Neos vs SmartAdvocate: how they compare in 2026

The tier structure settles this axis more cleanly than any editorial judgement could: NeosAI is sold only in Platinum, the top of three plans, and the two plans beneath it are complete, purchasable case management systems without any AI at all. Essentials at $109 per user per month carries document management, intake forms and workflows, standard reporting and email and calendar sync; Premium adds OCR, PDF editing, Bates stamping and redaction. The vendor sells its own product without the models, at a published price, which is the plainest possible evidence that the platform stands on its own. What the models do add is a genuine capability rather than a convenience: Demands Plus assembles a medical chronology across every page of records with dates, providers, diagnoses, treatments and outcomes extracted and each entry linked to its source page, flags breaks in the treatment timeline, and drafts the demand from that chronology. Alongside it sit AI Document Generation, Summarization, Extraction, Dynamic Layouts and Chat. Graded level with Opus 2, Casepoint, Consilio and Nextpoint on the same reasoning: the engine of real capabilities, layered on a workflow system with a forty-year lineage through Needles and TrialWorks that plainly functions without it.

Fifteen named AI features drive capabilities a firm pays for, on a case management system that predates all of them and works without any of them. The engine claim is real for several: demand letter generation from medical chronologies and depositions, medical chronology generation from unstructured records, a data extraction engine that pulls court deadlines and coverage details into fields, and autonomous agents for recurring tasks. The vendor's own framing is built in, not bolted on, and natively integrated rather than a separate application or browser extension. Two things hold it at this band. The platform underneath is a mature system covering intake, matters, documents, calendaring, workflow, texting, e-signature, portals, reporting and an API, and none of that needs a model. And the commercial treatment is that of an add-on: the terms group Built-In AI Tools with texting and e-signature as things purchased separately, and reserve a right to monitor AI usage for fair use and to impose an additional charge based on usage, limit it, or suspend it. A capability the vendor can meter and switch off is a layer, not the product.

The grounding apparatus is real, described, and verifiable by the reader, which is what separates this from the assertions common in the lane. Every chronology entry is linked to its source down to the document and page; every claim in a demand draft carries a citation that opens the exact source in Neos, described as a document, a note, or a value on the case. The retrieval method is described rather than implied, in two passes: the chronology is assembled from the records and establishes the facts, then the demand is drafted from the chronology rather than from raw records, drawing additionally on emails, call logs and notes already on the file. What is missing is any measurement. No accuracy figure, no test set, no failure modes named, and the Demands Plus FAQ asserts that a draft 'is accurate because every claim is cited to its source and an attorney reviews it', which is an argument from process rather than a number an outsider could test. The published figures are throughput, not accuracy: a first draft in about fifteen minutes, roughly thirty including attorney review, against four to eight hours by hand. Note also that the sources cited are the case record rather than primary legal authority; this product retrieves no law, so the citator and good-law limbs of the A band do not apply to it and the grade rests on the limbs that do.

Quality is asserted in the product descriptions and disclaimed in the agreement, with nothing measured in between. The marketing claims accurate document descriptions, clean searchable hyperlinked medical timelines, and tailored demand letter drafts. Hyperlinking in the chronology is the only grounding mechanism located, and it is described as a feature of the output rather than as a verification path with a stated behaviour. The agreement runs the other way and is more specific than the marketing: it states that the provider makes no warranties regarding the accuracy, currency, completeness or usefulness of AI content, that any reliance is solely at the subscriber's own risk, and that the subscriber is fully responsible for reviewing and verifying accuracy, legality and appropriateness before use. Searched the AI tools page and the complete terms of service on 31 Aug 2026 and located no accuracy figure, no test set, no evaluation, no error rate and no hallucination disclosure. One adjacent disclaimer belongs here because it concerns the same reliability question in non-AI form: the pre-installed statutes of limitation, municipal pre-suit notice requirements, rules and alerts are stated to be for illustrative purposes only, used at the subscriber's own risk, with a recommendation to verify and recreate them.

The supervision commitment is unusual in this corpus because it sits in the customer agreement rather than only in marketing. Section 4(a) of the Terms and Conditions for Neos Customers provides that use of any AI Tool is entirely at the customer's sole discretion and that the Platform will not implement any results obtained from AI Tools unless the customer directs it to do so. Product material matches it: nothing is final until an attorney approves, no demand sends without sign-off, and the generation interface offers regenerate, edit or approve. The scope limit is stated with unusual specificity for a plaintiff product, twice on the Demands Plus page: the tool drafts and does not decide causation, damages, standard of care, or whether a case has merit, which stay with the attorney and the expert. The review surface is concrete rather than rhetorical, since verification means opening a cited source rather than reconstructing where a number came from. Held off A by the two things the band asks for that are absent: no threshold is published at which the system stops or declines, and nothing published addresses what happens after the system is wrong.

A contractual review duty sits beside product features that are designed to operate with no reviewer at all, and nothing reconciles them. The duty is clear: the agreement makes the subscriber fully responsible for reviewing, verifying and ensuring the accuracy, legality and appropriateness of all AI-generated content before using or relying on it. Some features are built around that, with email and SMS assistants drafting responses for attorney or staff review, and the negotiation tools described as surfacing information while leaving legal judgment with the attorney. Against that, three features are autonomous by description. The Intake AI Voice Agent autonomously engages with prospective clients, collects case details and triggers communications. Customisable AI Agents are deployed to run recurring operational tasks without a described checkpoint. The data extraction engine writes court deadlines and coverage details directly into case fields. Searched the AI tools page and the full agreement on 31 Aug 2026 and located no threshold at which any of these stops, no confidence or uncertainty signal, no description of where a human intervenes in the autonomous paths, and no correction route. A review obligation that the product's own architecture bypasses is what holds this at this band.

