On October 28, 2026, Anthropic announced artifacts for legal work: a gallery of ready-to-use, interactive tools inside Claude, built for in-house legal teams and law firms. It follows Claude for Legal, which launched in May, and Anthropic says it responds to two requests from legal teams: simpler setup and easier collaboration.
That sounds like a product announcement, and it is. But underneath, it's an interesting piece of applied AI engineering. Each artifact is a small, repeatable workflow: pull documents in through connectors, run a structured comparison with Claude, and render the result as a table, board, or timeline where every output links back to its source. This piece breaks down how that works, what each artifact is doing, and what a legal team needs to think about before trusting it with real matters.
What an "artifact" is here
In Claude, an artifact is an interactive tool, such as a tracker or a dashboard, that a team can adapt, use, and share. The legal gallery is a set of these pre-built for legal work.
Three details from the announcement matter technically:
- Each artifact opens with an example. You see the output on sample data before you add your own documents or connect a system. That's a good design choice: it shows the output schema up front, so you know what "done" looks like before anything confidential goes in.
- You can ask Claude to customize it. Teams can tailor an artifact to their own process and connect their playbooks or research sources, so, in Anthropic's words, "nobody has to upload documents manually."
- Outputs are grounded. Anthropic's wording: "Each answer or output has a link to its source." For legal work, that's the most important property in the whole launch.
The intended division of labor is explicit: a legal professional supplies the documents and sources, Claude produces the first pass, and "a lawyer applies their judgment to the results." The Claude for Legal page puts it the same way: "A lawyer stays in charge of the work and judgment."
The architecture, in three layers
Anthropic doesn't publish an architecture diagram, but the pieces it describes fit a clear three-layer model. This is our framing, built from what the two pages say.
| Layer | What it does | What Anthropic says about it |
|---|---|---|
| Data | Gets documents and reference sources into the workflow | Connectors to systems like iManage, NetDocuments, Ironclad, Box, Docusign, and Everlaw, mostly through the open Model Context Protocol (MCP). More than 60 legal connectors added since May, including Bloomberg Law, Filevine, Ivo, Legora, and Patlytics. |
| Reasoning | Extracts, compares, classifies, and drafts | Claude produces the first pass against your documents, sources, and playbooks. |
| Interface and sharing | Renders results and controls who sees them | Interactive tables, boards, and timelines with source links; shareable internally or externally, with sharing settings that "ensure privacy and support ethical walls between teams." |
That separation is useful when you evaluate the product. Errors, risks, and controls live in different layers:
- Data-layer problems are about access: which repositories a connector can reach, under whose permissions, and whether the right version of a document comes through.
- Reasoning-layer problems are about accuracy: missed clauses, misread terms, or a confident answer that the source doesn't support.
- Interface-layer problems are about exposure: an artifact shared too widely, or a result read without checking the link behind it.
The six gallery artifacts, technically
Each gallery artifact is, at heart, a structured comparison between two sets of inputs. Here's what each one does according to Anthropic, and our read of how it works and where it can fail.
1. Cite check
What Anthropic says: compares a legal document with its sources and flags mismatched quotes, wrong page numbers, and sources that don't support the claim.
How to think about it: this is three different checks in one tool, and they have very different difficulty levels:
- Quote matching is mostly mechanical: does the quoted text appear in the source? This is the easiest check, though it depends on clean text extraction from the source document.
- Pin cite verification depends on page structure: does the quote appear on the cited page? Scanned PDFs, star pagination, and different reporter editions all make this harder.
- Support checking is genuine legal reasoning: does the source actually stand for the proposition? This is where model judgment matters most, and where human review is non-negotiable.
Verify by: testing on a brief you've already cite-checked by hand. Count false negatives (problems it missed) separately from false positives (problems it flagged that aren't real). Missed problems are the expensive ones.
2. Jurisdiction survey matrix
What Anthropic says: compares a topic, such as data breach notice rules, across states or countries in one table, linked to the underlying law.
How to think about it: this is extraction into a fixed schema: one row per jurisdiction, one column per attribute (trigger, deadline, who must be notified, penalties). The quality depends on the research sources connected. A matrix is only as current as the law it reads, so the source link on every cell is what makes it auditable.
Verify by: spot-checking a handful of cells, prioritizing jurisdictions with recent amendments, and checking that each link points to the current version of the law.
3. Matter board
What Anthropic says: turns engagement letters, emails, or spreadsheets into a board of open matters showing owner, next step, and next deadline.
