Paxton AI vs Vincent AI: how they compare in 2026
These two are shortlisted against each other constantly and they answer different buyers. Vincent AI runs on vLex's corpus of more than a billion documents across 100 plus countries, with United States primary law from Fastcase, and it quantifies that coverage by jurisdiction and court. Paxton AI is the self serve option. It publishes a real price, $499 per user per month, and Vincent publishes none. The deepest difference is confidentiality. Paxton's own footer states that communications with it are not protected by attorney client privilege or as work product, which is honest and is also the thing a firm has to reckon with. Vincent documents matter level segregation, inheriting a firm's own ethical walls through Clio.
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.
The models are the product. Research, drafting, document analysis, medical chronologies and the citator are all generative or machine learning capabilities, and there is no underlying practice system that would function without them.
The AI is the product. Vincent is sold as an AI assistant, running more than 20 generative workflows across research, litigation and transactions, with a pipeline that combines generative and rules based components and vector search over the underlying data. The vLex database it retrieves from is the grounding source rather than a separate product that would stand without the models.
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.
The disclosure is substantive, with a real gap. The vendor built its own citator and publishes precision, recall and F1 tables for it. It publishes measured accuracy of 93.82 percent on average on the Stanford Legal Hallucination Benchmark, broken out by task including case existence, citation retrieval and court identification. It also publishes 94 percent on the Stanford Casehold dataset of 2,400 examples for the citator. Both are public academic datasets, so an outsider can in principle run them again. Two gaps remain. The published results date from 2024, and no republication against the current product was found. The AI Confidence Indicator is a confidence signal rather than an explicit path for declining to answer.
Independent measurement, a named verification mechanism and grounding a user can open are all present. Vincent AI took part in the February 2025 Vals Legal AI Report, entering six tasks, and scored between 53.6 and 72.7 percent. It beat the lawyer baseline on document question answering at 72.7 percent, transcript analysis at 64.8 percent and document summarization at 58.9 percent. The evaluator also praised the product for refusing to answer where it lacked sufficient data rather than giving a hallucinated answer, and noted that those refusals lowered its scores. Every research output cites primary sources with direct links and carries a Cert Citator check on whether the cited law remains valid. The vendor claims randomized controlled trials showing 3.67 times more reliability than leading language models, with no published method located. Two limits apply. The vendor describes the Vals results as meeting or surpassing human benchmarks in four of five tasks, while the evaluator reports six tasks entered and three above baseline. And vLex agreed to join the later Vals legal research study, then withdrew before results were published, saying the study was not designed for enterprise AI tools.
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.
The product is user directed rather than agentic, and the vendor states plainly across the site that it provides self help services at the user's specific direction and cannot give advice, which is an oversight posture of a kind. The AI Confidence Indicator is a real review aid. What is not published is a described control structure: what the system does on its own, what a lawyer must approve, and what happens after an output is wrong.
A real published commitment with genuine review surfaces, short of thresholds. The product is positioned as a first draft assistant, with the vendor's chief strategy officer stating publicly that lawyers still bear responsibility for validating research results even when using an AI assistant. Review surfaces are concrete: a user sees the sources an answer rests on and can exclude specific sources, every output carries citations with direct links and a Cert validity check, and the Clio Operate integration gives every interaction an audit trail with access scoped as narrowly as a single matter. The vendor frames this as giving lawyers a clear path back to the source before an answer reaches a client, negotiation, filing or courtroom. Not located: what any workflow does unaided, a threshold at which it stops or escalates, and what the vendor commits to when an output is wrong.
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.
An unlabeled customer logo strip and three attributed testimonials from named individuals at named organizations stand in for deployment evidence. No figures, no dates, no described method, and no case study were found. A 2025 Legalweek Leaders in Tech Law award and a 22 million dollar Series A are recognition and funding rather than production evidence. Individual logos in the strip were not corroborated.