This is a deep and specific set by the standards of the pull. Named individuals at named firms carry named figures: Bettina Altizer of Altizer Law at a 60 per cent efficiency lift, Kelly Martin of Latham Martin at 80 per cent less data entry, Matthew Shad of Schad and Schad at 90 per cent faster file closing, Sheila Hiestand of McCoy and Hiestand at 50 per cent faster open to settlement, Edward Flynn of Finney Injury Law at twice the medical record review speed, Seth Green of Green Injury Lawyers at $3,000 saved per device setup, and Eva Herman Marcin of Herman Law tracking more than 4,000 active litigation cases. Further attributed customers include Paulson and Nace, Kalfus and Nachman, The Cochran Firm, The Samuels Firm, Queener Law, Becker Kellogg and Berry and Pacheco and Couceiro, alongside portfolio figures of 1,450 plus firms and 15,000 plus attorneys and staff. What holds it at B is the A band's method limb rather than its attribution limb: not one of the figures carries a stated basis, a measurement window or a definition of what was counted, so a reader cannot assess how 80 per cent less data entry was arrived at. The case studies library was not opened this session; under the load-bearing test it would not move the grade, because even a dated case study would supply the date the A band wants and not the method it also requires.

A large adoption figure, no attributed outcome located, and one contractual provision a reader should weigh before crediting any testimonial. The published claim is that the software is used by more than 20,000 legal professionals nationwide, from solo practitioners to Big Law and from private practice to municipalities, which is a scale statement about the vendor rather than a result at a customer. A reviews page exists and was not opened on 31 Aug 2026, and no case study with a named organisation, a figure or a date was located on the pages read. The provision that bears on this axis is the mutual non-disparagement clause in the terms of service, which bars each party from making any statement that could reasonably be expected to negatively affect the other's reputation, credibility or goodwill, whether true or false, to clients, press, investors or on any public forum, including anonymously or through third parties, and survives termination for three years. A customer under that obligation cannot publish criticism, which means published praise cannot be read as freely given. The reviews page is the rebuttal route and should be read alongside this clause rather than instead of it.

Substantive and readable before signing, on both limbs the B band names. Confidentiality is handled at section 9 of the Neos Terms and Conditions with a defined standard of care, an access restriction to personnel with a legitimate need to know who are bound by equivalent obligations, and a commitment at 9(d) to give advance notice of any compelled disclosure so the disclosing party can seek a protective order. The privacy policy, last updated 29 June 2026, adds a specific access limitation: Assembly does not access content entered into the software except to provide technical support as requested or required, as compelled by a legal or regulatory authority, or on an aggregated non-personally-identifiable basis. On training use, section 8(b) confines the processing licence over Customer Data to what is necessary to provide the Platform or to the customer's written instructions, and section 4(b) confines the Input licence to generating Output. Two things keep it off A. Privilege and work product are nowhere addressed by name on any surface read, which is a conspicuous silence in a product that ingests medical records and drafts demands. And section 8(c) grants Assembly a perpetual, irrevocable, sublicensable and transferable licence to process Aggregate Data, defined as aggregated and de-identified customer data, for any lawful purpose, which sits in tension with the marketing commitment that data is never used to train models.

A real confidentiality clause with one strong limb, undercut by an express training permission and silent on segregation. What is published: subscriber data is defined as Confidential Information; each party must protect it to the standard it applies to its own similar information and in no event less than reasonable care; disclosure is limited to employees with a need to know who are informed of the obligations; and the compelled-disclosure limb is genuinely protective, requiring the receiving party to assert the privileged and confidential nature of the information against the third party seeking it and to cooperate in obtaining a protective order narrowing the scope. Security is committed only as reasonably acceptable security and integrity standards, and the subscriber administers its own usernames and passwords. What holds this down. The agreement expressly permits the provider to use client data in aggregated, anonymised or de-identified form to develop, train and improve machine learning models, which is a use of client material rather than a protection of it. Nothing describes separation between matters or clients, no ethical wall model appears, and privilege is named only as something the vendor will assert against a third party demand rather than as something the platform is designed to preserve. Confidentiality survives two years, shorter than most in this category. The privacy policy was not opened.

A real position on advice versus tooling, published in the agreement rather than buried in a footer disclaimer. Section 5(d) of the Terms and Conditions states that Assembly does not offer and does not purport to offer legal or other professional advice and expressly disclaims any warranty that use of the Platform will ensure the customer's compliance with any requirement. Section 4(d) adds that Output is not intended to be and should not be used as a substitute for legal analysis and judgment, that the customer will independently evaluate Output for accuracy including through human review before using or sharing it, and section 13 puts verification of accuracy and completeness before acting squarely on the customer. The product copy draws the same line concretely by enumerating what the system does not decide: causation, damages, standard of care, and case merit. Short of A on two counts. Jurisdiction limits are commercial rather than professional, the contractual Territory being the United States and its territories with no engagement with state-level variation, and the client-facing intake surfaces, which include dynamic intake questionnaires and automated lead emails and texts reaching prospective clients, carry no located disclosure that the respondent is not dealing with a lawyer.

