Thomson Reuters Document Intelligence vs Zuva: how they compare in 2026
Thomson Reuters Document Intelligence and Zuva both sell contract analysis built on their own pretrained extraction models rather than on a general purpose language model. Document Intelligence, formerly ThoughtTrace and owned by Thomson Reuters since 2022, is a full application. Its models were trained by Practical Law attorney editors, and around them sit a document library, retrieval by meaning, obligation reminders, dashboards, a Word add in with playbooks and a feed into HighQ. Zuva, built by the team behind Kira Systems, sells the extraction itself. More than 1,400 pretrained fields run through a public API, a self serve review tool called Analyze, and a diligence service staffed by former large firm M&A lawyers. Zuva publishes its prices, from $10 a document in Analyze, lets the customer set retention down to none, and runs on Azure in the United States or Europe or in a private or self hosted deployment. Thomson Reuters publishes no price or hosting region for Document Intelligence, and its terms for the product let it train its products on customer data.
At a glance
All 15 axes, side by side
The same grid applied to every vendor in the index, graded from public sources. Hover a grade to see what the letter means on that axis.
AI Centrality
How much of the product is actually AI. Whether the machine learning is the mechanism the buyer is paying for or a feature layered onto conventional software, and whether the vendor is specific about which is which.
Document Intelligence is a contract repository and analysis platform built around pretrained extraction models. The models identify thousands of provisions across a suite of domain specific models, from dates, obligations and payment details in commercial contracts to curtailment rights, production guarantees and take and pay clauses in wind and solar agreements. They return results as structured fields that HighQ's help pages call Thoughts, the extracted provisions and sub provisions. The same models drive retrieval by meaning, intent and context across thousands of documents, automatic labeling of document types on bulk upload, and the AI generated table of extracted contents in the Word add in. Around the models sit features that work without them. These include relational libraries in place of folders, shared bookmarks of saved results, notifications and calendar reminders, dashboards, playbooks with preferred and fallback language, and a low code and no code workflow builder. Thomson Reuters markets no generative drafting, chat or summarization feature for Document Intelligence.
Machine learning is the product in every form Zuva sells. The API is a REST service for extracting structured data from contracts and classifying documents, and Zuva Analyze puts the same extraction behind a self serve interface. Diligence Reviews, the third offer, are sold as AI assisted contract analysis with lawyer review, so even the service is built on the extraction. The models come from Kira Systems, and generative AI now adds answers and summaries in Analyze. Remove the models and nothing is left: there is no document management, workflow or storage product underneath. The security overview says the API is not meant as a long term repository.
Citation Accuracy and Hallucination Disclosure
Whether the vendor publishes measured accuracy on citations and assertions, grounds output to primary sources, and says plainly what its system does when it does not know. Legal has a documented public record of fabricated citations reaching filed briefs, so an untested claim of accuracy is not evidence.
Document Intelligence extracts provisions from the customer's own contracts, so its output points to text in documents the customer holds rather than to legal authority. Thomson Reuters publishes no precision, recall or error figure for its extraction models and describes no test set. The product page promises truly accurate results based on the context and intent of the text. The April 2023 launch release quotes Thomson Reuters' chief product officer for legal technology saying automated review improves accuracy, with no measurement behind either claim. The 50 percent figure Thomson Reuters cites measures faster retrieval and review against manual methods, not accuracy, and the features page says contextual retrieval ensures nothing important is overlooked. Editor users can edit and confirm extracted provisions under section 5.2.2 of the Legal Product Specific Terms, and fields shown in HighQ are limited to those defined in the Document Intelligence application. The Thomson Reuters General Terms of September 2026 call AI output probabilistic and say it may at times be incomplete, inaccurate, out of date or biased.
Zuva's extraction returns the contract's own text. The extraction documentation describes each result as an instance of the requested text in the document, so a reader checks a finding against the clause it came from. The API page contrasts the fields with general purpose language models: the same input gives the same extraction every time. No recall, precision or error rate is published for the fields. The Analyze page says other systems give maybe 70 to 80 percent accurate results and that internal testing found faster review with comparable accuracy, but it gives no test set or figure. Analyze's generative summaries are not described at all. The out of the box fields were trained on publicly available English documents such as EDGAR filings.
Autonomy and Oversight Model
What the system decides on its own, what a lawyer must approve, and whether the vendor documents where the review point sits. A tool that drafts under review and a tool that files without one are different products and different risks.
Extraction runs on documents a user uploads or a HighQ site sends, and results flow into the library, dashboards, notifications and, for HighQ customers, AI Hub and iSheets. The Legal Product Specific Terms split users into Editor users, who upload and delete documents, create and modify tags and attributes, and edit and confirm provisions, and Viewer users, who have read only access. Section 3.3 of the General Terms of September 2026 says output must be reviewed by a licensed professional or someone with the right credentials before reliance, use or distribution. It makes the customer solely responsible for verifying the accuracy and legality of output. Thomson Reuters' AI Usage Policy of January 2026 bars using AI output for automated decision making, or as a substantial factor in decisions with legal or similarly significant effects, without that review. Playbooks in the Word add in carry escalation contacts for clauses outside approved positions. Thomson Reuters publishes no setting that holds extracted provisions for confirmation before they feed alerts or dashboards, and no point at which the system hands a document back to a person.