Scale claims and testimonials stand in for deployment evidence. The vendor states Vincent is trusted by 8 of the 10 top law firms worldwide and publishes an attributed customer testimonial about the relationship, and independent benchmark participation is real. But a share of top firms is a market claim without names, and the Vals result is test set performance rather than a production outcome. Searched the vendor site, the Vincent product pages, the Clio press material and the vLex news section and located no named customer paired with figures, a date and an assessable method.
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.
Confidentiality is asserted in general terms, and the documents behind it point the other way. The home page describes a secure closed model prioritizing the privacy of legal inquiries. The sitewide footer states that communications with the vendor are not protected by attorney client privilege or as work product. The terms of service require a user to warrant that submissions do not constitute confidential information, in a section scoped to questions, comments, feedback and other information about the Services rather than to uploaded documents. The privacy policy is scoped to website and account information, lists no document or prompt data, and states that the vendor does not process sensitive information, on a product that takes in medical records. No training prohibition, retention period for uploaded documents or segregation by matter was found in any of them.
Substantial published commitments cover retention and segregation, and the vendor's own training position is missing. The customer selects the retention window for conversation logs and files, defaulting to one year, after which data is permanently deleted. Segregation is documented at matter level through the Clio Operate integration, which the vendor says inherits the firm's existing permissions, ethical walls and audit trail, with access scoped as narrowly as a single matter. Vincent itself is described as multi tenant with each client's data logically isolated, and role based access control governs internal access. Zero retention agreements are stated with the language model providers, and Clio says no sensitive data is kept by the model. The vendor site, the Vincent pages, the security knowledge base article and the Clio press material, searched, said nothing about whether vLex itself may train on customer content. Provider retention and vendor training are different questions, and the material answers only the first.
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.
Paxton publishes a real position on the advice line in a footer on every page. It states that Paxton is not a law firm and not a substitute for an attorney, that communications are not protected by attorney client privilege or as work product, and that it gives no advice, explanation, opinion or recommendation about legal rights, remedies, defenses, options, selection of forms or strategies. Jurisdiction is addressed in part: the terms state the service is not intended for use where that would be contrary to local law. Nothing published addresses a lawyer's competence and supervision duties. The self help framing sits oddly against marketing aimed at attorneys, so it is not clear who the disclaimer is written for.
A real position on the lawyer's continuing responsibility is published in substance rather than as a disclaimer. The vendor positions Vincent as a first draft assistant. Its chief strategy officer has said publicly that lawyers still bear responsibility for validating research results when using an AI assistant, and the Cert validity check is framed as a clear path back to the source before an answer reaches a client, negotiation, filing or courtroom. The audience is professional throughout, from solo to global law firms, corporate legal departments, courts, government agencies, bar associations and academia, with no consumer surface located. Competence and supervision duties are not addressed as such, and no jurisdiction limits are stated despite coverage across more than 100 countries where professional rules differ. The responsibility statement appears in trade coverage quoting an executive rather than in a standing published policy.
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.
No governance position for model behavior was found: no named internal owner, no testing regime before release, no responsible AI statement, and nothing on uneven output across matter types, parties or populations. The trust center lists security control families, which is a different subject. The white papers index was not reviewed item by item and may say more.
Principles and evaluation are published without a governance mechanism a buyer could audit. Real elements exist. The vendor took part in an independent benchmark and published the results, including unflattering ones. It describes a hybrid architecture combining generative and rules based components for reliability, claims a program of randomized controlled trials, and describes Vincent Studio as building firm processes into workflows with consistency and governance. The vendor site, the Vincent pages, the security knowledge base and the Clio press material, searched, showed no AI governance framework, no AI management certification such as ISO 42001, no named owner of model governance, no testing gate before release, and nothing on uneven output across matter types, parties or populations.
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.
A generic privacy policy covers the website without addressing what happens to documents and prompts after processing. The security page does publish real controls: encryption in transit and at rest, quarterly access reviews on a least access principle, vendor risk management and a security development lifecycle, and the open trust center lists control families. But the privacy policy is scoped to account and marketing data, its retention section speaks only to the life of an account, and the terms say only that certain transmitted data is maintained to manage performance of the service. No retention period for uploaded documents, no subprocessor list, and no incident or breach notification practice is published.