A real position on advice versus tooling, published where it binds, with a conspicuous gap around the one product that most needs it. The agreement states plainly that AI-generated content and data are not intended to constitute legal advice and should not be construed or relied upon as such, and pairs that with a verification duty making the subscriber fully responsible for the accuracy, legality and appropriateness of AI output before use. The statute-of-limitations disclaimer does similar work for non-AI content, telling subscribers that pre-installed deadline data is illustrative only and that they should verify or recreate it against the law of their own jurisdictions. What is absent: no bar or ethics authority is named anywhere, including ABA Formal Opinion 512; no jurisdictional limit is stated; and nothing addresses whether an attorney-client relationship arises. The gap that matters most is specific. The Intake AI Voice Agent autonomously converses with prospective clients about their potential claims, which is the situation in which an unauthorised practice question actually arises, and nothing published addresses what that agent may say, whether it discloses that it is not a person, or how a firm using it discharges its own obligations. Checked the AI tools page and the full terms on 31 Aug 2026.

Nothing whatever was located. Searched the home page, the Neos product page, the pricing page, the IT Leaders solution page, the Demands Plus page, the Terms and Conditions for Neos Customers, the legacy Terms of Service and the privacy policy, together with the navigation and footer of every one of them, on 31 August 2026: there is no responsible AI page, no governance statement, no accountable owner named inside the vendor, no pre-release testing regime described, no ISO 42001 or equivalent, and no disclosure of uneven output across matter types or populations. The absence is more pointed here than the grade alone conveys. This is a product that extracts diagnoses and treatment histories from medical records into structured case fields, flags treatment gaps for attorney attention, drafts the demand that opens settlement negotiation, and feeds case-value reporting from prior settlements on the file. The category's decisive question of what a valuation basis is trained on and whether that basis is disclosed is not answered anywhere located. Security controls, which this vendor does publish, are a different subject and are graded separately.

No governance position is published for a system whose output affects legal outcomes. Searched the AI tools page, the home page and the complete terms of service on 31 Aug 2026 and located no responsible AI page or principles, no accountable owner for model behaviour, no pre-release testing or evaluation regime, no published result, no model documentation, and nothing whatever on bias, fairness or uneven output. The section of the agreement headed SmartAdvocate Artificial Intelligence Tools is not governance: it disclaims liability for output, allocates verification to the subscriber, states that output is not legal advice, and reserves a right to train on de-identified client data. A separate clause reserves a right to monitor AI usage for fair use and to charge, limit or suspend it, which is a commercial control rather than a behavioural one. The absence carries more weight here than on most records because of what the products do: an autonomous voice agent converses with injured people about their claims, a settlement value assistant compares an attorney's valuation against case data, and a negotiation assistant assesses the arguments of opposing counsel in real time. Each is a setting in which uneven performance across claimants would matter, and none is addressed.

Most of the ground is covered and the one canonical gap the B band names is the gap here. Incident practice is contractual and specific: section 8(e) of the Terms and Conditions requires either party becoming aware of a Security Breach to notify the other no more than 48 hours after becoming aware, with delay permitted only where law or a law enforcement instruction requires it. Retention of AI material is the strongest single commitment on the record, the vendor stating that input data is deleted immediately after a response is generated and that no data is held beyond the specific AI transaction. Access control is documented as role-based and individual permissions, with each firm's data isolated in its own Azure environment, and section 8(d) obliges both parties to maintain safeguards to prevailing industry standards. What is missing is a named subprocessor list: the privacy policy refers only to unnamed third-party vendors providing telecommunication, data hosting and data storage, and then disclaims all responsibility and liability for their negligent acts or omissions in capitals. General record retention is also left vague, the privacy policy reserving an archived copy as required by law or for legitimate business purposes with no period stated, while the deletion terms sit in a decommission schedule published at its own URL that was not opened this session.

Retention and deletion are specified with real numbers and everything around them is thin. The agreement commits the provider to back up subscriber data daily and retain backups for seven days, to supply a backup copy on request where the subscriber provides suitable media, and on termination other than for cause to make an SQL export available on written request against a service fee of not less than five hundred dollars, after which the provider has no obligation to hold data for more than seven days and will then irrevocably delete it. Those are stated periods a buyer can plan around, and the deletion commitment is unambiguous. Beyond that the position is asserted rather than described: security is committed only as reasonably acceptable security and integrity standards, the marketing offers bank-caliber data security with constant monitoring, HTTPS in transit, and two-factor authentication that is expressly optional rather than required. Searched the AI tools page, the home page and the full agreement on 31 Aug 2026 and located no encryption-at-rest statement, no access control model, no incident response or breach notification practice, no subprocessor list, and no data processing addendum, none of which exists in the footer inventory. The privacy policy was not opened and is the rebuttal route.

A complete and specific allocation of loss that a buyer can read in full before signing, which puts it well above the several vendors in this lane that publish no customer agreement at all. Section 12(a) gives a real indemnity, with Assembly defending and holding the customer harmless against third-party claims that Assembly Materials infringe intellectual property, subject to a defined Exception carve-out, with remedies of procuring a licence, modifying the materials, or terminating with a pro rata refund. Section 11(b) warrants that Services will be performed in a competent and workmanlike manner with reperformance and then refund as the exclusive remedy. Section 14(b) caps aggregate liability at fees paid or payable in the preceding twelve months, and 14(c) lifts that cap for fraud, gross negligence, willful misconduct, violation of law and the indemnification obligations, which is a named carve-out set rather than a bare cap. It stops short of A because of what the vendor does not stand behind: section 13 disclaims any warranty as to the quality, accuracy, currency or completeness of what the Platform produces, section 4(d) states that Assembly does not represent that Output will be accurate and complete, no insurance is named, and the indemnity reaches infringement rather than a wrong answer. Note for the reader that the separate legacy Terms of Service governing Needles and TrialWorks is materially worse on this axis, capping at one month of fees and disclaiming indemnities outright.