Zuva names human review as part of Diligence Reviews. The Diligence Reviews pricing page lists review by former large firm M&A lawyers as part of every engagement. No page says what the reviewers check, how much of each document they read, or how an extraction they reject is handled. Zuva Analyze is described in three steps, upload, extract and export, with nothing on who reviews the export or what a user should verify. The API leaves oversight entirely to the developer building on it. The product is extractive rather than agentic: it returns text found in the document and does not draft, send or decide anything. A missed provision produces no output to review, and nothing published says how the service or the software guards against that. The API page says the same input gives the same extraction every time.
Operational and Outcome Evidence
Named, dated evidence that the product works in production at real firms or legal departments. Case studies with figures and identified customers count. Unattributed testimonials and launch announcements do not.
Tim Custer, a senior vice president at Apache, is quoted on the product page saying work that would have taken a team days, weeks or months now takes minutes and hours, with no figure and no date. A 2022 customer infographic carries anonymous quotes reporting a project cut from two years to two months and a consent to assign clause found in two minutes after a month of manual review. Others report work of 20 to 25 days done in a few days and a $4 million find. The corporate legal page quotes an unnamed M&A general counsel on saving hundreds of hours of manual review. Thomson Reuters says customers have cut information retrieval and review time by more than 50 percent against manual methods, a figure first published in the April 2023 launch release. It also says the models received more than 15,000 hours of training and maintenance in the second half of 2022. No figure is tied to a named customer or comes with a method, and the newest dated customer material is from 2023.
Zuva names Microsoft's legal team as an API customer. The API page says Microsoft's legal team used Zuva inside Power Automate and reports a reduction of roughly 75 percent in time spent on routine contract requests. Its target is to go from a three day service level to about three minutes. No date, period, volume or method stands behind the figure, and the target is stated alongside the result. The Analyze page features a case study of an in house lawyer at a global motorsports company and reports internal testing of 40 to 80 seconds a page against about three minutes manually. The claim that 18 of the world's top 25 M&A law firms used the same technology describes Kira Systems at its sale, not Zuva's customers. Zuva was formed when Kira Systems was sold to Litera in 2021, taking a copy of the underlying AI and about 30 of Kira's staff.
Privilege and Confidentiality Posture
How client confidences are handled: attorney client privilege and work product treatment, segregation of one client matter from another, whether client data trains any model, and what the vendor commits to in writing rather than in marketing.
Section 5.4 of Thomson Reuters' Legal Product Specific Terms, version 1.0 of July 2026, lets Thomson Reuters use, store, copy, share, transfer and process customer data to provide Document Intelligence. Given the machine learning nature of the service, it may also use customer data to train, update, modify and adapt its products. The resulting Machine Learning Adaptions belong solely to Thomson Reuters, and the customer grants a royalty free, irrevocable and perpetual license to its data to secure those rights. Deidentified data Thomson Reuters collects or generates is its sole property, and the customer may have no access to it. The General Terms bar training of generative models and large language models on customer data, and the Legal AI Product Specific Terms bar training of generative or foundational models. Both rank below the Legal Product Specific Terms, and both address generative models, while Document Intelligence runs extraction models. The General Terms treat customer data as confidential information during the term and for five years after it ends. Privilege, work product and separation between clients or matters go unaddressed.
Zuva's DocAI and AI Trainer License Terms set out who owns customer material and how Zuva may use it. Under the license terms the customer owns its documents, customer data and outputs (section 6), and Zuva treats them as confidential information held to at least a commercially reasonable standard of care (section 8). Section 2(g) confines Zuva's improvement right to usage data, defined as information about access to and use of the service. The security overview says model storage does not rely on the content of the documents used to train custom fields, and only the creating token can reach an account's data. No express training prohibition appears in the agreement. Nothing treats privilege or work product, and nothing covers what the generative AI provider behind Analyze summaries may keep. Custom fields are trained on the customer's own documents.
UPL and Professional Responsibility Posture
Whether the vendor is clear that it supplies a tool rather than legal advice, who its audience is, and how it addresses unauthorized practice of law, competence and supervision duties, and jurisdiction limits. ABA Formal Opinion 512 is the reference point. Where the advice line is not the duty a product raises, the axis is read through the nearest professional duty it does raise: judicial conduct rules and the reviewing duty for products sold only to courts, and the duty to bill for time actually spent for products that draft time entries.
Section 14.3 of the Thomson Reuters General Terms, version 6.0 of September 2026, says Thomson Reuters provides its services for informational purposes and not as legal, financial, tax, accounting, compliance or other professional advice. Section 3.3 requires output to be reviewed by a licensed professional, or someone with the right credentials, before reliance, use or distribution. The AI Usage Policy of January 2026 bars using AI output to assist in giving legal advice unless such a professional reviews it first, and bars presenting output as solely human generated. Document Intelligence's marketing invites law firms and in house teams to advise with confidence. It also sells to buyers outside the legal function, including corporate asset management teams, contract managers and renewable energy developers, owner operators and investors, with no separate disclosure for them. Nothing in Document Intelligence material addresses competence or supervision duties, and jurisdiction limits are unstated.
Diligence Reviews are sold in place of law firm diligence work, to buyers who are not all lawyers. The home page and the review service's pricing page offer sell side diligence and disclosure schedules with lawyer review, to buyers including chief financial officers and private equity firms. No page states whether that work is legal advice, whether Zuva is a law firm, whether a lawyer and client relationship arises, or who is responsible for the disclosure schedule a seller signs. The accuracy disclaimer in the license terms limits output rather than drawing the advice line. Analyze is addressed to legal teams and the API to developers, and neither says where its output stops and a lawyer's judgment begins. Under the Kira sale, Zuva will not sell interface based contract analysis directly to law firms until later in 2026.