Substantive published policy covering most of the ground, at a level of specificity few records here reach. Published in the vendor's security knowledge base: customer selectable data retention for conversation logs and files defaulting to one year with permanent deletion after, role based access control restricting internal access to authorized personnel under documented circumstances, logical isolation per client in a multi tenant application, encryption described, and regular independent assessments with real time monitoring to identify vulnerabilities. A trust center is published for obtaining certification reports. Zero retention agreements bind the language model providers. Not located: a named subprocessor list, and an incident or breach notification practice.
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.
Liability is addressed only through a standard limitation clause that disclaims the exposure the product creates. The published terms disclaim all warranties, including any representation about the accuracy or completeness of content, and cap total liability at the amount paid in the six months before the claim. They require the user to indemnify the vendor and state that the vendor has no liability for loss or corruption of transmitted data. No indemnity running to the customer, warranty on output or insurance position was found. The allocation of loss is published, and a buyer can read it before signing.
No published indemnity, liability cap, carve out, warranty on output or insurance position is published, and no customer terms of service or master agreement is published as published on the property.
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.
No integration into document management, matter management, contract lifecycle or filing systems was found in vendor material, and the platform navigation lists none. A third party review reports a Microsoft Word add in; that is not vendor material. The product is positioned as a browser based tool for solo and small firm users.
Real integrations exist, are named, and include the legal specific ones that matter. Named: Microsoft Word, Microsoft Outlook, and iManage, the document management system a firm facing research product most needs, plus a deep integration with Clio Operate under which Vincent runs inside the firm's existing permissions, ethical walls and audit trail. That last is integration at the permission layer rather than the file layer, which is materially deeper than a connector. Vincent Studio adds a no code workflow builder for enterprise customers to embed firm processes. Not located: per integration documentation describing what each connector moves, in which direction, and what an administrator configures, and no integrations index page was reached.
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.
The service is hosted, and the tenancy model is not stated. The terms of service state that the service is hosted in the United States and that a user accessing it from elsewhere consents to transfer and processing there, which is a real residency statement. Not found: any tenancy model, regional option, single tenant or private deployment, or statement of where processing happens as distinct from where data is stored.
The tenancy model is stated and residency is not. The vendor publishes that Vincent is a multi tenant application with each client's data logically isolated and secured, which is a clear and honest statement of the deployment model and is more than several records here disclose. Searched the vendor site, the Vincent pages, the security knowledge base and the Clio press material and located no hosting provider, no named regions, no customer selectable residency, and no statement of where processing happens as distinct from where data is stored. For a product serving more than 100 countries and marketed on cross border capability, published residency options would be expected and none is published.
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.
Badges appear with no scope, no date and no report available. There is an open trust center that anyone can reach. It lists SOC 2, ISO 27001 and HIPAA as compliant and GDPR as in progress, and names security control families. It carries no attestation report, coverage period, audit scope or auditor name. The security page says the vendor rigorously adheres to these standards, which is adherence rather than attestation. The pricing page describes the platform as HIPAA certified, though HIPAA has no certification scheme. Its annual plan card cites ISO 20701, which is not an existing standard, while the monthly card on the same page cites ISO 27001. Separately, section 1 of the terms of service states that the services are not tailored to comply with industry specific regulations including HIPAA, and that a user whose interactions are subject to such laws may not use them. The product sells medical chronology and medical billing features alongside its marketing claim of HIPAA compliance.
Certification is named, dated and reachable. SOC 2 certification was obtained in January 2025 and ISO 27001 certification in June 2025, both stated with month and year rather than as undated badges. A vLex Trust Center is published as the route to the official certification reports for a firm's due diligence, alongside an account representative, which makes it a self serve request flow rather than a sales gate. The vendor also says its security systems undergo regular independent assessments and real time monitoring. The auditing firm and the audit coverage period are not published.
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.