Liability is addressed at length, entirely in one direction, and the agreement contradicts itself on the central number. Two different caps appear. The limitations section states that in the event of any breach by the provider, liability will not exceed the subscription fees paid in the previous three months. The arbitration section states that total liability for any claim arising out of the agreement or the services is limited to the fees paid in the twelve months preceding the event. Nothing reconciles them, and a buyer should raise it before signing. Around that: all warranties other than authority are expressly disclaimed; the AI section disclaims all liability for any content, data, recommendations or outputs generated by the AI tools, including those of third-party vendors; the limitations clause purports to exclude liability for negligence, carelessness and gross negligence alike; the only indemnity runs from subscriber to provider covering use, misuse or non-use of the software; and the subscriber expressly represents that it will not bring or implead the provider into any litigation regarding claims arising out of its practice of law, which is a covenant not to sue rather than a limitation. No vendor indemnity, no warranty on output, no insurance position and no service level commitment with a remedy were located. Disputes go to individual JAMS arbitration in Nassau or Suffolk County, New York, with a class waiver and no punitive damages.

The integrations are real and several are named, but the marketing and the agreement point in different directions and the agreement decides it. The IT Leaders page sells 'a documented API and built-in integrations', while section 2(c) of the Terms and Conditions is headed No API Access and provides that the Agreement grants no rights to access or use APIs, SDKs, connectors or other development tools, and that any such access is subject to Assembly's approval in its sole discretion and to additional terms. A buyer cannot rely on the API the marketing offers. On the built-in side the named set is substantial, with native Microsoft 365, Outlook, SharePoint and OneDrive, plus Case Status, Hona, RingCentral, Zoom Phone, DocuSign, Zapier, YoCierge and LawPay, and a handful carry a one-line description of what moves, such as calls logged to the case and documented in the Communications Center. But on the home and product pages they are presented as a logo strip, no documentation an implementer could use was located on any public surface, and nothing states what syncs in which direction or what a firm must configure. The knowledge base sits behind a separate support portal and was not opened; given section 2(c) it would not lift the grade to A.

Breadth is claimed and named connections exist, with the depth on a page that was not opened. The vendor states more than 175 integration partners across artificial intelligence, e-signature, accounting, lien resolution, technology and reporting services, medical record retrieval, answering service and call tracking, and marketing, which for a plaintiff practice covers the operational estate that actually surrounds a case. An open API is offered. Independent listings name Microsoft Outlook, DocuSign, Docubee, QuickBooks Online, CallRail, Keap and Sinch MessageMedia, with connections described to court docket systems and medical record retrieval services. The most current integration surface is published on the vendor's own page: an MCP server that lets a firm's technical team connect external AI platforms to SmartAdvocate case data and documents, which points outward at the buyer's own AI stack rather than inward at a fixed connector list. One caveat belongs on the record because it affects reliance: the agreement provides that any integration may be withdrawn or cancelled at the provider's sole discretion, with only best efforts to give thirty days' notice, and that the provider makes no representation about any third party's service or cost. The integrations and partners page was not opened on 31 Aug 2026.

This axis asks two questions and the vendor answers one of them well and the other not at all, so neither the B nor the C band fits cleanly and the lower is taken with the reason recorded. The deployment model is stated clearly and repeatedly: cloud only, hosted on Microsoft Azure, with each firm's data isolated in its own Azure environment and no on-premise option offered, which is a real tenancy statement and more than most records in this pull carry. Residency is the failure. No region is named anywhere, nothing distinguishes where processing happens from where data is stored, nothing states where the Azure AI services that run NeosAI are located, and no residency options are offered at any tier. What is published actively cuts against a residency commitment: the privacy policy states that personal information may be transferred to, collected, processed and stored in the United States and other locations where Assembly or its service providers operate. The contractual Territory of the United States and its territories is a restriction on where the customer may use the software, not a commitment about where the data sits. The C band's words about tenancy are false here and the note carries that; the grade reflects a residency position that does not exist.

Two genuine deployment models are published, and the self-hosted one answers the residency question by handing it to the customer. SmartAdvocate is sold both as a hosted cloud service and as a server-based installation on the subscriber's own infrastructure, each with its own technical specifications page, and the vendor states that with the server-based version the server and data can be located anywhere. That is a real choice rather than a marketing distinction, and the agreement supports it commercially by giving a subscriber moving from cloud to server a credit equal to fifty per cent of the subscription fees paid in the preceding twelve months. For a firm with residency or control requirements, taking the data in house is the answer, and few vendors in this pull offer it. What is not published is anything about the cloud side: no hosting provider is named, no region or data centre location is given, no residency option or commitment appears, and no tenancy model is stated. Searched the AI tools page, the home page and the full agreement on 31 Aug 2026; the cloud technical specifications page was not opened and is the rebuttal route.