AI Governance and Bias Disclosure
Published governance over model behavior: who owns it inside the vendor, what is tested before release, and what is disclosed about disparate output across matter types, parties, or populations.
Thomson Reuters publishes data and AI ethics principles, including commitments to strive to keep human involvement in its AI products and to treat people fairly. Its Data and AI Governance and Security Program whitepaper of March 2025 places governance with a Responsible AI and Data Trust team. It requires AI models used for a business purpose to be registered in a central Responsible AI and Product Compliance Hub. It also runs data impact assessments and model risk assessments, with ethics controls covering human oversight, fairness and automated decisions. Section 3.2 of the General Terms commits Thomson Reuters to policies on bias, accuracy, explainability, transparency and human oversight in developing its AI. Practical Law attorney editors build, train and maintain the Document Intelligence models. Thomson Reuters' ISO/IEC 42001 certificate, valid from August 2026 to March 2028, covers Westlaw, CoCounsel Essentials, CoCounsel Tax, Audit and Accounting, CLEAR Investigate and Practical Law, and does not cover Document Intelligence. No testing result, bias finding or named owner is published for these models.
Zuva states where its training data comes from and publishes research on how its models work. The Analyze page describes the fields as pretrained and tested. The science page lists about ten research papers by Zuva's research team on document review, interpretability and how lawyers work with AI, published on the principle that research advances through free dissemination. Zuva publishes no responsible AI policy and no named person accountable for model behavior. Nothing describes what is tested before a field or model is released, and nothing reports how extraction quality varies across document types, languages or drafting styles. The out of the box fields were trained on publicly available English documents such as EDGAR filings, and custom fields on the customer's own documents.
AI Safety and Data Stewardship
Retention, deletion, access control, and what happens to prompts and documents after they are processed. Whether the vendor states its subprocessors and its incident practice, or leaves the buyer to assume.
Thomson Reuters' Data Security Addendum, version 2.0 of November 2023, commits to notify the customer of a security breach without undue delay and within 72 hours of discovery, followed by a root cause analysis. Customer data is encrypted at AES 256 in transit and at rest, and user credentials are never stored in clear text. The Data Processing Addendum, version 3 effective February 2025, deletes customer personal data on request, except where law or Thomson Reuters' retention policies require it to be kept. Data kept on those grounds is isolated from further active processing. It points to subprocessor lists on Thomson Reuters' web pages, where product lists exist for Casetext, CoCounsel Core and Practical Law and none for Document Intelligence. The enterprise wide list of January 2022 names 20 providers, among them Amazon Web Services, Microsoft, Pendo and Salesforce. Editor users can delete documents, and customer data may pass to Thomson Reuters' service providers and to partners the customer enables. No retention period for documents or extracted provisions is published.
Zuva's security overview and license terms set out retention, deletion, encryption and segregation. The customer sets how long uploads stay and can delete them early, access to each account's data is limited by token, and the security overview describes encryption in transit and at rest. Section 8(d) of the license terms commits Zuva to physical, managerial and technical safeguards and refers to a separate data protection addendum. No named subprocessor list is published, and no incident notification period is stated. The privacy policy covers registration and site tracking and does not address uploaded documents. Uploaded documents expire after seven days by default, and the expiry can be set from one hour to fourteen days.
AI Liability and Recourse
What the vendor stands behind contractually when its output is wrong. Indemnities, caps, carve outs, insurance, and whether any of it is published or only reachable through a negotiated agreement.
The Thomson Reuters General Terms, version 6.0 of September 2026, cap each party's yearly liability at the amount payable for the affected service in the 12 months before the first event giving rise to the claim. In the first year the cap is 12 times the average monthly charge. Breaches of confidentiality, data protection or information security obligations carry a cap of twice that amount. Lost profits, revenue, goodwill and indirect, consequential and punitive damages are excluded. Section 3.2 says Thomson Reuters is not liable for the accuracy, legality or reliability of any output, or for decisions taken on it. Thomson Reuters defends the customer against third party intellectual property claims about the services, excluding combinations with customer data and unauthorized changes. It warrants that the services conform in all material respects to the documentation, with repair, modification or replacement as the remedy and termination if that fails. No indemnity or warranty covers extracted provisions, and no insurance is stated.
Zuva's license terms set a liability cap with carve outs and an intellectual property indemnity, and disclaim the accuracy of results. The DocAI and AI Trainer License Terms and Conditions, last updated 11 May 2022, are governed by Ontario law. Section 7(a) commits Zuva to indemnify the customer against third party claims that the services, used as the agreement allows, infringe intellectual property rights. Section 10 caps each party's liability at the amount paid in the six months before the claim, and carves fraud, willful misconduct and intellectual property infringement out of the cap. Section 9(b) disclaims all other representations and warranties, including any warranty that results will be accurate or reliable. These terms cover the API and the software. The Diligence Reviews pages publish no engagement terms, so nothing says what Zuva stands behind when a disclosure schedule prepared with its lawyers is wrong.
Practice Systems Integration Depth
How deeply the product reaches into the systems legal work already lives in: document management such as iManage and NetDocuments, Word and Outlook, contract lifecycle management, matter management, e-billing, and court filing systems.
