, organising it into a useful structure and drawing a legal conclusion. A product may help with the first two while the third requires authoritative sources and professional judgment. A polished final narrative can conceal where that boundary was crossed.

The PDF distinction that matters for a legal bundle

Anthropic’s current upload documentation describes chat uploads up to 500 MB, up to twenty files per chat and PDFs up to 1,000 pages, subject to further constraints. Project files have a different stated size limit.

More importantly, it says visual PDF analysis applies to documents of one hundred pages or fewer; longer supported PDFs are processed as text without the visual elements. Non-PDF documents also have text-extraction limitations. These are documentation figures checked on the review date, not a promise of complete comprehension.

For legal work, that can change the result. A signature, handwritten annotation, chart or screenshot may contain something material that is not represented adequately in extracted text. Do not assume the system inspected it merely because the overall PDF uploaded successfully.

Capacity, context and recall are different things

A defensible workflow tests retrieval of the difficult detail, not only the main theme. Ask about a small exception in the middle of a long file and verify the answer against the original. A model’s confidence does not substitute for that check.

Three hypothetical long-document tests

A chronology with competing dates

Create a synthetic correspondence file containing an event date, a later document date, an undated note and a correction. Ask for a table preserving those distinctions and giving exact source locations.

The human ledger should identify the correct relationship between the records. The test fails if the system resolves the uncertainty by inventing a date or promotes one person’s allegation into an established event. Check omissions as well as fabricated entries.

A long PDF containing a material visual element

Use an invented test pack with a chart or handwritten annotation that changes the interpretation. Compare a text-focused task with an explicitly visual task using formats supported by the documented workflow.

The purpose is to establish what information the selected processing path actually makes available. Do not turn a successful answer on a short visual PDF into a claim about a much longer bundle processed differently. This is a proposed test, not a reported Claude result.

A draft revised after decisive new evidence

Prepare a fictional issue summary, then introduce a later source that contradicts a material assumption. Ask for a revised version and a list of changed conclusions with reasons.

A good output should identify the new source and qualify the previous account. It should not preserve an earlier narrative simply because it has become established conversation context. Retain the original, proposed revision and human-approved version for comparison.

Page references, quotations and evidential status

The file guidance distinguishes PDF viewer page positions from printed page numbers. In a legal workflow, record both where necessary. “Page 42” can mean a different document location from bundle page B42 or the number printed on a scanned letter.

For quotations, require exact text and a location another person can open. Check punctuation, omissions and attribution. A plausible paraphrase should not be presented inside quotation marks.

Maintain separate labels for allegation, documentary record, professional opinion, admission and judicial finding. Claude can help organise those labels, but it cannot decide truth from the tone of an account. An apparent contradiction may have an innocent explanation or reflect different dates, sources or questions. The JSH Law digital-evidence guide develops that issue.

England and Wales research needs an authoritative route

Claude’s general reasoning and search functions should not be assumed to provide the same source entitlement and editorial infrastructure as a professional legal database. If the output states the law, inspect the original legislation, judgment or procedural rule.

Specify the jurisdiction and date. Require uncertainty where the source pack is incomplete. A US authority with a similar factual theme is not automatically relevant to an English procedural application, and a real judgment can still be misread.

For family proceedings, keep the particular order, remedy and procedural stage in view. A broad statement about child welfare may not answer the test governing an enforcement application or a request to reopen findings. Do not let a long, coherent document conceal that mismatch.

Personal and commercial data arrangements differ

Anthropic’s consumer-terms update distinguishes individual services from commercial arrangements and describes model-improvement choices and associated retention. Do not assume an individual paid plan has the same terms as an organisational deployment.

The commercial retention explanation also distinguishes API processing from services with persistent storage. A standard API retention statement should not be copied across to every Claude chat, project or file.

For Enterprise, custom retention controls provide configurable arrangements with documented conditions. Establish what applies to conversations, projects, artifacts, logs and connected tools. No-training, retention and access are separate parts of due diligence.

We have not examined a customer contract, audit report or the complete deployment’s data path. Assess UK GDPR, confidentiality, privilege, support access, subprocessors and international processing for the intended use. The hub’s governance section provides a shared framework.

Enterprise controls and the integration boundary

The current Enterprise help page describes organisational controls including audit-related functions, identity management, retention and workplace connectors. Confirm which controls are included in the actual plan and how the organisation configures them.

A connector changes what the assistant can see. An action-capable workflow changes what it can do. Restrict both to the necessary scope and keep approval before consequential external communications or changes.

Also test what happens when a user leaves, a permission changes or a source is corrected. A project containing copied material may need separate treatment from the original repository. Do not assume that revoking access in one place removes every retained copy or generated summary.

Pricing: the seat fee may be only the beginning

The current pricing page and Enterprise explanation describe different individual and organisational arrangements. The public US Pro price is US$20 monthly, with a discounted annual option; local pricing and tax may differ.

The reviewed Enterprise pricing page describes a US$20 seat fee per month on annual billing plus usage at API rates. The help page describes minimum seat arrangements and distinguishes self-serve from sales-assisted purchasing. Confirm those conditions, currency, minimum commitment and actual usage model before treating the headline as an affordable single-user plan.

For long-document work, usage cost can matter. Repeatedly processing a large file, generating revisions and running connected tasks may cost more than a superficial seat comparison suggests. Ask for a realistic cost model using the proposed workload and set appropriate spending controls.

Older contracts and seat types may differ. We have not negotiated a quotation or inspected a billing account. A published plan description is not a guarantee of the final terms offered to a UK organisation.

Measuring value without an unsupported ranking

Compare the same synthetic tasks and final review standard with ChatGPT, a source-backed product such as Lexis+ with Protégé, or a disclosure platform such as Everlaw where the latter is the actual requirement.

Record source fidelity, material omissions, correction time, export quality and permissions. A longer summary is not necessarily better. A shorter answer that clearly identifies missing information may be safer and more useful than a comprehensive-looking account that fills the gaps.

Our assessment: Claude is a credible candidate for supervised reading, synthesis and drafting trials. Its document-processing distinctions and usage model need more attention than a simple “large context” claim suggests. It should not be treated as a legal database, an authentication service or an adjudicator of disputed family evidence.

Does a thousand-page upload mean visual review of every page?

No. The reviewed documentation distinguishes visual processing of shorter PDFs from text-only processing of longer ones. Check the selected workflow and current limits.

Does Enterprise pricing include unlimited free usage?

The reviewed usage-based arrangement separates seat access from metered consumption. Confirm the actual plan and terms rather than assuming the seat fee covers all work.

Sources and methodology

Reviewed on 23 October 2026: file documentation, Enterprise plan, pricing, consumer-terms update, commercial retention and custom Enterprise retention. This is documentary analysis; no comparative benchmark, security audit or live-case experiment was performed. Recheck after changes to processing limits, retention, connectors or billing. Return to the research hub for the wider framework.

If you need help turning a long family-court file into a clear chronology, issue schedule or document plan, see JSH Law’s preparation support. Book a Consultation to agree a defined task and a suitable route for any confidential papers.