The attestation is claimed in the vendor's own name rather than inherited, which distinguishes it from the pattern this lane keeps producing, but none of the evidence a buyer would need is reachable. Assembly states that it maintains SOC 2 Type II certification with quarterly penetration testing, repeated on the home page, the IT Leaders page and the product FAQ. Against the A band, everything that would make it verifiable is absent: no auditor is named, no coverage period or report date is given, no scope statement says which systems or which Assembly products the certification covers, no penetration test summary is published, and no trust portal or report request route was located in the navigation or footer of any of the eight surfaces read on 31 August 2026. There is no gated portal either, so the middle access tier does not arise; there is simply nothing to request. This is the same shape recorded on Bloomberg Law and lands in the same place, with the small difference that the Type is stated here and was not there. Quarterly penetration testing is a real operational detail and is what keeps this above a bare badge claim.

No independent security attestation of any kind was located. Searched the AI tools page, the home page, the full terms of service and the complete footer inventory on 31 Aug 2026. There is no SOC 1, SOC 2 or SOC 3, no ISO certification, no HIPAA or CJIS claim, no penetration testing statement, no named auditor, no trust portal and no security page anywhere on the property; the footer carries only a privacy policy and the terms of service. What stands in its place is adjectival: bank-caliber data security, constant monitoring by the vendor's IT professionals, HTTPS so that data is encrypted in transit, and two-factor authentication described as something the software can be used with rather than as a requirement. The agreement's own security commitment is a standard of reasonableness, undertaking to maintain the security and integrity of the software and subscriber data using reasonably acceptable security and integrity standards. For a platform holding the case files of more than 20,000 legal professionals, including medical records and settlement information, the absence of any third-party attestation is the finding, and it is the sharpest gap on this record.

The provider is named and the architecture is described, which is the first B limb almost exactly. The IT Leaders FAQ states that NeosAI uses Microsoft Azure AI services under a zero-data-retention commitment, that input data is deleted immediately after a response is generated, and that no client or medical records are exported to a separate AI vendor and no data is stored outside Neos. That is a clearer account of where inference happens relative to the customer's data boundary than most vendors give, and it is corroborated by the platform's own Azure hosting. What is missing is the other half of the A band. No specific model is identified anywhere, the vendor referring only to Azure AI services as a family; nothing states which model version is in use for which feature; and there is no commitment to notify customers when any of it changes, which matters more than usual on a two-month release cadence the vendor advertises. A Microsoft case study published on the vendor's blog was named but not opened this session; it could not reach the A band, since the change-notification limb would still be unmet.

Providers are named in the agreement, which puts this ahead of most of the pull, and nothing follows them. The artificial intelligence section states that the tools include content and data generated using OpenAI technology, proprietary technology developed by SmartAdvocate, and other third-party AI technologies, and adds that third-party vendors whose technologies integrate with or use data within the platform offer further AI tools to customers. Naming a foundation model provider in a customer contract is rare, and only one other vendor in this pull does it. Separately, the MCP server is described as letting technical teams connect external AI platforms, with Anthropic Claude given as the example, so a buyer can also see which external systems the architecture is built to accept. What is not established: no model or version is identified, the phrase other third-party AI technologies leaves the set open-ended, no inference location is stated, no subprocessor list exists anywhere on the property, and no commitment to notify customers of a change of model or provider was located. The agreement instead reserves a general right to upgrade, modify, replace or remove features on thirty days' notice where a change materially and adversely affects use.

Real pricing for part of the range with the upper tiers withheld, which is the B band verbatim. The pricing page names three tiers, publishes a figure for the entry tier at $109 per user per month billed yearly, states the unit of charge, and carries a full feature comparison table across all three covering practice management, NeosAI, document management, communications, intake and support. It also publishes a floor most vendors leave to a sales call, stating that Neos is built for firms with three or more users and that smaller practices will be routed elsewhere, and it names five paid enhancements including ePayments through LawPay with no activation fee. The FAQ repeats that pricing starts at $109 per user paid annually. The specific limitation a buyer should notice is that the published figure buys the tier without the AI: NeosAI is a Platinum feature and Demands Plus requires Platinum plus the add-on, and neither Premium nor Platinum carries a number, so the price of the capability this index grades is not published at any level. Nothing states what implementation adds, and the legacy Terms of Service is where the incidental charges appear, including a $250 data retrieval fee.

The unit of charge and a set of real ancillary figures are published, and the rate itself is a phone call. What a buyer can establish without contacting anyone: the model is a per-user subscription against a Work Order specifying the number of users and level of service, with users addable or removable at any time and reductions effective from the next billing cycle; payment is in advance with a credit card held on file under a non-revocable charge authorisation; unpaid amounts accrue interest at the lesser of 1.5 per cent per month compounded quarterly or the legal maximum; reinstatement after six months of non-payment costs the arrears plus twelve months up front plus a one thousand dollar restoration fee; a data export on termination costs not less than five hundred dollars; support outside 9am to 5pm Eastern is chargeable, as is support found not to arise from the software, at the provider's standard hourly rate; and moving from cloud to server-based earns a credit of fifty per cent of the prior twelve months' fees. Those are specific and unusual to publish. What is nowhere published is the subscription rate: there is no pricing page on the site, and the vendor directs buyers to call a business development specialist for a tailored quote. The agreement also makes pricing confidential, barring either party from disclosing the terms including pricing without consent.