Document Intelligence connects to HighQ through the HighQ AI Hub, set out step by step in HighQ's help pages. A HighQ system administrator enters a client ID, secret token and application URL generated under API clients in Document Intelligence, over the API at api.thoughttrace.com. Files sent from HighQ come back as Thoughts, the extracted provisions and sub provisions, and Facts, the metadata users assert, in AI Hub and iSheets. The customer must hold both products, each Document Intelligence workspace connects to one HighQ instance, HighQ's IP addresses must be allowed by Document Intelligence support, and HighQ fields are limited to those defined in Document Intelligence. A Microsoft Word add in compares drafts against a standard template, applies playbooks and alerts on redlines and comments. A low code and no code workflow builder, embedded from Workato under Thomson Reuters' third party terms, offers prebuilt connectors to common systems that are not named. No connection to iManage, NetDocuments, SharePoint or Salesforce is described.
Zuva's API documentation is public and written for developers. The documentation covers file submission, OCR, classification, language detection, field extraction and custom field training, with a quick start, a Python SDK tutorial and a public API reference among the guides. A tutorial covers the Zuva Power Automate connector. That is enough for an engineer to build the extraction into another system. Analyze exports results to Excel or Word for analysis across documents. No connector to a document management system such as iManage or NetDocuments, a CLM or a matter management system is published. Nothing describes what a Power Automate flow moves back beyond the extraction results, so integration depth depends on engineering by the customer. The API is also offered in a private cloud or self hosted on Azure or AWS for an added fee.
Deployment Model and Data Residency
Where the software runs and where the data sits. Multi tenant cloud, single tenant, private deployment, on premises, and whether region of residence is a published option or an enterprise conversation.
Document Intelligence is delivered as software as a service from app.thoughttrace.com, the ThoughtTrace domain it kept after the 2022 acquisition. HighQ's help pages give app.thoughttrace.com/myworkspace as an example application address, one per Document Intelligence workspace, and api.thoughttrace.com as the API address. Thomson Reuters publishes no hosting provider, region, tenancy model or residency option for Document Intelligence. Its Data Processing Addendum covers international transfers of personal data across Thomson Reuters services under applicable data protection law, with separate terms for European personal data. Files a HighQ customer sends for processing leave HighQ for the Document Intelligence application, whatever region the HighQ instance runs in. No on premises or private cloud option is described, and where extraction runs, as distinct from where documents are stored, is unstated.
Zuva runs as a shared cloud service, in a private cloud or self hosted. The API servers run on Microsoft Azure in the United States and Europe, according to the security overview. The API is also available in a private cloud, including a different Azure region, or for self hosting on Azure or AWS for an added fee. The license terms place hosting, in section 2(d), in the jurisdiction the customer selects and user authentication in Canada or the European Union, and the FAQ confirms hosting on the Zuva cloud or on premises. For extraction, storage and processing locations are clear across the shared cloud, a private cloud and the customer's own infrastructure. Nothing states where the generative AI behind Analyze answers and summaries runs, or whether it stays within the selected region. The security overview says the API is not meant as a long term repository.
Security Certifications and Trust Center
Independent attestation a buyer can pull without a sales call: SOC 2, ISO 27001, penetration test summaries, a trust center with current reports and named scope rather than a badge image.
Thomson Reuters' trust center publishes security profiles for CoCounsel, HighQ, Practical Law and other legal products, and none for Document Intelligence. Its ISO/IEC 27001 certificate, CERT 001130 issued by MSECB and valid to October 2028, lists products in scope by site, including HighQ, CoCounsel and Contract Express, and does not list Document Intelligence or ThoughtTrace. The ISO/IEC 42001 certificate, CERT 002034, valid to March 2028, also leaves it out, and no SOC 2 report is stated for the product. Section 3.2 of the Data Security Addendum lets a customer review an executive summary of any independent assessment or certification Thomson Reuters makes available for a service, and section 3.1 allows one security questionnaire a year. Section 2.9 says Thomson Reuters or a third party it appoints may periodically run penetration tests, with a summary of the results available on written request. Systems are monitored for threats and vulnerabilities on an ongoing basis.
Zuva states SOC 2 Type II certification across its site. The home page, the privacy policy footer and the API security overview all state the certification, and the overview's SOC2 compliance section says the Zuva API has obtained it. No auditor, report period, scope statement or route to request the report appears on those pages, the Analyze page or the documentation index, and no trust portal is linked. The overview does describe controls in some detail: TLS 1.2 in transit, 256 bit encrypted storage, disaster recovery, security testing and response, and personnel security, under named information security policies. The overview also says each account's data is permissioned so that only the token that created it can reach it.
Model Supply Chain Disclosure
Which models sit underneath, whose they are, where they run, and whether the vendor commits to telling customers when that changes. A legal buyer inherits every dependency it cannot see.
Document Intelligence runs on Thomson Reuters' own extraction models, described as a suite of domain specific models pretrained by Practical Law attorney editors and maintained by Thomson Reuters with no upkeep from the customer. Thomson Reuters says the editors spent more than 15,000 hours building and training them in six months, and that it continues to grow and scale its model training. No third party model provider, large language model or model version is named for Document Intelligence. Thomson Reuters' generative AI third party terms page lists Amazon Web Services, Anthropic, Google, Microsoft and OpenAI across its portfolio without tying any of them to this product. Where the models run is unstated. Section 5.4 of the Legal Product Specific Terms lets Thomson Reuters update, modify and adapt the models with customer data, and nothing commits it to tell customers when a model changes.