No model provider is named, no architecture is described, no subprocessor list exists, and there is no commitment to notify customers when the supply chain changes. The vendor refers to operating within a secure closed model without identifying what sits underneath.
The vendor describes its architecture and the terms binding its providers without naming them. Published: a hybrid pipeline combining generative and rules based AI, vector search organizing more than one billion documents, and zero retention agreements with large language model providers. That tells a buyer the commercial terms that govern the model layer. What it does not tell them is who is in it. Searched the vendor site, the Vincent pages, the security knowledge base and the Clio press material and located no named model provider, no statement of where models run, no subprocessor list, and no commitment to notify customers when the supply chain changes.
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.
Real pricing is published for part of the range, and Enterprise pricing is withheld. The pricing page publishes a figure, a unit and a term: $499 per user per month, or $2,999 per user per year billed annually, with the plan contents itemized. Signup is self serve with a 7 day free trial, and the terms of service disclose the exact card authorization hold of $499. The Enterprise plan is custom and based on volume, with no figure, floor or seat minimum published. What implementation adds is not stated for either plan.
No pricing page is published, no rate is published, no unit of charge is stated and no tier structure appears on the surfaces reached. Every commercial path located terminates in a demo or contact request. No third party pricing figure is published either, beyond a general observation in review material that pricing may be competitive against the two largest incumbents, which is not a figure.
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.
Segment and practice coverage is described with substance. Named segments are solo attorneys, small and midsize firms and in house teams, with dedicated pages for personal injury, family law, employment law, criminal law and corporate law. The research coverage page names every federal and state database included by jurisdiction and court, which tells a buyer where research coverage stops. Not found: any treatment of government or court use, and any statement of which firm sizes or practice areas the product is not built for.
Coverage is quantified and specific. The data spans more than one billion legal documents across more than 100 countries, US primary law from Fastcase, and more than 850 million court records behind profiles of judges, opposing counsel and parties. A Compare Jurisdictions workflow and a 50 State Survey workflow exist because coverage across jurisdictions is the product's stated strength, and the evaluator independently called its support for international matters a significant strength for global firms. More than 20 prebuilt workflows span research, litigation, transactions and litigation intelligence. Segments run from solo and small firms, through the Fastcase heritage, to global firms, corporate legal departments, courts, government agencies, bar associations and academia. Limits are published too: the evaluator noted its study covered only a small part of US capability, and the vendor said document summary tasks do not draw on its legal database.
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?
Read the terms of service dated 27 Feb 2026, the privacy policy dated 20 Feb 2026 and the security page on 29 Aug 2026. No located term or policy addresses training on customer content either way. The contribution license clause grants the vendor rights to access, store, process and use information the user provides, following the privacy policy, and the privacy policy is scoped to website and account information and does not mention uploaded documents, prompts or model training. Home page language about a secure closed model is not a training commitment.
The model provider layer is addressed and the vendor's own position is not. Published: zero retention agreements with the language model providers, and a Clio statement that no sensitive data is retained by the model. Those are retention commitments binding third parties, and retention and training are different questions. Searched the vendor site, the Vincent product pages, the security knowledge base article and the Clio press material on 29 Aug 2026 and located no statement of whether vLex itself may use customer content to train or improve its own systems, either way.
Recorded as silent under the rule that a value is never inferred from the absence of a contradiction, and specifically not inferred from a provider side retention term.
Prompt and Output Retention
How long does the product keep what a lawyer typed, and can that be set to zero?
The terms of service acknowledge that data transmitted to the service is maintained for the purpose of managing performance, with no period stated and no customer control described. The privacy policy retention section speaks only to personal information for the life of an account and does not address documents or prompts. No zero retention option was located as of 29 Aug 2026.
The customer selects the retention period for conversation logs and files, the default is one year, and after the selected period the data is permanently deleted. That answers whether the window is disclosed, whether the customer controls it, and what happens at the end. No material says that zero retention is a selectable setting, which matters to a firm that wants nothing kept at all rather than kept briefly.