Coverage is documented precisely and, unusually, so are the limits, which is the part of the A band most vendors fail. Six practice areas have their own pages: personal injury, auto accident, medical malpractice, mass tort, workers' compensation and premises liability, with nursing home abuse covered through a named customer and a further post for other practice areas. Five roles have their own pages: managing partners, attorneys, firm administrators, paralegals and IT leaders. The boundaries are stated rather than left open. The pricing FAQ says plainly that Neos is built for firms with three or more users and routes smaller practices to a different Assembly product, the Terms and Conditions confine the Territory to the United States and its territories, and the positioning declines general practice outright, the comparison section contrasting Neos with platforms built for many practice areas and optimised for none. The in-house and government limb of the A band does not apply to a product sold exclusively to plaintiff firms and is not graded. One inconsistency is recorded rather than graded: the privacy policy describes Assembly's customers as law firm, corporate, government and other, which does not match the plaintiff-only positioning on every Neos surface.

Segment coverage is documented in detail and the boundary is drawn once, in the agreement rather than the marketing. Thirteen practice areas have their own pages: employment law, personal injury, mass torts, workers' compensation, veterans affairs, municipalities, social security, solo practitioners, family law, criminal defence, fraud, corporate law and in-house legal teams. Firm size is addressed from solo practitioners upward, and the vendor states its users run from solo practice to Big Law and include municipalities, with a claimed base of more than 20,000 legal professionals. The product is offered in two deployment forms so a firm can choose based on its own infrastructure. One real limit is stated and it is worth recording because so few vendors state any: the agreement provides that the unlimited document storage is solely for documents directly related to individual cases opened in the application, and expressly may not be used to store e-discovery documents or data, or data not tied to a specific case. That tells a buyer where the product stops. What is still absent is any statement of jurisdictional coverage, any practice area identified as unsupported, and any matter size threshold. The individual practice-area pages were not opened on 31 Aug 2026.

The agreement permits aggregated and de-identified use of customer data for any lawful purpose, which the marketing does not mention. The no-training commitment is stated plainly on the IT Leaders solution page, and the FAQ there attributes it to a DPA and BAA that are not published. The published Terms and Conditions for Neos Customers approach the question from three directions: section 8(b) confines the Customer Data processing licence to providing the Platform or the customer's written instructions, section 4(b) confines the Input licence to generating Output, and then section 8(c) grants Assembly a perpetual, irrevocable, sublicensable and transferable licence to process Aggregate Data, defined as customer data that has been aggregated and de-identified, for any lawful purpose. That last licence is the operative permission and no opt-out from it was located. The qualifier is that it runs on aggregated and de-identified material rather than identifiable client content. Moved from policy-never on 1 September 2026.

The agreement expressly reserves the right rather than restricting it. Section 19(d) of the Terms of Service, effective 1 January 2026, provides that any data, information or other content input into or otherwise provided through the Software may be used for the purposes of developing, training and improving machine learning models, artificial intelligence systems and other related technologies. Two qualifiers are stated and both are recorded here rather than treated as the answer. The permission is limited to aggregated, anonymised or de-identified form, and the provider undertakes not to use client data in a manner that identifies the client or its customers and not to disclose it to third parties except as required to provide the Software or as otherwise permitted under the agreement. What is not stated anywhere is any means of declining. Searched the AI tools page, the home page and the complete terms of service on 31 Aug 2026 and located no opt-out mechanism, no consent step, no configuration setting and no route to request exclusion, so a subscriber that objects has no published alternative to not using the platform. The privacy policy was not opened.

The vendor publishes a zero-retention position on AI transactions, stating that data is never held beyond the specific AI transaction and that no data is stored outside Neos, with the commitment attributed to Microsoft Azure AI services operating under zero data retention. The period is specific and immediate, and nothing indicates the customer can configure it, which is why this records as a disclosed fixed window rather than customer-controlled. Note the scope: this addresses what the AI layer retains, not what the case file retains, since outputs are written back into the matter and persist there as the firm's own record.

Retention is stated with real periods for the data lifecycle and nowhere for AI interactions specifically. The agreement commits the provider to have subscriber data backed up daily and retained for seven days, and to supply a backup copy on request where the subscriber provides suitable media. On termination other than for cause, and provided all sums are paid, an SQL export of subscriber data as at the termination date is available on written request against a service fee of not less than five hundred dollars, after which the provider has no obligation to hold the data for more than seven days and will then irrevocably delete it. Where the provider terminates for breach it disclaims any obligation to maintain the data or provide access to it at all. Nothing published addresses prompts submitted to the AI tools or the content those tools generate as a separate class: no retention period, no customer-configurable setting and no zero-retention option was located on the AI tools page or in the full agreement on 31 Aug 2026. The privacy policy was not opened.

Neos is the system of record rather than a layer over a separate document management system, so there is no external access model for it to inherit and it maintains its own. That model is described at two levels: tenant separation, with each firm in its own Azure environment, and within the firm role-based and individual permissions set so that the right people see the right cases. Nothing located ties NeosAI retrieval to those permissions, so whether a chat query or a generated document respects a case-level restriction at query time is unstated.

Nothing published addresses separation between matters or between clients. Searched the AI tools page, the home page and the complete terms of service on 31 Aug 2026 and located no ethical wall concept, no matter-level permission model, no statement that a user can be restricted to particular cases, and nothing about whether the AI features respect any such restriction. Access control is handled at the account layer and delegated entirely to the customer: the subscriber is responsible for providing and administering usernames and passwords for all its employees and other users, each user must have valid credentials, and the subscriber is responsible for all activity occurring under its accounts. The question has particular weight for this product because several AI features read across the case file and because the natural-language query tool is described as working across SmartAdvocate data and documents without any stated scope limit. Nothing was located describing whether that query is bounded to a single matter.

