Why lawyers use Claude instead of legal AI, and where it breaks

By HAQQ Team · 2026-10-01 · 10 min read · Ai-legal-tech

For AI-fluent lawyers, the real rival to legal AI is a Claude setup they built themselves. Why they stay, where it breaks, and what legal AI must beat.

The competitor legal AI keeps underrating

Search for "Claude for lawyers" and the legal tech vendors who rank give much the same answer: fine for a first pass, buy a legal tool for the real work. Clio's guide, updated on 16 September 2026, gives this verdict: "For research or drafting that depends on verified law or matter facts, use a legal-specific, matter-aware AI tool that cites its sources." We have published versions of that argument ourselves, in legal AI vs generic AI.

For a growing group of lawyers, that answer misses what they actually run. Their alternative is a general model they already pay for, set up with their own projects, their own uploaded files and their own checks.

The surveys show how large that group has become. In 8am's 2026 Legal Industry Report (more than 1,300 responses, fielded September to October 2025), 69% of legal professionals said they use general-purpose AI tools for work, up from 31% the year before, while 42% reported work-related use of legal-specific AI. Thomson Reuters' 2025 Generative AI in Professional Services Report, across 1,702 respondents in legal, tax, risk and government roles, found that 41% personally use publicly available tools such as ChatGPT and 17% personally use industry-specific tools. Among solo and small firms that had used AI, Clio's 2025 report (survey fielded June 2024) found generic non-legal tools such as ChatGPT in use among 57% of solo and 54% of small-firm respondents, level with legal research platforms at 54% and 56%.

Personal use of general-purpose vs legal-specific AI at work — Two surveys, two samples. Each figure is scoped exactly as its source words it.
General-purpose AI	69%
Legal-specific AI	42%
Publicly available tools	41%
Industry-specific tools	17%

8am 2026 Legal Industry Report: 1,300+ legal industry respondents, fielded September to October 2025. Thomson Reuters 2025 Generative AI in Professional Services Report: 1,702 respondents across legal, tax, risk and government roles.

The habit has a public manifesto too. In March 2026, Artificial Lawyer reported that a US lawyer's essay on X, "The Claude-Native Law Firm", had been viewed over 7 million times. Its sharpest line: "This is the capability gap that no specialized legal AI product can match."

We should declare an interest. HAQQ is a legal AI company, and our own FAQ describes HAQQ as "model-agnostic", running on "multiple LLMs, including Claude". So we are not here to argue that Claude is bad. It is good, which is exactly the problem legal AI has to solve, and why this post concedes before it criticises.

Why they stay: three reasons we concede

A price with a ceiling

As of October 2026, the Claude pricing page lists Claude Pro at $20 a month billed monthly. That is one predictable line in a budget, and solo budgets are tight: Clio's 2025 report found that solo firms put about 1% of their total expenses into software, while small firms spend around 2%.

Per-credit legal AI works differently. HAQQ's Legal AI Chat is priced by usage, and our own pricing page says the uncomfortable part out loud: "it is not possible to know the exact credit cost of a request before sending it." For a lawyer who already knows what a long prompt with three attachments is worth to them, a known monthly ceiling is easier to live with than an unknown per-request cost.

The concession has a limit, and Anthropic states it. Every Claude plan has usage limits that reset on a rolling five-hour window, paid plans add weekly limits, and "there's no fixed message count". The price is flat. The amount of work it buys is not.

Even so, a capped bill is easier to judge than a variable one, and judging AI spend is hard even for well-funded teams. In Axiom's 2026 survey of 528 in-house legal leaders, 83% said they cannot measure whether their AI spending is working.

Sources they chose themselves

Claude projects are, in Anthropic's words, "self-contained workspaces with their own chat histories and knowledge bases", where you can upload "relevant documents, text, code, or other files". Projects are available on every plan; free accounts are limited to five.

The careful setup fills that knowledge base with official text. A UAE lawyer can download federal laws from UAE Legislation, which describes itself as "the official platform of the legislations of the UAE government". When the model quotes an article, the lawyer knows which file it came from, because they put the file there.

That is a better habit than trusting a corpus you cannot inspect. It is close to the argument we make in why we argue for primary sources.

A workflow they own

On 16 September 2026, Anthropic merged Claude Cowork and chat into one Claude, rolling out to Pro and Max plans first. The announcement says what Cowork could do is now "available from any conversation, with the context, skills, and connectors you already have". Anthropic's legal page describes practice-area plugins as "extensible with your team's playbook and house style".