Zuva's extraction models are its own, carried over from Kira Systems. The about page says Zuva's AI is the technology developed at Kira Systems since 2011 and retained at the sale. Its research team works on the underlying models and on wrappers around general purpose AI. The Analyze page calls the fields purpose trained machine learning enhanced with generative AI for structured answers and summaries. Custom fields are trained on the customer's own documents, and the extraction models run on Azure in the region the customer selects. Nothing names the provider or model producing the generative answers and summaries, where it runs, or what that provider keeps. No commitment to tell customers when a model or provider changes is published. Custom field training is part of the API and runs under the DocAI and AI Trainer License Terms.
Commercial Transparency
Whether a buyer can learn what this costs without entering a sales process: published rates, the unit being charged, what sits behind an enterprise tier, and what implementation adds.
Thomson Reuters publishes no price for Document Intelligence. Buying runs through a free demo request, a contact form and a sales phone line, and a representative replies within one business day. The Legal Product Specific Terms say the scope and parameters of users are set in the ordering document and define two user types, Editor users and Viewer users, which sets the shape of the subscription without a figure. No tiers, modules or add on prices are named. The HighQ integration requires a Document Intelligence subscription alongside HighQ, and the Workato connectors run under Workato's embedded terms. SoftwareOne's software marketplace lists Document Intelligence with Thomson Reuters as vendor and no price. No implementation or training fee is stated, although Thomson Reuters runs product training and support for Document Intelligence customers.
Zuva prices its self serve products and its review service on its own pages. Analyze costs $10 per document with a $5,000 minimum, the API is free for small volumes and then charged by the document and by the field, and most Diligence Reviews deals run from $10,000 to $75,000. The pricing structure charges Analyze by the document, charges the API by the document and by the field after a free tier, and sets Diligence Reviews in fixed fee bands with stated turnaround times. Enterprise volume, private cloud, self hosting and rush delivery are left unpriced. From the published figures, the cost of a review project, a batch of contracts or an integration can be worked out without contacting Zuva.
Firm and Practice Coverage
Who the product is actually built for. AmLaw, midlaw, small firm and solo, in house departments, government and courts, and which practice areas are supported rather than merely claimed.
Document Intelligence names three buyer groups, law firms, in house counsel and legal operations, and corporates from energy to finance, including asset management teams and contract managers. Thomson Reuters runs a dedicated page for corporate legal departments and one for renewable energy. The renewable energy page lists power purchase agreements, interconnection agreements, operations and maintenance contracts, and site leases and option agreements. It counts more than 450 out of the box provisions for wind and solar agreements, from curtailment rights to forecasting requirements. M&A due diligence is the lead law firm use, the subject of the April 2023 launch release, alongside commercial contract review and drafting in Word. The demo forms take law firms by size, corporate legal departments, financial institutions, other businesses and government. Languages, jurisdictions and contract types outside commercial and energy agreements are not stated.
Zuva sells three offers to three kinds of buyer. Diligence Reviews are for the sell side of M&A, covering pre screens, disclosure schedules and execution status, for buyers such as chief financial officers, private equity firms selling portfolio companies and M&A advisors. Zuva Analyze is for legal teams and in house lawyers working on M&A, contract portfolio management and compliance. The API is for developers building contract automation into their own products. The field library covers more than 1,400 provisions and more than 225 document types. Zuva states two limits. Under the Kira sale, Zuva agreed not to sell interface based contract analysis directly to law firms until later in 2026. And the out of the box fields were trained on publicly available English documents. Nothing states which governing laws or contract types beyond that population the fields have been tested on. Analyze exports its results to Excel or Word.
The 12 legal signals, side by side
Recorded rather than graded. These are the questions a practitioner has to answer before a tool touches a client matter, and the answers are taken from public material only.
Client Data in Training
Can material a lawyer puts into this product be used to train a model?
Section 5.4 of Thomson Reuters' Legal Product Specific Terms, version 1.0 last modified in July 2026, applies to Document Intelligence subscribers. Given the machine learning nature of the service, Thomson Reuters may use customer data to train, update, modify and adapt its products, and the resulting Machine Learning Adaptions are solely Thomson Reuters property. The customer grants a royalty free, non exclusive, irrevocable and perpetual license to its data to secure those rights, and deidentified data belongs to Thomson Reuters.
No opt out is offered. The General Terms and the Legal AI Product Specific Terms bar training of generative or foundational models and large language models on customer data, and both sit below the Legal Product Specific Terms in its order of precedence.
A public vendor page says neither Zuva nor any other customer may use a customer's fields and data without permission, and the agreement does not address training. The API security overview states the commitment, and the standard fields were trained on public filings. The license terms (11 May 2022) never name training; the only reuse right they give Zuva covers usage data about access to the service, so the commitment is a policy rather than a contract term.
Prompt and Output Retention
How long does the product keep what a lawyer typed, and can that be set to zero?
Section 3.6 of the Thomson Reuters Data Processing Addendum, version 3 effective February 2025, deletes customer personal data on request, except where law or Thomson Reuters' data retention policies require it to be kept. Retained data is isolated and protected from further active processing. Editor users can delete documents in Document Intelligence, and no retention period is published for documents, extracted provisions, facts users assert or backups.