Ethical Walls and Matter Segregation
Does retrieval respect the firm’s ethical walls, or can the model read across them?
Searched the home page, platform pages, security page, the open trust center, the terms of service and the help center index on 29 Aug 2026. No vendor material addresses ethical walls or matter level segregation. No document management system integration was located either, so there is no source system permission model for retrieval to inherit. The security page documents least access controls for vendor personnel, which is a different question.
In the Clio Operate integration, the vendor says Vincent works within the permissions, ethical walls and audit trail already set up in that system. Access can be scoped as narrowly as a single matter or as broadly as the firm, and every interaction is visible. Retrieval therefore runs inside the firm's existing access model rather than a parallel one the firm must keep aligned, and ethical walls are named as such. Two limits apply.
The inheritance is documented for the Clio Operate integration specifically. No equivalent statement was located for Vincent used on its own or through the iManage connector, where the product's own multi tenant logical isolation and role based access control apply. Enforcement for each user at query time is implied by scoping to a single matter but is not stated in those words.
Third Party Request and Subpoena Notice
If someone subpoenas the vendor for a firm’s data, does the firm hear about it first?
Read the terms of service and the privacy policy in full on 29 Aug 2026. Neither contains a government or law enforcement request clause. The privacy policy section on when and with whom information is shared names only business transfers such as a merger or acquisition. No commitment to notify a customer of a third party request was located, and no transparency report exists.
Searched the vendor site, the Vincent product pages, the security knowledge base article and the Clio press material on 29 Aug 2026, and no published customer agreement or data processing agreement was reached. No clause committing to notify a customer of a government or law enforcement request for their data was located, and no transparency report was located. Noted for a reader: the published retention control means a firm can limit how much history exists to be requested, which reduces the exposure without addressing the notification question.
Primary Law Corpus Provenance
Where does the law in this product come from, and does the vendor have the right to use it?
The vendor publishes a research coverage page that names its source databases individually rather than by jurisdiction alone. It lists the United States Code, the Constitution, the Code of Federal Regulations, every federal circuit and the Supreme Court, and six named administrative tribunals. For each state and territory it lists the statutes, administrative code, court rules and named courts. No license or public domain basis is stated for any of it, and no update cadence or lag is published.
The sources are named and the rights basis is stated. The data is vLex's own global legal database combined with US primary law from Fastcase, both now owned by the same company, and the vendor says directly that the legal authority behind Vincent's research is owned rather than licensed from another party. Scale is quantified at more than one billion documents across more than 100 countries, plus more than 850 million court records.
Owning the data removes any dependency on a license. Update frequency is the gap: the vendor site, the Vincent pages and the Clio press material, searched on 29 Aug 2026, gave no update frequency or lag for any jurisdiction and no statement of completeness by jurisdiction.
Good Law Verification
Does the product tell you when the authority it just cited has been overruled?
The vendor built and surfaces its own treatment signal, the Paxton AI Citator, described as patent pending, which reports whether a case has been overruled, affirmed, questioned or cited by later cases and returns explanations in Bluebook format. The method is described and quantified: evaluated against the Stanford Casehold dataset of 2,400 examples with published precision, recall and F1 figures by case type. It is the vendor's own signal rather than a licensed commercial citator such as Shepard's or KeyCite. The published evaluation dates from 2024 and no republication against the current product was located.
Vincent answers the good law question directly. Every research output carries a Cert Citator check on whether the cited law remains valid, shown with the authority rather than as a separate step. Cert has a traceable origin: it comes from Judicata software that Fastcase acquired before the vLex merger, and the same company now owns it. Because the vendor owns the citator rather than licensing it, no outside party can withdraw it.
Coverage outside the United States is published and uneven. Clio's help article on verifying case law, updated 16 September 2026, shows positive and negative treatment counts and a negative treatment flag for cases in Canada, the UK and Ireland. In most other jurisdictions a case shows only how often it has been cited, which says nothing about whether it is still good law. The Canadian launch of 23 September 2026 adds checking of existing citations against Canadian authorities.