The commitment appears twice and in both a policy and a contract. The privacy policy states that unless prohibited by law the vendor will notify the customer promptly on receiving a subpoena or other legal process seeking access to their data. Section 9(d) of the Terms and Conditions goes further on mechanism, requiring advance notice of a compelled disclosure to the extent permitted by law so the disclosing party can seek a protective order, and requiring reasonable steps to minimise the extent of the disclosure. No transparency report or count of requests received was located, which is what separates this from the top value.

One of the stronger compelled-disclosure provisions located in this pull, and it is contractual. Subscriber data is Confidential Information under the agreement, and where a party is legally compelled to disclose it, that party must first assert the privileged and confidential nature of the information against the third party seeking disclosure and cooperate fully with the other party in protecting against disclosure and in obtaining a protective order narrowing its scope and use, disclosing only as and to the extent necessary if that protection is not obtained. A related provision requires no less than fifteen days' prior written notice to the other party where legally permissible, so the customer has a window to seek protective relief rather than learning after the fact. Two limits: no transparency report or disclosure statistics were located on 31 Aug 2026, and the confidentiality obligations survive only two years after termination, which is short for material that may remain privileged indefinitely.

This signal does not fit the product and the honest value is the absence. Neos retrieves no primary law: the corpus its AI works from is the firm's own case file, the records, notes, emails, call logs and case values the firm has already put there, and the vendor is explicit that Demands Plus draws on the live case rather than an external body of material. Searched the home page, the Neos product page, the Demands Plus page, the IT Leaders page and both agreements on 31 August 2026; nothing identifies any external corpus, licensed or otherwise, because none is claimed.

Most of what the AI reads is the firm's own case file, where provenance is not in issue, but the platform does ship a body of legal data and its provenance is unstated. The software is supplied with pre-installed statutes of limitation, municipal pre-suit notice requirements, notice of claim requirements, rules, alerts, forms and picklists. Nothing published identifies where that material came from, which jurisdictions it covers, how it is compiled or on what basis it is licensed. The agreement instead disclaims it, stating it is provided for illustrative purposes only, that the subscriber uses it at its own risk, and that the provider is not responsible for maintaining, correcting, changing or updating it. One provision is worth recording because it describes a corpus-building mechanism running the other way: the provider reserves the right, if it develops the technology, to receive automatically and remotely the changes a subscriber makes to that pre-installed data, for the stated purpose of providing accurate data to its subscribers, and the subscriber consents in advance and agrees not to block the transfer without payment or credit.

Not applicable to this product class and recorded as the absence rather than skipped. Neos does not retrieve or cite legal authority; the citations its demand drafts carry point to documents, notes and values on the case file, not to cases or statutes, so there is no subsequent history for a citator to check. Searched the Demands Plus page, the Neos product page and the pricing feature table on 31 August 2026 and located no citator, treatment signal or good-law claim of any kind.

No citator or currency check is claimed, and the product class does not raise the question in its usual form since the AI reads the firm's own case records rather than primary law. Recorded as not addressed with the reason stated. The analogue that does bite is deadline currency, and it is addressed only by disclaimer: the pre-installed statutes of limitation, municipal pre-suit notice requirements and rules carry an express statement that the provider is not responsible for maintaining, correcting, changing or updating them, and the subscriber bears an affirmative obligation to notify the provider within ten days of discovering an error or learning of an update in any jurisdiction. For a plaintiff case management system, whether a limitation period is current is the closest equivalent to whether authority is still good law, and the answer published is that keeping it current is the firm's job. Searched the AI tools page, the home page and the full agreement on 31 Aug 2026.

What is documented is a categorical refusal of judgment rather than an inability-to-ground abstention path. The Demands Plus page states twice that the tool drafts and does not decide causation, damages, standard of care, or whether a case has merit, and that those remain the attorney's and the expert's. A second feature is adjacent to uncertainty: treatment gaps, meaning breaks in the treatment timeline, are detected and surfaced for review rather than smoothed over. Nothing located describes what the system does when it cannot ground a claim in the record, and no confidence or grounding score is exposed, so the specific failure this signal tracks is not addressed even though abstention behaviour is.

Nothing published describes what any of the AI tools does when it is unsure. Searched the AI tools page, the home page and the complete terms of service on 31 Aug 2026 and located no abstention path, no confidence or uncertainty indicator surfaced to a user, no threshold at which a feature declines to act, and no statement of behaviour where the case file does not support the output requested. The agreement handles the risk by allocation rather than description, disclaiming any warranty as to accuracy, currency, completeness or usefulness and placing full responsibility for verification on the subscriber. The absence is most consequential on the autonomous features, since the Intake AI Voice Agent conducts a conversation with a prospective client in real time and the customisable agents run recurring tasks unattended, and in neither case is there a published account of what happens when the system encounters something it cannot handle.

Searched the AI Hallucination Cases database maintained by Damien Charlotin, and secondary trackers and audits reporting from it, on both the product name Neos and the company name Assembly Software on 31 August 2026. No court order, opinion or disciplinary record naming either was located; the tools named in reporting on the database are general-purpose assistants and legal research products. This is a statement about the public record rather than a finding about the product. Note for the reader that the failure mode this signal tracks fits Neos poorly: it generates no citations to legal authority, its citations point to the firm's own case record, and its principal output is a demand letter addressed to an insurer rather than a filing.