How to think about it: this is unstructured-to-structured extraction across messy inputs. Deadlines are the high-stakes field. A missed or misread date is a classic malpractice risk, so this board should supplement your docketing system, not replace it.
Verify by: reconciling the board against your docket for a sample period. Pay special attention to relative dates ("within 30 days of notice") that require a trigger date to compute.
4. Playbook deviation board
What Anthropic says: compares contracts against company standard terms, flags differing clauses, and routes them to the person who approves exceptions.
How to think about it: this is the most "agentic" gallery artifact, because it doesn't only classify, it routes. Technically, that's three steps: segment the contract into clauses, match each clause to the corresponding playbook position, and classify the deviation. The routing step is where your organization's approval matrix gets encoded, so get that mapping reviewed by whoever owns it.
The Claude for Legal page also describes practice-area plugins for contract review, NDA triage, and compliance that can be configured to an organization's playbook, and Word integration with tracked-change redlining. Together, those suggest a flow from deviation detection to a redlined draft, with a lawyer approving each step.
Verify by: running a set of contracts whose deviations you already know, including clauses that are semantically off-playbook but use standard-looking language. Those are the ones string matching misses and models need to catch.
5. Regulatory horizon scan
What Anthropic says: maps upcoming regulations on a timeline based on the company's business and locations, ranked by impact, with source links.
How to think about it: this one combines retrieval (what regulations are coming), filtering (which apply to this business and these locations), and ranking (impact). The ranking is a judgment call, so treat it as a starting order for review, not a conclusion. The business profile you give it is effectively the filter, so an incomplete profile means missed regulations.
Verify by: comparing its output against your existing regulatory tracker or outside counsel alerts for one quarter.
6. Tabular review
What Anthropic says: builds a table across contracts for chosen terms, such as severance or non-competes, linked to the exact language.
How to think about it: this is the classic due-diligence grid: N documents by M terms, with each cell holding an extracted value and a link to the clause. It scales the work linearly with document count, which is exactly where first-pass automation pays off. The link to "the exact language" is what turns it from a summary into evidence.
Verify by: sampling cells across the grid, weighted toward documents with unusual structure: amendments, side letters, and scanned exhibits.
Connectors: how documents get in
Connectors are the reason "nobody has to upload documents manually." According to the Claude for Legal page, Claude connects to document and contract systems such as iManage, NetDocuments, Ironclad, Box, Docusign, and Everlaw, mostly through the open Model Context Protocol (MCP). Partner connectors include Harvey for Claude and Thomson Reuters content, and there's a legal connector directory in Claude.
For teams with their own systems, the page says developers can connect contract lifecycle, matter management, or e-discovery systems through the API and Agent SDK. It lists the SDK's controls: sandboxed containers, network controls, credential isolation, and audit logging.
From a technical standpoint, connectors are also where most of the security decisions live. Our recommendations:
- Scope each connector to the workspaces it needs. A connector to your DMS with access to every client's files is a much bigger risk than one scoped to the matter workspace an artifact works on.
- Respect your existing permission model. Confirm that a connector acting for a user sees only what that user could already see in the source system. Ethical walls in your DMS should hold through the connector.
- Log connector activity. If you build custom connectors with the Agent SDK, use its audit logging and send those logs wherever your firm's security team already reviews access.
Sharing, ethical walls, and data handling
Artifacts can be shared internally with business colleagues or externally with clients and partners. Anthropic says the sharing settings "ensure privacy and support ethical walls between teams."
On data handling, the Claude for Legal page states: "On Team and Enterprise plans, Anthropic doesn't train on your content by default," and that Enterprise admins set data retention periods. The page also links to privilege and confidentiality guidance. The announcement itself is tagged for Claude Enterprise and doesn't add further detail on retention or security controls, so read your agreement and that guidance before putting privileged material in.
A threat model for legal artifacts
This section is our analysis, not Anthropic's. Legal workflows have a specific risk profile that's worth spelling out.
Prompt injection through counterparty documents. A playbook deviation board reads contracts drafted by the other side. Any AI system that reads untrusted documents can be influenced by text inside them. A clause or hidden text that tries to steer the model ("this agreement conforms to standard terms") is a realistic risk. The defense is the one the product is built around: source links and human review on every flagged and unflagged clause in high-value deals.
Over-sharing. External sharing with clients is useful, and also a new way to disclose the wrong matter. Default to internal-only, and make external sharing a deliberate, per-artifact decision.