A buyer researching elsewhere should expect a citation count rather than a treatment signal. Verified 27 September 2026.
Refusal and Uncertainty Behavior
What does the product do when the answer is not in the corpus?
The product exposes an AI Confidence Indicator, announced alongside the vendor's benchmark results and described as helping a lawyer judge the reliability of a response. Searched the site, platform pages, benchmark posts and help center index on 29 Aug 2026 and located no explicit no answer or abstention path, and no published statement of what the product does when it cannot ground an answer.
Abstention was observed independently, and it cost the vendor points. The February 2025 Vals evaluator praised Vincent AI for refusing to answer questions where it lacked sufficient data rather than giving a hallucinated answer, and recorded that those refusals lowered its scores. A product that accepts a lower benchmark result rather than guess is showing the behavior in question, and the evidence comes from an outside evaluator with nothing to gain from the finding.
The vendor itself does not document an explicit path for declining to answer. The vendor site, the Vincent product pages and the vLex news section, searched on 29 Aug 2026, showed the behavior as observed by others rather than described by the vendor, so it is evidenced but not promised.
Fabricated Citation Record
Does a public court record exist addressing fabricated or hallucinated legal citations in output from this product?
One court record mentions this product, and the court's findings address the lawyer's failure to verify rather than the product's output. In Flycatcher Corp. Ltd. v. Affable Avenue LLC (S.D.N.Y., No. 24 Civ. 9429, opinion and order of 5 February 2026), counsel told the court that after running arguments through vLex he put his motion brief through Paxton AI as a cite checker and then through a further AI check. He said the cite checker introduced citation errors that he took for corrections; the court found that explanation not credible, found that he never cite checked the brief himself, struck the filings and entered default judgment against his client.
The court made no finding about Paxton AI's own output. The AI Hallucination Cases database maintained by Damien Charlotin was searched on 2 October 2026 and returns this matter as the only record naming the product.
Three court records mention vLex as a tool counsel used, and each addresses the lawyer's failure to verify rather than the product's output. In Flycatcher Corp. Ltd. v. Affable Avenue LLC (S.D.N.Y., No. 24 Civ. 9429, opinion and order of 5 February 2026), counsel said he ran some arguments through vLex before using a separate cite checker; the court found he never cite checked the brief himself and struck the filings.
In Brooks and Brooks Inc. v. Patel (Minnesota District Court, Hennepin County, No. 27-CV-24-11194, order of 8 September 2025), counsel admitted failing to confirm a citation through VLex Fastcase and was fined $5,000. In Fletcher v. Experian Information Solutions (5th Cir., No. 25-20086, 18 February 2026), the court named Casetext and vLex among counsel's sources and sanctioned counsel for not verifying. All three name vLex rather than Vincent AI, and none makes a finding about the product's output.
Auto 61/2026 of the High Court of Justice of Galicia (7 July 2026) lists Vincent AI among verified legal AI tools when fining a lawyer, without naming the tool used. The AI Hallucination Cases database maintained by Damien Charlotin was searched on 2 October 2026.
Bar Guidance Alignment
Has the vendor engaged in public with the ethics opinions its buyers are bound by?
Searched the home page, platform and practice area pages, the news and insights index, the white papers index and the webinars page on 29 Aug 2026. No engagement with any named ethics opinion was located, including ABA Formal Opinion 512 and state bar guidance. The sitewide footer carries a specific unauthorized practice disclaimer, which is a statement about the vendor's own role rather than engagement with the guidance its buyers are bound by.
Searched the vendor site, the Vincent product pages, the vLex news section and the Clio press material on 29 Aug 2026. No engagement with any named ethics opinion or professional guidance was located, including ABA Formal Opinion 512 and state bar guidance. Worth recording as context rather than credit: the company has a substantial bar association presence inherited through Fastcase, which supplied research tools to state bar memberships, so a relationship with bar associations exists without engagement with their AI guidance being published.