No court order, opinion or disciplinary record naming this product has been located as of 31 Aug 2026. Instrument searched: the AI Hallucination Cases database maintained by Damien Charlotin, which tracks decisions worldwide where a court addressed hallucinated AI content and records the tool implicated where known, alongside several independent 2026 sanctions trackers and practitioner guides, searched on the company name and the product name. This is a statement about the public record on the date shown rather than a clearance. The exposure profile differs from a research tool: the AI here drafts demand letters and extracts deadlines from documents rather than citing legal authority, so the failure this signal tracks would more likely reach a court through a filing drafted elsewhere. The nearer risk on this product is a missed or mis-extracted limitation date, which would surface as a malpractice claim rather than as a hallucination case.

Professional responsibility is engaged with in general terms and no guidance is named. Section 5(d) of the Terms and Conditions states that Assembly does not offer or purport to offer legal or other professional advice and disclaims any warranty that use of the Platform will ensure compliance, section 4(d) requires the customer to evaluate Output independently including through human review, and the Demands Plus material enumerates the judgments reserved to the attorney. No ethics opinion is named anywhere located, including ABA Formal Opinion 512, and no state bar guidance is mapped or referenced. Searched both agreements, the privacy policy and the five product and solution pages on 31 August 2026.

No professional conduct authority is named anywhere. Searched the AI tools page, the home page and the complete terms of service on 31 Aug 2026 and located nothing citing ABA Formal Opinion 512, any state bar guidance on generative artificial intelligence, or any rule governing supervision of AI-assisted work. The agreement states the substance of one duty without sourcing it, providing that AI content is not intended to constitute legal advice and that the subscriber is fully responsible for verifying accuracy, legality and appropriateness before use. The gap is sharpest around client contact rather than drafting: the Intake AI Voice Agent autonomously converses with prospective clients and collects details of their potential claims, which engages questions about solicitation, disclosure that the caller is not a person, and the firm's supervisory duty over a system speaking in its name, and none of those is addressed on any page located.

Time savings are claimed prominently and repeatedly, and the client side of the equation is not addressed anywhere located. The published claims include a demand draft in about fifteen minutes against the four to eight hours the vendor says the task takes by hand, described as roughly thirty times faster, alongside customer figures of 80 per cent less data entry and 90 per cent faster file closing. Nothing published addresses what happens to a bill or a fee when AI-assisted work compresses, and no per-matter record of AI-assisted work is described. The buyer here is typically a contingency practice, where the hourly question takes a different shape, but the vendor engages with neither shape. The expense tracking, invoice generation and case cost recovery features address the recovery of case costs, which is a separate question from this one.

Efficiency is the pitch and the client's bill is not addressed. The published claims are general rather than quantified: the software streamlines processes, makes everyday tasks easier to manage, increases profits and efficiency, and reduces manual work. Searched the AI tools page, the home page and the complete terms of service on 31 Aug 2026 and located no per matter record of AI-assisted work intended for fee purposes and no guidance on billing, fee or client disclosure treatment where AI-generated work informs what a client is charged. The question is live for this buyer base, since a contingency firm recovers case costs from the settlement and an AI-drafted demand letter is work that would otherwise be billed or absorbed. One provision touches cost without touching disclosure: the agreement reserves the right to monitor AI tool usage for fair use and to impose an additional charge based on usage, so the cost of the AI to the firm is itself variable and unstated.

The material is stated to exist but sits behind an agreement. The IT Leaders FAQ says that Assembly's DPA and BAA document the zero-data-retention commitment, which names the artifacts without publishing them, and no request route or self-serve portal was located. What is published and would help a firm answering a client's AI clause is real but partial: Microsoft is named as the AI service provider, the vendor states that no client or medical records are exported to a separate AI vendor, and the zero-retention position is specific. What is absent is the artifact the value set turns on, a current subprocessor list; the privacy policy refers only to unnamed third-party vendors for telecommunication, data hosting and storage.

None of the artifacts a client's AI clause asks for exists on this property. Searched the complete footer inventory, which carries only a privacy policy and the terms of service, together with the AI tools page and the home page on 31 Aug 2026: there is no subprocessor list, no data processing addendum, no security page, no certification of any kind, no trust portal, no AI policy or disclosure pack, and no stated route for requesting any of them. What a firm could forward is limited to the terms of service themselves, which do contain two provisions a client might accept, the compelled-disclosure protections and the naming of OpenAI technology as a source behind the AI tools, and one it would very likely query, the express permission to use de-identified client data to train machine learning models. A firm asked by a client which third parties process its matter data would have no list to point to and no one identified to ask.

Some of the elements a disclosure would need are produced as a by-product of how the drafting works, which is more than most products in this pull can say. Every claim in a demand draft carries a citation that opens the exact source on the file, so what was retrieved for a given passage is recorded in the document itself, and the document stays attached to its case and can be reopened. Two elements are missing. No model is identified, so the model-used limb cannot be answered from the product. And while the vendor states that an attorney reviews, edits and approves every demand before it goes out, nothing located indicates that the approval is captured as an exportable record of who verified what and when, as distinct from a workflow gate.

Nothing published would help a lawyer disclose or certify AI use. Searched the AI tools page, the home page and the complete terms of service on 31 Aug 2026 and located no model identification or versioning beyond the general naming of OpenAI technology and other third-party AI, so which system produced a given output cannot be established; no audit trail or log of AI invocations is described; nothing records who reviewed or accepted generated content; and no export, template or guidance exists for a court's standing order on AI use or for a certification of verification. Medical chronologies are described as hyperlinked, which links an entry to its source record and is the only traceability located, but nothing states that the link set is exportable or that it survives outside the platform. The gap widens on the autonomous features, since neither the voice agent conversations nor the custom agents are described as producing any record a firm could later produce.