Under the Legal Product Specific Terms the license to customer data that secures Thomson Reuters' Machine Learning Adaptions is perpetual and irrevocable, and deidentified data stays with Thomson Reuters. The Data Security Addendum adds encryption at AES 256 in transit and at rest.
The customer sets retention, and keeping nothing is offered. The API page says retention can be none, a week or whatever the customer needs. The security overview says uploaded documents expire after seven days by default, that the expiry can be set from one hour to fourteen days, and that files can be deleted earlier through the Files API without restoration. The Analyze page says documents are deleted when the customer chooses.
Ethical Walls and Matter Segregation
Does retrieval respect the firm’s ethical walls, or can the model read across them?
The Legal Product Specific Terms define two user roles for Document Intelligence. Editor users upload and delete documents, manage tags and attributes and edit and confirm provisions, and Viewer users have read only access, with the customer administering accounts. Relational libraries group documents without folders, and bookmarks of saved results can be shared with other users. Each Document Intelligence workspace connects to a single HighQ instance, where HighQ's own site and folder permissions govern who sees returned provisions.
No ethical wall, information barrier or restriction by matter or client is described for Document Intelligence, and nothing says whether access follows the permissions of a document management system.
Zuva documents its own access model at the level of the API token. The security overview says all data in an API client account is stored and permissioned so that only the token that created it can access it. Other customers and Zuva itself are barred from using a customer's data without consent. Nothing describes walls between matters inside a customer or any link to document management permissions.
Third Party Request and Subpoena Notice
If someone subpoenas the vendor for a firm’s data, does the firm hear about it first?
Section 13.2 of the Thomson Reuters General Terms, version 6.0 of September 2026, lets either party disclose confidential information when required by law, regulation, subpoena or court order. Where legally permitted, the disclosing party must first give the other prior notice and reasonable assistance to contest or limit the disclosure. Section 1.3 names customer data as confidential information. Section 3.4 of the Data Processing Addendum refers requests and complaints from regulators about customer personal data to the customer, to the extent permissible.
No transparency report is published, and the Data Security Addendum commits to notify a security breach within 72 hours.
The license terms commit to prompt written notice of a compelled disclosure where the law permits. Section 8(c) covers subpoenas, civil investigative demands and requests in legal proceedings for confidential information, which includes the customer's documents and data, and requires reasonable efforts to preserve confidentiality and cooperation with a protective order. Section 2(d) places hosting in the jurisdiction the customer selects. Zuva publishes no transparency report.
Primary Law Corpus Provenance
Where does the law in this product come from, and does the vendor have the right to use it?
Document Intelligence's answers come from the customer's own contracts, extracted by models that Practical Law attorney editors built, trained and maintain. Thomson Reuters says the editors devoted more than 15,000 hours to training and maintaining the models in the second half of 2022. It markets the result as AI powered by Practical Law, its own content service with more than 650 attorney editors. The documents the models learned from, and the rights basis for them, are not described.
Playbooks in the Word add in hold the customer's approved clause positions. Section 5.4 of the Legal Product Specific Terms adds customer data to the material Thomson Reuters may use to adapt its products.
Zuva's fields work on contracts the customer supplies rather than on primary law. The API FAQ says Zuva's legal knowledge engineers train the fields on publicly available documents such as EDGAR filings, all in English. Zuva names no collection beyond that example and states no rights basis or update cadence for it. Custom fields are trained on the customer's own documents, and the generative layer that adds answers and summaries in Analyze is not described.
Good Law Verification
Does the product tell you when the authority it just cited has been overruled?
Document Intelligence extracts provisions and data from contracts and does not cite case law or legislation, so the subsequent history of legal authority sits outside the product. Thomson Reuters' citator, KeyCite, belongs to Westlaw and CoCounsel, and no Document Intelligence material connects it to extracted provisions. Playbooks in the Word add in compare clauses with the customer's approved language rather than with authority.
Nothing describes how a provision tied to a repealed rule, a changed regulation or an outdated market position is flagged.
Zuva extracts provisions from contracts and classifies documents, and it does not retrieve or cite case law or statutes, so a check of subsequent history sits outside the product. Its fields return the contract's own text rather than legal authority. Nothing describes how a provision that depends on a rule since changed is flagged in an extraction or in the generative summaries in Analyze.
Refusal and Uncertainty Behavior
What does the product do when the answer is not in the corpus?
Thomson Reuters publishes nothing on what Document Intelligence shows when a provision is missing from a contract or the models cannot place it, and no confidence indicator is described. Editor users can edit and confirm provisions under the Legal Product Specific Terms, which gives a manual correction path, and HighQ fields carry only what Document Intelligence returns. The General Terms say AI output may at times be incomplete, inaccurate, out of date or biased, and that it must be reviewed by a licensed professional or someone with the right credentials before reliance. Marketing promises that contextual retrieval ensures nothing important is overlooked.
The extraction returns nothing for a field it does not find. The field extraction documentation says each requested field comes back with zero or more extractions, each an instance of the text in the document, and the API page says the same input gives the same extraction every time. No confidence measure is described for an extraction, and Zuva describes no abstention behavior for the generative answers and summaries in Analyze.
Fabricated Citation Record
Does a public court record exist addressing fabricated or hallucinated legal citations in output from this product?
