Drafty AI vs Visalaw AI: how they compare in 2026

The product is a chain of model driven steps that turns a client's evidence into a filing. Smart Upload reads uploaded documents and fills in case details, and the drafting interview asks only about facts the matter does not already hold. The Strategy Outline engine maps the evidence to each legal element of the claim, builds an evidence matrix showing which elements have little support, detects the forum and flags deadlines, bars and weak points, each with a citation. From there one Draft action routes the matter by case type. Generic matters go to the Drafting Studio interview, while RFE, NOID and NOIR responses, asylum, withholding and CAT briefs, I-601A waivers, expert letters, affidavits, exhibit lists and translations each run in a dedicated flow. The Document Improvement Studio rewrites an existing document paragraph by paragraph. Matters and clients are the only parts that work without the models, and the product is sold for what the models draft.

The models drive the research and drafting the vendor leads with, on a platform with real value beyond AI. Research in natural language over the immigration library, drafting of petitions and briefs, legal analysis, chat over uploaded documents, translation and summarization are model work, and the pricing is built around them. Around them sit projects organized by matter, team permissions and audit trails, electronic signature, a certified translation option and a large curated library, which the company's connector for Claude exposes read only without Visalaw's own models. Remove the models and a curated immigration research library remains.

Drafts are built with live legal research tailored to the jurisdiction, and attorneys quoted on the site point to its circuit specific case law. Every section in review shows the generated text beside the citations and supporting sources tied to it, and the authorities a draft relied on are gathered into a table of authorities that carries into the exported Word file. After generation, citations the system failed to match to its sources are flagged in that table for a closer read. The matching tolerates a different pinpoint or format, so a flag means a citation missing from the sources or a match that failed to complete. Drafty presents this as an aid rather than a guarantee and tells attorneys to verify every authority and fact whether flagged or not. The research sources behind the drafts are not named, and no accuracy rate, test set or measured error figure is published.

Answers are grounded in named sources, with citations on every output, and nothing is measured. The research library covers agency materials from USCIS, DOS and DOL, among others, board decisions such as BIA and AAO, AILA and AILALink secondary content, and more than ten million cases from over 2,000 federal, state, territorial and tribal courts, described as more than 99 percent of published precedential US case law. Every output includes citations for cross checking, the user chooses which libraries are in scope, and each answer names the ones it used. The README for the connector for Claude, published 17 September 2026, adds that results are ranked by authority, that cited authorities are checked against the library catalog, and that case law comes through CourtListener. No citation accuracy rate or evaluation is published, and the claim that the product ensures accuracy every time rests on no measurement.

Every matter moves through the same visible stages, Outline, Draft, Review and Drafts ready, and nothing becomes a Word document until an attorney finalizes it. Drafts come back in sections that each start as Pending. The attorney reads the text with its sources, edits it, rewrites it with AI help or marks it Reviewed, in any order, and the matter header counts the drafts still awaiting review. Unmatched citations stop in the table of authorities for a closer look rather than passing silently. Notice responses run a separate Verification step before export, with automated checks for issues that lack linked evidence and for unlinked exhibits. Exhibit lists are approved by the attorney before export, and the Document Improvement Studio applies no change the attorney has not accepted. Where an interview question is skipped, the draft carries a placeholder rather than an invented fact. The terms make the attorney responsible for reviewing and approving each document as their own before it goes to any agency or court.

The attorney stays in control by written position, with real review surfaces and a supervision rule for staff, and no published control structure. The Platform Agreement makes the customer responsible for evaluating output for accuracy, and the about page says technology should support professional judgment rather than replace it. Staff use is tied to a supervising attorney with audit trails, and the firmwide offer of support staff seats extends that model rather than loosening it. Drafts are generated for editing in the platform, and a customer story from Kuck Baxter describes the product supporting early drafting, with senior attorneys handling judgment, revision and client strategy. No categorical limit on what an output may be used for, and no check before a drafted petition leaves the platform, is published.

Immigration attorneys Patrick Kolasinski, Joe Michelotti and Vilerka Bilbao and legal consultant Luis Bares are quoted by name, praising brief drafting, circuit specific law and the RFE drafting studio, without their firms or any figures. An unnamed user reports saving more than 60% of the time spent drafting cover letters. The headline claim is a 60% or greater cut in case preparation time, described as verified by customers, with no customer, sample or method behind it. A worked example on the home page assumes nine hours a week reclaimed at a $300 hourly rate. Coverage in ABA Journal, Bloomberg Law and AILA and membership of NVIDIA Inception are listed as press and programs. No dated deployment at a named firm with measured results is published.

