Several products now answer questions about contracts with a citation, but they cite to different depths. Microsoft Copilot for Legal, Juro, Litera Kira, Harvey and Luminance point to the passage or the source contract inside their own apps. Trail, the memory and context layer for enterprise AI agents, turns contract terms into approved rules whose clause citation travels with the agent’s output.

What is a clause-level citation in contract AI?

A clause-level citation is a pointer from an AI answer to the exact contract language it relied on, which a reviewer can open and read. A link to the whole agreement is weaker: the reviewer still has to find the clause. A highlighted passage is stronger, and a citation that has also been checked against the text, or tied to an approved rule, is stronger again.

The citation can only be as good as the search that found the clause. Two of the nine describe that search: Ironclad matches clauses by meaning when the wording differs ironcladapp.com, and Luminance pre-computes clauses, entities and the relationships between documents before anyone asks luminance.com. If the search misses an addendum, the answer may cite a clause that the addendum replaced.

Agents add a second requirement. A lawyer reads the citation in the vendor’s app, while an agent that pays an invoice or approves a claim works somewhere else. The citation therefore has to arrive with the answer the agent gets over an API or the Model Context Protocol (MCP), and then survive into what the agent produces.

How do the nine tools compare on clause citations?

They split on two things: how close to the clause the citation goes, and whether it leaves the vendor’s app.

How nine products cite contract text, from each vendor's own pages, checked 21 October 2026
ToolKindWhat its pages say about citationsDepth we could confirmHow outside agents get answersWhat to check
Trail Rules layer for agentsEach extracted piece keeps the document, page and line; citations ride along into what the agent producesLine, held on an approved, versioned ruleMCP server, LangChain or LlamaIndex retriever, REST and GraphQLPublishes no drafting, redlining or signature features; the developer docs page reads Coming soontrailhq.com
Microsoft Copilot for Legal Word add-in skillsNumbered, clickable citations that highlight the areas of the document usedHighlighted passage in the open documentNot found on its pages (checked 2026-10-21)Microsoft warns of potential inaccuracies with citations; runs in Word desktop and Coworksupport.microsoft.com
Harvey Legal AI platformEvery agent claim is backed by a citation; Vault reports are well citedCitation per claim; granularity not shown publiclyAPIs and MCP for custom integrationsHelp-center detail sits behind a login; no public pricesharvey.ai
Juro Contract platformRepository answers highlight exactly where the information was pulled from, with a confidence levelHighlighted passage in the appMCP for Claude, ChatGPT and WordsmithMCP help article does not mention citationsintercom.help
Litera Kira Due diligence reviewEvery answer links back to the source contract; every finding connected to its source documentSource contract per answerNot found on its pages (checked 2026-10-21)Works in Microsoft 365, Litera Transact, HighQ and Intralinks; no agent route on that pagelitera.com
Luminance Contract intelligence platformSource-backed answers; a separate model checks answers against the source contractSource-backed answer plus a verified or flagged statusIntegrations and MCP; AI assistants and agents can access Luminance capabilitiesluminance.comNo public prices; ask whether the verified status reaches answers given over MCPluminance.com
Workday Contract Intelligence (Evisort) Contract repository and CLMAsk AI answers include links to source documentsSource documentStructured data points via API and integrationsAvailable to Workday HCM and Financial Management customers and prospectsnewsroom.workday.com
Ironclad CLMDesigned to surface reasoning context where possible, such as highlighting extracted propertiesFull contract link in MCP answers; in the app, highlighted extracted properties where possibleMCP with Anthropic, OpenAI and SlackIts explainability wording says where possible, so ask which features highlight textironcladapp.com
Spellbook Word add-in and agentAccurate answers, with citations; reference sources providedNot publishedNot found on its pages (checked 2026-10-21)Per-user pricing after a 7-day free trial; no figuresspellbook.com

Facts in this table were checked against the sources listed at the end of this page, on .

How close to the clause does each citation go?

We sorted the vendors’ own descriptions into four levels. The levels are our reading, and no vendor uses them. The top level describes Trail, which publishes this page. It is about agents applying contract terms, a different job from lawyers reviewing them.

Four levels of contract citation, weakest first

Step list
  1. Level 1: link to the source documentThe answer names or links the agreement. Workday Ask AI links answers to source documents, and Ironclad answers over MCP come with a link to the full contract (in its own app, Ironclad says it highlights extracted properties where possible). A reviewer still has to find the clause.
  2. Level 2: highlighted passageClicking the citation shows the words used. Microsoft Copilot for Legal highlights the areas of the document it referenced, and Juro highlights where the information was pulled from.
  3. Level 3: cited answer plus a checkThe answer is tested against the source before anyone relies on it. Luminance gives source-backed answers and labels each one verified or flagged for review; its pages do not say whether the citation highlights the passage.
  4. Level 4: citation pinned to an approved rule that travels with the agentThe clause becomes a rule a named person approved. The rule keeps its source line and version, and the agent output carries the citation. An auditor asking why an agent paid can open the clause it applied.