Billing and Fee Posture
Does the vendor address what happens to the bill when the work takes an hour instead of six?
Vendor material is framed around time saved, describing streamlined document analysis to save time and accelerate decisions, with testimonials quoting time savings. Searched the site, platform pages, help center index and legal pages on 29 Aug 2026 and located no per matter record of AI assisted work intended for fee purposes, and no published guidance on billing, fee or client disclosure treatment.
Searched the vendor site, the Vincent product pages, the vLex news section and the Clio press material on 29 Aug 2026. No per matter record of AI assisted work intended for fee purposes was located, and no guidance on billing, fee or client disclosure treatment was located. Time savings are referenced in general terms, such as workflows turning hours of research into minutes, without a quantified savings claim tied to a matter.
Recorded as not addressed rather than at the savings claims value because the located framing is capability description rather than a savings figure. Noted for a reader: the Clio Operate integration records every interaction in an audit trail, which is the raw material for such a record without being published as one.
Outside Counsel Guideline Readiness
Can a firm get this vendor through a client’s AI clause without a bespoke negotiation?
The open trust center publishes compliance badges and security control families with no gate, which is a real access route. What a firm would need to satisfy a client AI clause is absent: searched the trust center, the security page, the terms of service and the privacy policy on 29 Aug 2026 and located no subprocessor list, no statement of which model providers see customer content, and no client facing consent or notification material. There is also no data processing agreement published.
A firm can assemble much of what a client AI clause asks for without a bespoke negotiation, though not all of it. Available without an agreement: a published trust center as the route to the official SOC 2 and ISO 27001 reports, dated certifications, a published retention policy with customer control and permanent deletion, a stated position that language model providers are bound by zero retention agreements, and a statement that the legal authority behind the product is owned rather than licensed.
The certification reports sit behind a trust center request or an account representative rather than being published. No subprocessor list, no statement of which model providers see client content, and no consent or notification pack for clients were located.
Court Disclosure Support
If a judge’s standing order requires an AI disclosure, can the product produce one?
Some elements of a disclosure record are available. The product returns citations with the relevant sources highlighted, and the citator produces a treatment result with a written explanation, so sources retrieved and their status can be shown per answer. Two elements are missing: the model used is not disclosed anywhere on the property, and no per document export covering model, sources and human verification together was located as of 29 Aug 2026.
Much of a disclosure record exists, assembled from several published features rather than offered as one. Through the Clio Operate integration every interaction is visible in an audit trail scoped to the matter. Every research output cites primary sources with direct links plus a Cert validity check, so what was relied on and whether it was still good law are both recorded in the work product. A user can see and exclude the sources an answer rests on, which leaves a trace of human intervention.
Two parts are missing. No single export for each document covering the model used, the sources retrieved and human verification was located, and no model is named in published material, so the model used could not be stated.
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.
- Third Party Request and Subpoena Notice
- Bar Guidance Alignment
Which one fits
Choose Paxton AI if
- Price transparency is the point. Paxton publishes $499 per user per month or $2,999 per user per year with plan contents itemized and a seven day free trial, and sells through self signup rather than a sales process. Vincent publishes no rate at all.
- You are solo or small and want to start today. Paxton's segments are stated as solo attorneys, small and midsize firms and in house teams, with dedicated pages for personal injury, family law, employment, criminal and corporate practice.
- You want a published position on the advice line. Paxton carries one on every page: not a law firm, not a substitute for an attorney, and no advice, opinion or recommendation on rights, remedies, defenses or strategy.
Choose Vincent AI if
- Your research crosses borders. Vincent discloses its coverage in detail: more than a billion documents across 100 plus countries, United States primary law from Fastcase, and dedicated Compare Jurisdictions and 50 State Survey workflows.
- Client confidentiality has to survive a partner's review. Vincent documents matter level segregation through its Clio Operate integration, running inside the firm's existing permissions, ethical walls and audit trail, and publishes a customer controlled retention window.
- You need document management and productivity connections. Vincent names Microsoft Word, Outlook and iManage. Paxton lists no document management, matter management or filing system connection.