The AI Hallucination Cases database maintained by Damien Charlotin, which records court decisions worldwide that address hallucinated AI content and the tool involved where known, has no entry naming Document Intelligence or ThoughtTrace. Document Intelligence extracts provisions from contracts and does not draft court filings or cite authority.
The AI Hallucination Cases database maintained by Damien Charlotin, which records court decisions worldwide that address hallucinated AI content and the tool involved where known, has no entry naming Zuva or Zuva Analyze. Zuva extracts contract text for diligence and portfolio review rather than drafting court filings, so its output does not ordinarily reach a brief.
Bar Guidance Alignment
Has the vendor engaged in public with the ethics opinions its buyers are bound by?
Thomson Reuters' AI Usage Policy of January 2026 bars using AI output to assist in legal advice, or as a substantial factor in decisions with legal effects, without prior review. The reviewer must be a licensed professional or someone with the right credentials. It also bars presenting output as solely human generated. The General Terms say the services are not legal advice and leave verification of output with the customer.
No bar opinion, ethics rule or professional conduct guidance is named in Document Intelligence material, including for law firms using it on client due diligence. Thomson Reuters' AI principles commit in general terms to human involvement and fair treatment.
Zuva's pages and license terms name no bar opinion, ethics rule or regulator guidance, and they do not refer to lawyers' professional or ethical obligations. That includes Diligence Reviews, the sell side service delivered with former large firm M&A lawyers to buyers such as chief financial officers and private equity firms. The license terms disclaim any warranty that results will be accurate or reliable. Analyze is addressed to legal teams and in house lawyers without guidance on supervising its output.
Billing and Fee Posture
Does the vendor address what happens to the bill when the work takes an hour instead of six?
Document Intelligence's pitch to law firms promises to reduce unbillable hours, and the April 2023 launch release says pretrained models matter to firms because training AI models is nonbillable work. Customers are said to cut retrieval and review time by more than 50 percent, and the corporate legal page quotes an M&A general counsel on saving hundreds of hours. Due diligence for a deal, the lead law firm use, is work billed to clients, and no Document Intelligence material addresses how AI assisted review is recorded, billed or disclosed to the client. In house and corporate buyers sit outside a fee relationship, while law firm use sits inside one.
The products are bought by in house legal teams, developers and companies selling a business, none of which bills a client for the work. Under the Kira sale, the interface products are not yet sold to law firms. The review service is marketed at about 80 percent below traditional law firm costs on a mostly fixed fee with no hourly billing, a saving aimed at the buyer's own cost.
Outside Counsel Guideline Readiness
Can a firm get this vendor through a client’s AI clause without a bespoke negotiation?
Thomson Reuters' Data Processing Addendum and Data Security Addendum can be read without an agreement, and the processing addendum points to subprocessor lists on Thomson Reuters' web pages. Those pages carry lists for Casetext, CoCounsel Core and Practical Law and none for Document Intelligence, and the enterprise wide list dates from January 2022. No model provider is named for Document Intelligence, and the trust center has no Document Intelligence profile.
Under section 3.1 of the Data Security Addendum a customer may ask once a year for a completed security questionnaire, and under section 3.2 may review executive summaries of independent assessments where they exist. The Legal Product Specific Terms let Thomson Reuters pass customer data to its service providers and to partners the customer enables.
Zuva's security overview, readable without an agreement, describes Azure hosting in the United States or Europe, customer set retention, token level segregation and encryption, and Zuva states SOC 2 Type II certification across its site. The data protection addendum that section 8(d) of the license terms refers to is not public. Zuva publishes no subprocessor list and does not name the provider behind the generative answers and summaries in Analyze, so the processing detail sits in an agreement signed with Zuva.
Court Disclosure Support
If a judge’s standing order requires an AI disclosure, can the product produce one?
Document Intelligence produces extracted provisions, dashboards and alerts from contracts rather than court filings. Editor users confirm provisions, and bookmarks save result sets that update as contracts are added, but no audit log, record of the model used or export showing how a provision was produced is described. HighQ customers receive Thoughts and Facts in iSheets, where HighQ's own audit logs apply. Thomson Reuters publishes no disclosure template or guidance on certifying AI assisted review, including for due diligence findings that reach deal documents or disputes.
Zuva extracts contract provisions for diligence and portfolio review and exports the results to Excel or Word, and its output is not a court filing. Nothing describes a record of which model produced an extraction, which fields a person reviewed, or how Diligence Reviews lawyers sign off their work. Uploaded documents expire after seven days by default unless the customer sets another period. Zuva offers no disclosure template.
The questions both sides leave open
Derived from the records above rather than written, so it cannot favor either vendor. Take these into both conversations and ask each side the same question.
- Good Law Verification
- Court Disclosure Support
Which one fits
Choose Thomson Reuters Document Intelligence if
- You want contract repository and analysis in one application. Document Intelligence pairs its extraction with a document library organized without folders, retrieval by meaning and context across thousands of documents, obligation and renewal reminders, dashboards, and a low code workflow builder with prebuilt connectors to other systems.
- You already run HighQ or draft in Word. HighQ's AI Hub sends files to Document Intelligence and receives extracted provisions and facts in iSheets, set up with an API client ID and secret. The Word add in compares a draft against a standard template, builds a table of extracted contents and applies playbooks with fallback language and escalation contacts.
- You review energy agreements or want models trained by lawyers. Practical Law attorney editors pretrained the models, and the renewable energy offer covers power purchase agreements, interconnection agreements and site leases, with more than 450 out of the box provisions from curtailment rights to forecasting requirements. Law firms are a named buyer group.