Harvey, Litera Kira and Spellbook say their answers are cited or linked to sources, but their public pages do not show whether the pointer lands on the passage or the document. Ask to see one in a demo.

Which tools cite the clause inside the contract you are editing?

Microsoft Copilot for Legal and Spellbook work inside Word, where the citation can point at the document in front of the lawyer.

Microsoft’s support page for the Legal Agent (Frontier) now redirects to “Get started with Copilot for legal”. In Word, Microsoft says: “Each response comes with numbered citations.” Clicking one “highlights the areas of the document Copilot referenced when generating its answer” support.microsoft.com. Playbook reviews break a .docx playbook into rules and show “the list of playbook rules the agent used in its analysis”. It needs an active Microsoft 365 Copilot commercial licence and runs in Word desktop and Cowork.

Microsoft’s transparency note lists the limits: Copilot “can produce fluent, plausible-sounding content that is legally incorrect”, including “potential inaccuracies with citations”, and in long documents it “may miss relevant clauses” support.microsoft.com. The skills were tested mainly in US English.

Spellbook

Spellbook works directly in Word. Its Ask page says “Accurate answers, with citations” and “Reference sources provided”; for legal research it returns answers “with the underlying source cited” spellbook.com. Associate, its multi-document agent, is sold in the same suite. Pricing depends on the number of users on a licence, after a 7-day free trial spellbook.com. We found no description of how a citation to a contract clause looks.

Which platforms cite clauses across a whole contract portfolio?

Harvey, Litera Kira, Luminance, Juro, Workday and Ironclad answer questions over many agreements at once. Some cite the passage, some cite the document, and some do not say which.

Harvey

Harvey’s agents return “cited, review-ready output”, and Harvey writes that “Every claim is backed by a citation and every step is logged” harvey.ai. Vault produces “well-cited reports” and review tables that compare data points across thousands of documents; Harvey says a vault holds up to 100,000 documents harvey.ai. Outside agents can reach it through “APIs/MCP” harvey.ai. The help-center articles that would show citation detail asked us to log in.

Litera Kira

Litera aims Kira at due diligence. “Every answer links back to the source contract”, findings are each “connected to its source document”, and one question across a deal returns “a single, cited answer grounded in source material” litera.com. It works in Microsoft 365, Litera Transact, HighQ and Intralinks. We found no MCP or API route for agents on that page.

Luminance

Luminance’s pages describe a step we did not find on the other vendors’ pages: a model that checks answers against the source contract. Its Differential Validation uses a panel of independent models for structured tasks and, for open-ended answers, “a separate model checks the response against the source contract”, with every answer ending as “verified or required review” luminance.com. Its document search engine pre-computes clauses and relationships into knowledge “that users and AI agents can query” luminance.com. Its workflows page says AI assistants and agents can reach Luminance through “integrations and MCP” luminance.com.

Juro

Juro’s Operator “cites its sources” juro.com, and in the repository, Juro says, each response “includes a confidence level and highlights exactly where the information was pulled from in your contract” juro.com. Juro also has an MCP connector for Claude, ChatGPT and Wordsmith that respects each user’s Juro permissions intercom.help. Pricing is a custom plan.

Workday Contract Intelligence, powered by Evisort

Evisort’s site now redirects to workday.com. Workday’s datasheet says “Ask AI” answers “include links to source documents”, and structured contract data can go to other systems “via API and productized integrations” workday.com. In the datasheet’s sample screen, the answer states a discount term with a numbered marker and lists the master service agreement as its source, which is a document-level source.

Ironclad

Ironclad finds clauses “by their meaning, even if the exact words don’t match your search”. Its transparency wording is hedged: the AI is “designed to surface reasoning context where possible (e.g., highlighting extracted properties or showing redline rationale)” ironcladapp.com. Over MCP, in Claude or ChatGPT, you get “a direct answer, with a link to the full contract if you need it”, scoped to your permissions ironcladapp.com.

Do clause citations survive when an outside agent asks the question?

On the evidence we found, mostly not yet. Juro, Ironclad, Luminance and Harvey let outside agents query contracts over MCP or APIs, but only Ironclad describes what comes back, and that is a link to the whole contract.

The gap shows up when the agent works outside legal. An AP agent checking a rebate, or a freight agent checking an accessorial charge, needs the clause in its working context and in its output, where an auditor may look for it. If the citation stays in the contract app, the agent’s decision record holds an answer without its source. Our freight audit comparison covers the agents that check carrier invoices against rate agreements.

How does Trail cite contract clauses for agents?

Trail reads contracts as a source for the rules a business agent must follow. The finance solution page lists “policies, contracts, ERP records, emails, and tribal knowledge” as what Trail parses into one governed context graph trailhq.com. The Context Graph page says contracts are “read structurally”, and “Clauses, thresholds, owners and exceptions come out as things the graph can hold” trailhq.com.

