Lawyers ask us two questions about the AI under ModulawAI more than any other: which model is it, and why that one. The answer to the first is Anthropic’s Claude, run through Amazon Bedrock so that client data stays inside our AWS environment and is never used for training. The answer to the second takes a little longer, and it is worth giving honestly, including the parts where a model on its own is not enough.
What Claude is genuinely good at for legal work
- Long documents. Claude reads very long inputs, an entire disclosure bundle or a two-hundred-page agreement, and answers questions about them without losing the thread. That is the capability behind tabular review, where one question is asked across every document in a set.
- Following instructions precisely. Told to answer only from the retrieved sources, to cite every proposition, and to say when the sources do not support an answer, it does so more reliably than most models we have tested. In legal work that discipline matters more than eloquence.
- Careful drafting. It produces first drafts that read like a competent junior wrote them: structured, hedged where hedging is right, and willing to flag an assumption rather than paper over it.
- Refusing to bluff, mostly. Asked something it cannot answer from the material in front of it, it is more likely than older models to say so. Mostly. Which brings us to the failures.
Where a raw model still fails a lawyer
None of the following is specific to Claude. They are properties of language models, and they are why “just use the chatbot” is bad advice for anything you will cite.
- It does not know your jurisdiction’s current law. A model’s training has a cut-off and no access to last month’s Court of Appeal decision or this year’s amendment. Ask it about the Evidence Act and it will answer from a general memory of evidence law, which may be a different country’s.
- Asked for an authority from memory, it can invent one. The case name will be plausible, the citation format will be right, and the case will not exist. This is the failure that has led to sanctions in several countries, and it is why an ungrounded model should never be the source of a citation.
- It does not know which matter you are in. A chat window has no idea who the client is, what was filed last week, or which precedent your firm prefers. Every answer starts from nothing.
- It cannot show its working unless you make it. A paragraph of confident prose with no links is unverifiable, and unverifiable is unusable in practice.
What ModulawAI builds around the model
The product is the part around the model, and it exists to remove each failure above.
- Retrieval before answering. When you ask a research question, ModulawAI first searches its own indexed corpus of primary law, judgments and statutes for the US, UK, Canada, EU, Nigeria, Ghana, Kenya and South Africa, and hands the relevant passages to Claude with an instruction to answer only from them. Every citation in the answer links to the source it came from, and the authority’s subsequent treatment is flagged. The model does the reading and writing; it is not the source of the law.
- The matter as context. Inside a case, the assistant knows the parties, the documents, the deadlines and the firm’s precedent library. A drafting request starts from your last good version of the document, not from a generic template.
- Structured outputs. Tabular review returns a grid, not prose, with each cell citing a passage, so the output can be sorted, checked and exported. Redlines in Word arrive as track changes so the lawyer accepts or rejects each one.
- Controls and an audit trail. Agent skills are scoped, critical actions can require a named approver, and every input, output and tool call is logged with the reviewer. That is what turns a capable model into something a firm can stand behind.
- Privacy by architecture. Claude runs through Amazon Bedrock inside our AWS environment. Prompts and documents are not sent to Anthropic’s consumer services, are not used to train any model, and are encrypted in transit and at rest.
The uncomfortable part: Claude has produced fake citations too
We would rather say this ourselves than have you find it elsewhere. In 2026 a US federal court sanctioned an immigration attorney who submitted fabricated legal quotations produced by Claude Sonnet 4 in an emergency habeas filing seeking to halt a client’s deportation. The model is not exempt from the failure mode described above, and any vendor telling you their underlying model does not hallucinate is either misinformed or selling.
The scale of the problem is now measurable. Damien Charlotin maintains a public database of court decisions involving fabricated AI-generated content, which recorded 2,041 cases as at 14 September 2026 across more than 40 jurisdictions, with 1,395 in the United States, 217 in Canada and 111 in Australia. The foundational case is Mata v Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023), a $5,000 sanction over six invented decisions. In England and Wales the Divisional Court in Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin) made wasted costs orders of £2,000 plus VAT against each of two lawyers and required them to self-report to their regulators after five non-existent authorities were cited.
The lesson is not that the model is unsuitable. It is that the model is the wrong layer to hold the law. Every one of those cases involved a lawyer asking a model to recall an authority rather than asking a system to retrieve one. In ModulawAI the retrieval happens first and the model never supplies the citation, only the reading and the writing. That is an architectural choice, not a prompt, and it is the difference between the two workflows.
What your regulator expects of you, not of us
ABA Formal Opinion 512, issued 29 July 2024, is the first formal US ethics guidance on generative AI. It requires competence in the tools used, an assessment before entering client information as to whether it could be disclosed to or accessed by others, billing that reflects time actually spent, and supervision of non-lawyers and third parties using these tools. It also treats the cost of an AI subscription as firm overhead rather than a client disbursement. The SRA’s warning notice of 17 August 2026 states that solicitors remain accountable for all work and advice delivered whether or not AI was used, that reliance on an AI output would not be a suitable defence, and that client information should only be entered into AI systems where appropriate contractual, technical and organisational safeguards are in place.
Running Claude through Amazon Bedrock inside our AWS environment, with no training on client data, addresses the confidentiality limb. Linked citations address the verification limb. Logging every AI action against the fee earner who reviewed it addresses the supervision limb. None of that transfers the professional responsibility, and no product can. Sources for this section: the AI Hallucination Cases database, ABA Formal Opinion 512, and the SRA warning notice on the misuse of AI.
Can I just use Claude directly for legal work?
For orientation, yes: understanding an unfamiliar area, explaining a concept, tidying your own prose. Treat it as a well-read colleague with no access to the library. For anything you will rely on, cite or file, no, and the same is true of any general model. The value is not the model’s knowledge of the law; it is the model’s ability to read the law you give it, quickly and carefully, and to show you where each answer came from. That is what ModulawAI is for.
You can also connect Claude, or any assistant that speaks the Model Context Protocol, to your own ModulawAI workspace, so it can read matters and documents under your permissions. See the MCP guide and the AI legal assistant.