Choose Zuva if
- You want to know the cost before you talk to sales. Zuva Analyze costs $10 a document with a $5,000 minimum, and the API is free for small volumes and then charged by the document and by the field. Most Diligence Reviews deals run from $10,000 to $75,000, with stated turnaround times.
- You want to control how long documents stay and where they run. The customer sets retention, with none available, and uploads expire after seven days by default. The API runs on Azure in the United States or Europe, in a private cloud in another Azure region, or self hosted on Azure or AWS for an added fee.
- You are building extraction into your own software. Zuva's public API documentation covers OCR, classification, language detection, field extraction and custom field training, with a Python SDK and a Power Automate connector. Each field returns zero or more instances of the contract's own text, and the same input gives the same extraction every time.
In summary
Thomson Reuters Document Intelligence
Thomson Reuters Document Intelligence is contract analysis software, formerly ThoughtTrace, sold to law firms, in house legal teams and corporate contract and asset managers. Its extraction models, pretrained by Practical Law attorney editors, identify thousands of provisions across domain specific models, including more than 450 for wind and solar agreements. According to the AI Legal Index, users look across documents by meaning and context, set obligation and renewal reminders, build dashboards and review drafts in a Word add in with playbooks. HighQ customers receive extracted provisions in AI Hub and iSheets. Thomson Reuters' Legal Product Specific Terms let it train its products on customer data under a perpetual license, and it publishes no price, hosting region or security certification covering the product.
Zuva
Zuva is contract analysis AI from the team that built Kira Systems, formed when Kira was sold to Litera in 2021. It sells more than 1,400 pretrained extraction fields three ways, as a developer API, as a self serve review tool called Zuva Analyze, and as Diligence Reviews delivered with former large firm M&A lawyers. According to the AI Legal Index, Zuva prices Analyze at $10 a document with a $5,000 minimum and most Diligence Reviews deals at $10,000 to $75,000. The customer sets retention, and keeping nothing is an option. The service runs on Microsoft Azure in the United States or Europe, in a private cloud or self hosted. Zuva states SOC 2 Type II certification and names Microsoft's legal team as an API customer.
Questions buyers ask
Thomson Reuters Document Intelligence vs Zuva: which is better for M&A due diligence?
Both extract provisions from large sets of contracts for due diligence with their own pretrained models. Document Intelligence suits a law firm or in house team that wants the review inside an application with retrieval, dashboards and a HighQ feed. Zuva suits a team that wants per document pricing and its own retention settings, or a seller that wants a fixed fee diligence service in which former large firm M&A lawyers produce disclosure schedules. From the AI Legal Index, based on each vendor's own published materials as of October 10, 2026. No vendor pays for placement.
Does Document Intelligence or Zuva train AI on customer contracts?
Section 5.4 of Thomson Reuters' Legal Product Specific Terms, last modified July 2026, says it may use Document Intelligence customer data to train, update, modify and adapt its products. The resulting Machine Learning Adaptions belong to Thomson Reuters. Zuva says on a public page that neither Zuva nor other customers may use a customer's fields and data without permission, and its license terms limit Zuva's improvement right to usage data about access to the service. From the AI Legal Index, based on each vendor's own published materials as of October 10, 2026. No vendor pays for placement.
Which one publishes its price?
Zuva does. Analyze costs $10 per document with a $5,000 minimum across both pay as you go plans, and the API has a free tier and then charges by the document and by the field. Most Diligence Reviews deals cost $10,000 to $75,000. Enterprise volume and private hosting are quoted. Thomson Reuters publishes no price for Document Intelligence, and its terms set Editor and Viewer users in the order without a figure. From the AI Legal Index, based on each vendor's own published materials as of October 10, 2026. No vendor pays for placement.
Where do Document Intelligence and Zuva host customer data?
Zuva's security overview puts its API servers on Microsoft Azure in the United States and Europe, and its license terms place hosting in the jurisdiction the customer selects. Private cloud and self hosting on Azure or AWS cost extra. Zuva does not say where the generative answers in Analyze run. Document Intelligence runs as a service at app.thoughttrace.com, and Thomson Reuters names no hosting provider, region or tenancy model for it. From the AI Legal Index, based on each vendor's own published materials as of October 10, 2026. No vendor pays for placement.
What do Document Intelligence and Zuva both leave unpublished?
Neither publishes a precision, recall or error rate for its extraction fields, or a test set behind its accuracy claims. Neither names a person accountable for model behavior or publishes release testing results. Neither publishes a subprocessor list for the product, commits to tell customers when a model changes, or addresses privilege and work product. Neither describes a record of which model produced an extraction. From the AI Legal Index, based on each vendor's own published materials as of October 10, 2026. No vendor pays for placement.
Neither vendor paid for inclusion, placement or a grade, and neither reviewed this page before it published. Everything on it comes from public material on the dates shown. How the index grades.
The two training positions sit in different documents. Section 5.4 of Thomson Reuters' Legal Product Specific Terms lets it use Document Intelligence customer data to train and adapt its products, and that section ranks above the general no training clauses, which cover only generative models. Zuva's commitment not to use a customer's data without permission is a policy, and its license terms limit its improvement right to usage data without naming training. Under the Kira sale, Zuva does not sell its interface products directly to law firms until later in 2026. Neither vendor reviewed this page.