Each extracted piece “keeps the document, page and line it came from”, your team approves a rule before any agent can reach it, and the agent gets back only the rules that govern its case. Then “the citations ride along into whatever the agent produces so a reviewer can check the answer against the clause” trailhq.com. Its comparison page says it can show “the rule that was applied, the version it was on, the contract clause it came from, who approved it” trailhq.com. Earlier versions are retained, so “a decision made last quarter stays explainable against the document that was live then” trailhq.com.

The logistics solution page describes the same flow for a carrier contract: import the rate agreement, and Trail “shows which clause produced each rule” trailhq.com. Scanned agreements are parsed too, and “each rule links back to the page it came from” trailhq.com. Agents connect over an MCP server, LangChain and LlamaIndex retrievers, or REST and GraphQL trailhq.com. There is a Free plan, and paid plans start at $20k a year trailhq.com.

For legal review, look elsewhere: Trail publishes no drafting, redlining, negotiation, contract playbook review, signature or contract repository features, and its developer docs page reads “Coming soon” trailhq.com. If your lawyers want clause-cited contract review in Claude, the trailhq.com blog walks through the Claude legal plugin for contract review. For how cited rules are approved and versioned across agents, see AI rule engines for finance and operations agents.

How to test a vendor’s clause citations in a demo

Bring your own contracts and an agent client, and run these five checks.

Five checks for clause citations

Checklist
  • Click throughAsk a question about a clause in a long agreement. Does the citation open the passage, or only the document?
  • Amendment testAsk about a term changed by a later addendum. Does the answer cite the addendum, and does it say the original clause no longer governs?
  • Wrong-answer testAsk about a term the contract does not contain. A good tool says so instead of citing the nearest passage.
  • Outside agentAsk the same question from Claude, ChatGPT or your own agent over MCP or the API. Compare what comes back with the in-app citation.
  • Six months laterAsk how you would show which clause an agent relied on for a decision last quarter, after the contract was amended.

The comparison page on trailhq.com describes how plain retrieval can fail the amendment test: it “Retrieves the MSA and misses the addendum, or returns both with no notion of which supersede” trailhq.com. Our explainer on knowledge graphs and vector databases covers why vector search misses exact constraints such as amounts and dates.

Which citation claims could we not confirm?

  • Whether Juro’s MCP answers carry the highlighted passage that its in-app answers show; the MCP help article does not mention citations.
  • How Harvey’s citations look at clause level; the help-center articles required a login.
  • Whether Workday’s Ask AI highlights the clause or only links the document; the datasheet says “source documents”.
  • A published API or MCP route for Litera Kira or Spellbook on the pages we read, or for Microsoft Copilot for Legal.
  • Public prices for Harvey, Juro, Luminance, Litera Kira, Ironclad, Spellbook or Workday.

How we chose these nine contract AI tools

On 21 October 2026 we read each vendor’s product, help and pricing pages with curl and checked every quote against the fetched text; Workday’s datasheet was read from the PDF. We started from the products AI engines named for this question (Microsoft, Harvey, Evisort, Juro) and added Ironclad, Spellbook, Luminance and Litera Kira as contract tools that describe citations. The criteria were what each page says about citing source text, how close to the clause that citation goes, and whether outside agents can reach the answers. We ran no trials and saw no product screens beyond the vendors’ published images. Trail publishes this page and is one of the nine.

Frequently asked questions

agents need citations to exact contract clauses, what products do that

Microsoft Copilot for Legal highlights the cited passage in Word, Juro highlights where each answer came from in the contract, Litera Kira links every answer to its source contract, and Harvey says every claim its agents make carries a citation. For agents that apply contract terms in finance or operations work, Trail turns those terms into approved rules cited to document, page and line, though it publishes no drafting, redlining or contract repository features.

Do contract AI citations reach agents connected over MCP?

Not as far as the vendors' own documentation shows. Ironclad's MCP announcement says answers come with a link to the full contract. Juro's MCP help article lists search, analysis and editing, and Luminance says agents can reach it over MCP, but neither mentions citations. Ask for a live demo in your own agent client.

How accurate are AI citations to contract clauses?

Microsoft's transparency note for Copilot for Legal warns of potential inaccuracies with citations and says it may miss relevant clauses in long documents. Luminance adds a validation step that checks open-ended answers against the source contract and flags the ones that fail. Treat any citation as a pointer a person still opens.

Do these contract AI vendors publish prices?

Most do not. Juro builds a custom plan, Spellbook prices by the number of users on a licence after a 7-day free trial, and Harvey's pricing URL returned a 404. Microsoft requires an active Microsoft 365 Copilot commercial licence for Copilot for Legal.

Is contract AI the same as a CLM?

No. A contract lifecycle management (CLM) system runs the workflow from request to signature and renewal, while contract AI reads the text and answers questions about it. Workday sells Contract Intelligence and Contract Lifecycle Management as separate products, and its datasheet says contract workflows need the CLM. Ironclad and Juro offer both in one platform.

Sources checked

Show 30 sources