Skills are where the fluent lawyer puts their judgment. The essay Artificial Lawyer quoted describes skills that encode "my analytical frameworks, my preferred formats, my voice, and my judgment".

A general model also does not stop at the edge of law. Claude is, in Clio's description, a family of "general-purpose AI models, designed for tasks such as writing, analysis, reasoning, and working with documents". Think of a construction claim that turns on a delay programme, or a shareholder dispute that turns on a valuation. The fluent lawyer asks about the critical path or the discount rate in the same conversation as the contract, without justifying a second tool to anyone. Legal tools are built to be good at law. Plenty of legal work is also about something else.

Where the setup breaks

The concessions are real. So are the limits, and each one shows up in public sources rather than in a vendor's opinion of a rival.

1. The verification burden stays with you

The ABA's Formal Opinion 512 (29 July 2024) does not sort tools into safe and unsafe. Its questions apply to generative AI tools "whether general purpose or designed specifically for the practice of law", including "What level of review of a GAI tool's process or output is necessary?" Anthropic's own legal plugin says the same thing in its disclaimer: "AI-generated analysis should be reviewed by licensed attorneys before being relied upon for legal decisions."

Legal AI does not remove that duty. What the configured setup changes is who maintains the sources. If you built the library, you own its upkeep, and the law moves underneath uploaded files.

Take the UAE's civil code. Federal Decree by Law No. 25 of 2025, issued on 1 October 2025, repealed the UAE's Civil Transactions Law of 1985 (Federal Law No. 5 of 1985) with effect from 1 June 2026, when the new law entered into force. A project knowledge base loaded with the 1985 code before then will still quote it, accurately and confidently. The model is faithful to the file. The file is no longer the law, and nothing in a static upload marks the difference.

Fabricated authority is a separate risk, and court decisions about it keep piling up; our running tally is in the AI legal hallucination audit.

2. Consumer terms, and what Heppner did and did not decide

Pro and Max are consumer plans. Anthropic's 28 August 2025 update to its Consumer Terms and Privacy Policy applies to "Claude Free, Pro, and Max plans" and not to services under its Commercial Terms, such as Claude for Work. Under it, Anthropic trains new models on data from those accounts "when this setting is on", and keeps that data for five years; users who decline keep the existing 30-day retention period.

That matters because of United States v. Heppner. Judge Jed Rakoff of the Southern District of New York ruled from the bench on 10 February 2026, and in a written opinion on 17 February, that roughly thirty-one documents a defendant generated with Claude were protected by neither attorney-client privilege nor the work product doctrine. Part of the reasoning was the privacy policy, which, the court said, provides that Anthropic uses inputs and outputs to "train" Claude and may disclose data to "third parties".

Scope it carefully. Heppner was a client who used Claude "on his own volition", not a lawyer, and his counsel "did not direct" him to do it. The opinion calls Claude "a publicly available AI platform" and does not say which plan or account he used. The court also left an opening: had counsel directed the use, Claude "might arguably be said to have functioned in a manner akin to a highly trained professional who may act as a lawyer's agent". Whether a lawyer's own Pro account would be analysed the same way is untested. A confidentiality argument that leans on consumer terms is still not one a careful lawyer wants to make. Our full Heppner analysis goes further.

3. One person's setup is not a firm's memory

Anthropic's own pricing page shows where the controls a firm usually writes into its AI policy sit. Team adds project sharing and, in Anthropic's words, "no training on your data by default". Enterprise adds SCIM, audit logs, custom data retention and role-based access on top of Team, and it is billed per seat "plus usage billed at API rates, so cost scales with the models and tasks your team runs." The flat price that kept the lawyer on Claude and the full set of firm controls sit on different plans.

There is a quieter gap too. On Pro, the setup lives in one person's account: their skills, their uploaded statutes, their checks. Project sharing is a Team and Enterprise feature, so when that lawyer is away, so is the setup.

4. Jurisdiction coverage, rebuilt per matter

Anthropic's legal plugin is candid about its defaults. Its playbook examples "reflect U.S. legal positions and jurisdictions (Delaware, New York, California)", and anyone under a different legal system "must customize the playbook". For a GCC lawyer, that customisation is the job.