Named customers are published with some figures, and no method. The customer stories page profiles Charles Kuck of Kuck Baxter Immigration in Atlanta, a past president of the American Immigration Lawyers Association whose large firm uses the platform for early drafting, Cyrus Mehta of Cyrus D. Mehta & Partners, and Ken Robinson, a solo practitioner who says it saves him four to six hours a week and let him lower his hourly rate. Five stories are published in all. The pricing page claims savings of more than ten hours per case and $60,000 a year. None of the figures is measured across a described sample of cases or firms, and the stories are not dated.

The data processing agreement, part of the terms every user accepts, limits use of client data to providing the service and bars Drafty and its subprocessors from using it to train, fine tune or improve any AI model. Client and attorney details are masked before a prompt reaches an AI provider, so the provider receives personal details only where a draft needs them. Providers hold inputs for up to 30 days for abuse monitoring and then delete them, and everyone who handles client data is bound to confidentiality. A published guide covers United States v. Heppner, the 2026 ruling on AI and work product, but the agreement itself does not address privilege or work product. Firms add users through team management, and nothing describes walls between matters or users inside a firm. Session replay runs inside the platform and can capture case content shown on screen. Input fields are masked, and staff view a recording only with the customer's authorization to resolve a support request.

Every element of the confidentiality picture is in the contract, and the vendor addresses privilege and work product directly. The Platform Agreement bars training by Visalaw and its model providers and treats customer content as the customer's confidential information for the term and ten years after. Isolation by organization reaches the shared vector index, and model providers may not retain or share content after a call. The Data Processing Addendum limits removal of identifiers to operational metadata. The connector for Claude does not log search queries and trains nothing. The company has also written publicly on how privilege applies to AI output.

Only attorneys licensed in the United States may use the product, and each must give a bar license number at sign up. The terms describe Drafty as an AI system designed to assist attorneys and disclaim the accuracy of AI generated content. Review, validation and compliance with professional conduct rules sit with the user, who must approve each document as their own before filing it with an agency or court. Scope is stated narrowly as immigration practice in the United States and nothing else. Blog guidance walks attorneys through Florida Bar Ethics Advisory Opinion 24-1, ABA guidance on AI and the Heppner ruling, and offers an AI use policy template and retainer language for firms. How the product supports an attorney's competence and supervision duties, as opposed to placing them on the attorney, is not set out in the terms.

A clear position on what the product is, who may use it and how staff use is supervised is published, without jurisdiction limits and with one claim a tool cannot make. The Platform Agreement calls the service a research and drafting tool whose output is not legal advice. The platform FAQ says the product is designed for immigration attorneys and professionals who are AILA members, and that paralegals and assistants may use it only under a supervising attorney's account with audit trails, which maps onto a lawyer's duty to supervise assistants who are not lawyers. No jurisdiction limit is stated, although the library now includes Canadian and global collections. The same FAQ asserts that the product ensures ethical compliance and meets professional responsibility standards, which only the lawyer using it can do.

No responsible AI policy, governance framework, accountable owner or testing program is published. A 2023 post describes a commitment to make AI practical and safe for immigration attorneys and to evaluate AI at every stage, without saying how. Drafty's ABA guidance summary lists fairness and avoiding bias as principles for lawyers using AI, and nothing describes how its own drafts are tested for uneven quality across case types, forums or client populations.

Values and data controls specific to AI are published, and a governance standard is in progress, without published governance of what the models produce. The about page states principles, including that every answer must be grounded in the law. The security page lists ISO/IEC 42001, the AI management system standard, as in progress and says its commitments align with it, and the Security Addendum has a section on AI model security. Under the Data Processing Addendum, the terms will be renegotiated if new AI regulation requires it. Nothing names who owns model quality, what is tested before a model or prompt change reaches firms, or what has been found about uneven output across case types, languages or applicants' nationalities, for a product drafting asylum declarations and hardship waivers.

Retention and deletion are contractual. Case data stays while the account is open, and on closure or earlier request Drafty deletes or returns it at the customer's choice within a reasonable period, with backups removed on the regular cycle. AI providers keep inputs for up to 30 days, and analytics recordings follow the analytics provider's settings. A security schedule in the agreement covers encryption in transit and at rest, role based and least privilege access for staff, and multifactor authentication on administrative and production systems. It adds staff training, vulnerability scanning, backups with a documented recovery process and logging of production access. Breaches are reported without undue delay. Subprocessors fall into six named categories, all in the United States and all barred from training on customer data, with prior notice of any change and a right to object. The named list is sent on request by email, and breach notice carries no fixed deadline.