In summary
Paxton AI
Paxton AI is a self serve legal AI assistant aimed at solo attorneys, small and midsize firms and in house teams. It covers research, drafting, document analysis, medical chronologies and medical billing summaries. The AI Legal Index records a published rate: $499 per user per month or $2,999 per user per year, sold through self signup with a seven day free trial. Paxton built its own citator and publishes precision, recall and F1 tables for it, alongside measured accuracy of 93.82 percent on the Stanford Legal Hallucination Benchmark. Its sitewide footer states that communications with the vendor are not protected by attorney client privilege or as work product.
Vincent AI
Vincent AI is a legal research and workflow assistant built on vLex's global database of more than a billion documents across 100 plus countries. It combines vLex international content with United States primary law from Fastcase and ships more than 20 prebuilt workflows. According to the AI Legal Index, its coverage is quantified by jurisdiction and court rather than claimed. It participated in the independent 2025 Vals Legal AI Report, opting into six tasks and beating the lawyer baseline on document question answering at 72.7 percent. Its Clio Operate integration runs inside a firm's existing permissions, ethical walls and audit trail, which is segregation at the permission layer.
Questions buyers ask
Paxton AI vs Vincent AI: which should I choose?
Vincent AI is the stronger choice on coverage, confidentiality and integrations. It quantifies its global coverage, documents matter level segregation through Clio Operate, and names Word, Outlook and iManage. Paxton's case is access: it publishes a real price of $499 per user per month, sells through self signup with a free trial, and is aimed at solo attorneys and small firms. If budget transparency and starting today matter more than global coverage, Paxton is the pragmatic pick. From the AI Legal Index, based on each vendor's own published materials as of August 30, 2026. No vendor pays for placement.
What does Paxton AI cost?
$499 per user per month, or $2,999 per user per year billed annually, with the plan contents itemized, signup self serve and a seven day free trial. The enterprise tier is withheld. Vincent AI publishes no rate at all. From the AI Legal Index, based on each vendor's own published materials as of August 30, 2026. No vendor pays for placement.
Is Paxton AI protected by attorney client privilege?
Paxton's own sitewide footer states that communications with the vendor are not protected by attorney client privilege or as work product. That is the vendor's published position, stated candidly, even though its home page describes a secure closed model prioritizing the privacy of legal inquiries. Vincent AI by contrast documents matter level segregation inheriting a firm's own ethical walls through its Clio Operate integration. From the AI Legal Index, based on each vendor's own published materials as of August 30, 2026. No vendor pays for placement.
Which one has been independently tested?
Both have measurement behind them, in different forms. Vincent AI participated in the February 2025 Vals Legal AI Report, opting into six tasks, second only to Harvey by task count, and beat the lawyer baseline on document question answering at 72.7 percent. Paxton publishes measured accuracy of 93.82 percent on the Stanford Legal Hallucination Benchmark broken out by task. It also built its own citator and publishes precision, recall and F1 tables for it. From the AI Legal Index, based on each vendor's own published materials as of August 30, 2026. No vendor pays for placement.
Which covers more jurisdictions?
Vincent AI, substantially. It runs on more than a billion legal documents across more than 100 countries with United States primary law from Fastcase and more than 850 million court records, and ships dedicated Compare Jurisdictions and 50 State Survey workflows. Paxton publishes a research coverage page naming every federal and state database included by jurisdiction and court, which is a concrete boundary statement limited to the United States. From the AI Legal Index, based on each vendor's own published materials as of August 30, 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 most important sentence on this page is one Paxton AI publishes about itself. Its sitewide footer states that communications with the vendor are not protected by attorney client privilege or as work product. That is candid, and it is also a live problem for a firm putting client facts into the tool. Read it as disclosure working correctly, not as a gotcha. These two also sit in different index categories, Paxton in general legal assistants and Vincent in legal research. They are paired because buyers compare them, not because they are the same kind of product. Neither vendor reviewed this page.