The UAE alone has at least three bodies of civil and commercial law. Federal laws are published on the UAE Legislation platform. The DIFC is "empowered to create its own legal and regulatory framework for all civil and commercial matters", on principles of common law. ADGM has "its own civil and commercial laws" under a framework "based on English Common Law". A matter with an onshore counterparty and a DIFC entity needs two libraries, assembled and kept current by hand, before the first prompt. Across the wider GCC, the work multiplies.

What a legal AI has to do to win this lawyer

If the configured Claude setup is the bar, the bar is specific. A legal AI that wants the AI-fluent lawyer has to clear all of it, not most of it:

A bill they can predict: a ceiling, or a cost shown before the work runs. "We will tell you afterwards" loses to a flat monthly fee.
Sources on the screen: every answer tied to the text it came from, with room for the lawyer's own uploads next to the platform's.
In-force status, not just text: flag the repealed code before the model quotes it.
Firm memory without the Enterprise trade: shared matters, permissions and an audit trail, without giving up a predictable price.
Jurisdictions that do not need rebuilding: onshore and free-zone law across the GCC, in Arabic and English, kept current by someone other than the lawyer on the matter.
No walls: the lawyer will still want the delay programme and the valuation in the same session. A tool that refuses to leave law sends them back to Claude.

We hold HAQQ to that list, and we do not clear every line of it today; the first line is on our own pricing page. Where we think HAQQ does clear the bar, and where plain Claude stays the better choice, is the subject of why a lawyer would pick HAQQ over plain Claude.

Key takeaways

For AI-fluent lawyers, the main alternative to legal AI is a general model they configured themselves, and the surveys show general-purpose use running ahead of legal-specific use.
Claude wins this lawyer on a price ceiling, on sources they chose and on a workflow they control. Those are real advantages, and pretending otherwise loses the argument.
The setup breaks on verification upkeep, consumer terms, firm controls and jurisdiction coverage.
On Anthropic's own pricing page, the firm controls sit on Team and Enterprise, and Enterprise is billed per seat plus usage, so the flat price and the controls do not come together.
To win this lawyer, a legal AI has to clear the whole bar: predictable cost, visible sources, in-force status, firm memory and GCC coverage without walls.

Sources & further reading

FAQ

Why do lawyers use Claude instead of legal AI tools?

For AI-fluent lawyers, a configured Claude setup wins on three things: a flat subscription price, sources they upload themselves into Claude projects, and a workflow they control with skills and connectors. Surveys show general-purpose AI use running ahead of legal-specific use: 8am's 2026 Legal Industry Report found 69% of legal professionals use general-purpose AI tools for work, against 42% reporting work-related use of legal-specific AI.

Is Claude Pro suitable for confidential client work?

That is a judgment for each lawyer under their own professional rules, so start with the terms. Anthropic's Consumer Terms cover the Free, Pro and Max plans, and on those plans model training depends on a setting the user chooses. On Anthropic's own pages, no training on your data by default is attributed to the Team and Enterprise plans, and audit logs, SCIM, custom data retention and role-based access are listed as Enterprise controls.

Does US v. Heppner mean using Claude waives privilege?

Not as a general rule. In United States v. Heppner (S.D.N.Y., written opinion of 17 February 2026), Judge Rakoff held that documents a defendant created with Claude on his own volition, which his counsel had not directed, were protected by neither attorney-client privilege nor the work product doctrine. The court relied partly on the privacy policy that Claude users consent to and left an opening for use directed by counsel. How the reasoning applies to a lawyer's own use is untested.

How much does Claude cost compared with legal AI?

As of October 2026, Anthropic's pricing page lists Claude Pro at $20 a month billed monthly, and every Claude plan has usage limits that reset on a rolling five-hour window. Claude Enterprise is billed per seat plus usage billed at API rates. Per-credit legal AI works differently: HAQQ's own pricing page says it is not possible to know the exact credit cost of a request before sending it.

Do general AI and legal AI change a lawyer's duty to verify?

Not under the ABA's guidance. Its Formal Opinion 512 frames its questions for generative AI tools whether general purpose or designed specifically for the practice of law, including what level of review of the output is necessary. Anthropic's own legal plugin says its analysis should be reviewed by licensed attorneys, and its default playbooks reflect US jurisdictions (Delaware, New York, California), so lawyers elsewhere must customise them.

Does HAQQ use Claude?

Yes, among other models. HAQQ describes itself as model-agnostic, using multiple LLMs, including Claude.