The stewardship set is published in contract detail, with only the subprocessor list held back behind a request. The Security Addendum commits to AES 256 encryption at rest and TLS 1.2 or higher in transit, least privilege access with multifactor authentication, annual penetration tests, stated windows for fixing vulnerabilities, and staff background checks. It promises notice of a security incident without undue delay and in any event within 24 hours of detection, with an independent forensic assessor at Visalaw's expense. Customer data is returned or deleted after termination, and new subprocessors come with advance notice and objection rights. The connector for Claude logs only a hash and encrypted copy of the access token and the name of each tool called, not the search terms. The subprocessor list sits in a Trust Portal that requires an access request.

The terms, last updated 5 October 2026 under Delaware law, exclude indirect, incidental, special, consequential and punitive damages, including lost data and profits, and say nothing about a cap on direct damages. Drafty makes no warranty on the accuracy or completeness of AI generated content and accepts no liability for loss caused by its third party AI providers or for submission of documents the user did not revise. The service is provided as is, with merchantability, fitness and non infringement warranties disclaimed, and disputes go to mediation and then binding AAA commercial arbitration. Users indemnify Drafty for infringing or defamatory material, misrepresentation, breach and claims by third parties arising from their use. Drafty gives no indemnity in return, and no insurance or output warranty is offered. The data processing agreement's security commitments survive the disclaimers, and its liability falls under the same terms.

A detailed liability position is published with named insurance, and nothing stands behind output. The Platform Agreement of 17 March 2026 gives a vendor indemnity against third party claims that the service infringes intellectual property or arises from Visalaw's violation of law. It caps general liability at twelve months' fees, with carve outs for confidentiality breaches and the indemnities, and sets a separate data breach cap at twice annual fees. It also commits Visalaw to carry technology errors and omissions and cyber liability insurance, with certificates on request, though the amounts of cover are not stated. Output accuracy is placed on the customer: the agreement says output is generated by AI and may contain errors, and the firm is responsible for evaluating it.

Drafty stands alone. Client documents are uploaded or linked through each matter's documents panel, and finished drafts download as Word files named for the client and matter type, ready to edit and file. No connection is offered to a case management system such as Docketwise, Clio or MyCase, to a document store, or to USCIS or EOIR filing. Drafty's own comparison with Docketwise presents the two as separate products a firm runs side by side. Enterprise plans add single sign on.

Named integrations span immigration practice systems, and one is documented in depth. The platform page lists partners including the eImmigration case management system, Clio and Docketwise, along with tools for email drafting, translation and file storage. It says research and drafts fit into case management and email workflows, and API access is available on request and included on Pro. The connector for Claude has a public README, dated 17 September 2026, giving the server address, the transport, OAuth 2.1 with PKCE and the single read only scope granted. It documents each of its seven tools, from semantic library search and lookup of an INA section or CFR provision to case law search and saved prompts. Every tool is limited to what the signed in user's organization can already see. What each practice system integration exchanges, and in which direction, is not documented to the same standard.

The service runs in the cloud on Amazon Web Services. Customer data is processed and stored in the United States under the data processing agreement, and every subprocessor, the AI providers included, is based there too. Transfers of data covered by the GDPR or UK GDPR rely on standard contractual clauses incorporated into the agreement. One deployment in the United States serves every plan. No other region, private deployment or residency choice is described, and the tenancy model is not stated. The security portal lists a recovery time objective of five hours.

Tenancy, hosting region, residency options and where AI processing happens are all stated in the contract. The Security Addendum says customer data is stored and processed by Visalaw and its vendors in US data centers, hosted by default in the AWS US East 1 and MongoDB US East 1 regions with replication only to other US regions, and that customers may request a specific region. The service runs on shared infrastructure with logical separation by organization at the application and database layers. Model providers receive content over encrypted API calls and may not keep it beyond the call. Encryption keys are managed by the providers, with no customer managed key option. The Data Processing Addendum calls the arrangement a dedicated tenant environment, so the two addenda describe tenancy differently.

No independent security attestation is held today. The security portal, run on SafeBase, describes the company as aligning with and in pursuit of the CCPA, the GDPR and SOC 2, and lists SOC 2 under compliance without a report, auditor or date. It sets out controls for data security, access monitoring, backups, erasure, bot detection, credential management and the development lifecycle, shows a SecurityScorecard rating, and lets a reviewer request access to restricted documents. Customers hold audit rights under the processing terms, which Drafty may meet with certifications, reports or a security questionnaire.