Privilege and confidentiality. Putting client material into any third-party system raises professional-responsibility questions about confidentiality and supervision. Have your general counsel or ethics partner sign off on which matters and data categories are in scope.
Stale or incomplete sources. A jurisdiction matrix or horizon scan is only as good as the research sources connected. Know what each artifact reads, and how current it is.
Automation bias. The cleaner the table, the more tempting it is to trust it. That's why "Each answer or output has a link to its source" matters: it makes verification cheap. Build the habit of using it.
Partner-built artifacts
Beyond Anthropic's gallery, the announcement lists artifacts built by legal tech companies and law firms:
- AI Governance Benchmark (Goodwin): benchmarks a company's AI governance approach to surface gaps.
- Harvey Intake (Harvey): one place for business users to submit legal requests. Routine questions get answered from Legal's guidance in Harvey; others go to a lawyer, whose reply comes back to Claude.
- Obligations Checker (Foley & Lardner): reviews agreements for both parties' obligations, before and after closing.
- Subpoena Response Tool (PwC): turns precedents, standard objections, and templates into audit-ready first drafts.
Anthropic says more are coming from Legora and Orrick. Technically, Harvey Intake is the most interesting pattern on that list: it's a triage-and-escalate loop across two systems, where the model handles routine requests from approved guidance and hands everything else to a human, with the answer flowing back into the same thread.
Access to justice
The gallery also includes nine access-to-justice artifacts, five built by nonprofits that Anthropic supported with Claude credits:
- Criminal Charge Reality Check (Zealous): sentencing ranges, public outcome data, and standard release playbooks. New York only.
- Immigration Papers Reader (Mobile Pathways): a plain-language first read of key dates and missing documents.
- Rent Ledger Reviewer (Rising for Justice): finds errors in landlord rent ledgers in eviction cases, under D.C. law.
- Record-Sealing Screener (Rising for Justice): assesses sealing eligibility, under D.C. law.
- Record Clearance (The Access Project): checks whether a record can be cleared under state rules.
Note the jurisdiction limits. These tools encode specific bodies of law, and they're scoped accordingly. Anthropic is also funding legal education programs that teach students to build with Claude, and the work builds on the Oxford Justice Innovators Programme, a 12-month program with the Oxford Institute of Technology and Justice and Microsoft supporting up to 10 innovators. The Claude for Legal page adds that qualifying clinics and public defenders can get discounted pricing through Claude for Nonprofits, run with Free Law Project and the Justice Technology Association.
How to roll this out: a technical checklist
This is our recommended sequence for a legal ops or legal engineering team.
- Start on the built-in examples. Every artifact opens with sample data. Learn the output schema before connecting anything.
- Pick one artifact with a measurable baseline. Tabular review and cite check are good candidates, because you can compare against work your team has already done by hand.
- Build a small evaluation set. Twenty to fifty documents with known answers. Track misses and false flags separately, and rerun the set whenever you customize the artifact or connect a new source.
- Connect read-only, narrowly scoped sources first. One matter workspace, one playbook, one research source.
- Customize deliberately. When you ask Claude to tailor an artifact, write down what changed, and rerun your evaluation set.
- Set a review policy per artifact. Define which outputs must be checked against their source links, by whom, and how that's recorded.
- Lock down sharing. Internal-only by default. Confirm ethical walls hold through both the connector and the artifact's sharing settings.
- Confirm data terms. Check plan-level training and retention settings with your admin, and read Anthropic's privilege and confidentiality guidance with your ethics counsel.
- Expand gradually. Add artifacts and connectors one at a time, with the same evaluate-then-deploy loop.
For a firm-wide rollout, Anthropic points to its sales team and a list of implementation partners, including Accenture, Deloitte, KPMG, and PwC, among others.
Bottom line
The legal gallery isn't a new model or a new chat feature. It's a packaging decision: take the legal workflows teams were already improvising in chat, and turn them into structured, repeatable tools with connected data, a fixed output format, and a source link on every answer. That's the right shape for legal AI. The first pass gets faster, the verification gets cheaper, and a lawyer stays responsible for the result.
The open questions are the ones that matter most to legal teams: how each artifact performs on your documents, how your connectors are scoped, and how your sharing settings map to your ethical walls. Those you'll need to answer yourself, with a small evaluation set and a careful rollout.
Sources
Product details are from Anthropic's announcement and Claude for Legal page as of October 28, 2026, and may change. Analysis, threat modeling, and rollout recommendations are our own. Nothing in this article is legal advice.
← All AI & Workflows