A named attestation is published with a contractual route to the report, and work in progress is stated as such. Section 1 of the Security Addendum says the security program is audited against SOC 2 Type II, lists ISO 27001 with a target of the fourth quarter of 2026, and commits to make outside audit reports available on written request. Section 10 adds penetration test summaries and questionnaire responses at no cost, and says Visalaw will adopt an equivalent standard if it ever drops an audit. The security page lists a dedicated chief information security officer, SAML single sign on, audit logs and IP allow listing, and points to a Trust Portal hosted on Drata that opens after an access request. The auditor, the SOC 2 report period and its scope, including whether the models fall within it, are not published.

Drafts come from third party AI providers called through their APIs, under contracts that bar training on inputs, with client and attorney details masked before a prompt leaves Drafty. The providers and models are not named, and nothing describes which model handles which flow. AI processing is one of six subprocessor categories in the agreement, all based in the United States. Prior notice of any added or replaced subprocessor is promised, with a right to object, which covers a change of AI provider. Use of the providers' models is also subject to their own usage policies, which the terms pass through to users.

The AI architecture and the obligations on model providers are set out in the contract with change notice, and the providers are named only behind an access request. Section 6 of the Security Addendum describes content passing to contracted model providers over encrypted APIs and bars them from training, from keeping content beyond a single call and from sharing it onward, and Visalaw commits to document which subprocessors receive data at each stage. The Platform Agreement and Data Processing Addendum give 60 days' notice before a new or changed subprocessor, with a right to object. The connector for Claude's privacy policy names one third party directly, CourtListener, which receives case law queries without account information. The model providers are listed in the Trust Portal, which requires a request; a 2024 announcement issued with AILA named GPT-4.

Prices are published in USD and bought without a sales call. The base plan is $250 a month for one seat and ten matters a month. Each further seat costs $100 a month, and matters cost $25 each for the first ten and $10 each above that, shared across the firm. A matter is one client case with unlimited drafts inside it, charged once at the first AI generation. Overage at $20 a matter is off by default. Yearly billing gives two months free and releases the year's matter pool up front, and seat or matter changes are prorated. A four day Starter Trial with two matters needs no card, followed by a seven day trial once a card is added. The calculator stops at 25 seats and 500 matters. Enterprise pricing, with volume matter rates, SSO, an uptime SLA and onboarding, is set on a call. Self serve plans carry no stated setup fee.

The product can be priced, and each plan's contents seen, before any sales conversation. Core is $220 a month per license, billed annually, for research, chat, document analysis, translation and a case management integration, with occasional drafting at $150 a draft. Pro is $480 a month per license and adds ten petition or brief drafts a month, about $26 a draft, and Enterprise is priced on request. A comparison table shows what each plan includes, and a return calculator lets a firm model savings from its own hourly rate and case mix, with its assumptions shown. For a solo immigration attorney, the annual cost and the break even point can both be worked out before buying. Only the Enterprise rate is withheld, and a two week free trial opens after a demo.

Immigration practice in the United States is the whole scope, for attorneys licensed there, from solo practitioners to firms large enough for an enterprise plan. Drafty puts that point at around ten users or 100 matters a month. Named case types include asylum, withholding of removal and CAT claims, I-601 and I-601A waivers, cancellation of removal, VAWA, U and T visas and BIA appeals. Motions to reopen and reconsider, RFE and NOID responses and family petitions are covered too, as are employment petitions such as EB-1A, NIW, EB-5, O-1A, P-1A, L-1 and TN. Limits are stated. Naturalization matters get a client filing checklist and cover page rather than a drafted brief, and court motions are offered only for immigration court matters. Forum detection frames work for USCIS, immigration court or appellate review. In house and government use are not addressed, which fits a product sold only to licensed immigration attorneys.

Segments, case types and limits are documented precisely. The product is for immigration attorneys who are AILA members, from solo practitioners to large firms such as Kuck Baxter. The platform page lists the case types supported for drafting: business petitions such as O-1A, EB-1A and H-1B, family matters such as I-485 and N-400, humanitarian work such as asylum, U visas and VAWA, and work across categories such as RFE responses and BIA appeals. The research library has separate US, Canadian and global collections. Anything outside immigration is outside the product by design, and the Claude connector's case law tool excludes state courts unless asked.