Mary Technology vs Wexler AI: how they compare in 2026

Remove the models and nothing remains to sell. The product is a reading and extraction engine: it takes the case files as they arrive, including poor scans and old formats, reads every page, extracts each individual fact rather than storing documents as embeddings, tags it with people, dates and events, links it to its source page, assigns it a status of supported, contradicted or unconfirmed, and surfaces contradictions, unexplained periods and material referenced elsewhere but absent from the production. There is no document management system, no review platform and no matter management layer underneath that a buyer would license on its own; documents are uploaded or pulled from systems the firm already has, and what Mary adds is the structured factual record and everything built from it. The vendor's own architectural claim reinforces it, distinguishing per-fact extraction with a traceable link to source from compression-based approaches. Checked 5 September 2026.

The models are the entire product. Fact extraction across a document set, contradiction detection, chronology construction and the Kim assistant are all model outputs, and the platform is positioned as reasoning at the scale of a million pages rather than as a repository with analysis attached. There is no document management, workflow or matter system underneath that would function without them. The company's own framing, a fact intelligence platform that analyses every fact in every document to show what the evidence supports, describes a capability that does not exist without the models. Checked 4 September 2026.

The verification architecture is documented in more detail than anything else in this lane, and nothing about it is measured. Grounding is per fact rather than per answer: every extracted fact is traceable to its source page, questions asked in plain language return answers with one-click citations to the exact facts and document pages, and the original source can be opened against any entry in a click. Beyond grounding, the product publishes an epistemic status on each entry, marking it supported, contradicted or unconfirmed, and it refuses to collapse competing accounts into a single answer, keeping contradictions and unexplained periods visible on the record. Coverage is itself reported, so a user can see what was processed, what informed an entry, what was reviewed but unused and what remains unprocessed. That combination answers the hallucination question structurally. What is absent is any test a reader can run: no accuracy figure, evaluation, test set or error rate is published. The nearest thing is a preference study, eight of nine litigators preferring Mary's output in a head-to-head matter review, at an unnamed global firm, undated and with no method, which is a preference measure rather than an accuracy one.

Grounding is real, documented and central to the product's claim rather than incidental to it. Answers are described as sourced facts, confidence-scored and grounded entirely in the customer's own documents, every fact is traceable to its source passage, and an arbitration case study records that the system explained where in a document it took each date or event from. Testable at trial is used as a design goal rather than a slogan. What is absent is a vendor-published measurement. The one accuracy figure on the site is a customer's own, quoted in a testimonial from an unnamed national firm reporting that in its testing the hallucination rate was less than 0.1 per cent and the tool was at least on a par with human review; that is a customer's test with no test set described, no methodology and no attribution to a named organisation, so it is recorded rather than credited as a published accuracy figure. No failure mode is named by the vendor anywhere.

The control structure is published in full and the answer to the threshold question is that there is no autonomous mode. Proposed updates remain pending until a person approves them, which is stated as a product mechanism rather than a promise, and the published workflow makes human review its own numbered stage: the team inspects sources, reviews contradictions and gaps, sets status and approves changes before anything becomes work product. What the system does alone is stated, reading every page and structuring people, dates, events and factual entries. What constrains it is stated, in terms the vendor puts plainly: Mary does not decide the case, it shows the record including what is missing and what does not add up so the lawyer can, and legal judgment remains with the lawyer. The review surfaces are the same ones that carry the product's value, since every entry can be opened against its source and its status set by hand. The agreement carries it into obligation at clause 11(d), requiring that all outputs be reviewed and approved by a qualified legal professional before any action is taken. What is still missing is the failure account: nothing describes what happens when an entry is wrong, or how a firm would detect a systematically mis-read document set.

Oversight is implied by the product's design and never described as a structure. The confidence score attached to each answer and the traceability of every fact to its source are genuine verification surfaces, and a testimonial describes the system generating follow-up questions of the kind a reviewer would expect from a trainee. But nothing published states what the system does unattended against what a lawyer approves, no threshold is given at which it defers or stops, no review workflow is described, and there is no written commitment anywhere that the models operate alongside a supervising lawyer. With no terms of service published, there is no contractual statement of the review obligation either. Searched the home page, the security page, the privacy policy and the site navigation on 4 September 2026; the platform page was not opened.

Aggregate scale claims and unattributed quotations stand in for a named deployment. The figures published are counts rather than outcomes: more than four thousand lawyers and more than one hundred and fifty firms, neither dated nor tied to any organisation. The one comparative claim is closer to evidence and still falls short, reporting that eight of nine litigators preferred Mary's output in a head-to-head matter review at a global law firm, with the firm unnamed, no date, no description of the matter and no method for how preference was assessed. Testimonials on the self-serve site are quoted without a name, a role or a firm, including an account that work which previously took half a day to a full day now takes under fifteen minutes. No customer is named on any surface read, and no dated deployment or measured outcome was located. A customer stories page exists in the navigation and was not opened in this pass; it is named here as the limit and is the cheapest available upgrade on this record. Trade press names Australian firms as customers; it is third party and was not used.

Real deployment evidence with substance, and the usual split between the names and the numbers. Named customers are unusually senior for a company this size: Clifford Chance, Goodwin, HSF Kramer and Burges Salmon appear as customers, with attributed quotations from Claire Freeman, a partner at Clifford Chance, and Tom Whittaker, a legal director at Burges Salmon, alongside an endorsement from Lord Wilson of Culworth. Relationships are dated through the insights journal, with Goodwin deepening its partnership on 3 June 2026 and Burges Salmon extending a firm-wide partnership on 11 May 2026. Figures exist and are specific, but attach to unnamed firms: an AmLaw 100 matter covering more than 250,000 pages with a 90 per cent average time saving and 100 per cent of users saying they would adopt it, and an international arbitration at a global firm where identified inconsistencies were relied on at the hearing. No figure is tied to any named firm, and the case study cards were not opened in this pass.

Segregation is published at the level this buyer segment actually needs, and the privilege limb is absent. The security commitment is specific and stated on the product page rather than buried: customer data is isolated by matter, with separate access and data boundaries for each matter, encrypted in transit and at rest, access following the firm's own permissions with single sign-on, and documents, matter content, prompts and outputs never used to train AI models. Matter-level isolation is what a litigation practice with screened matters requires, and very few records in this corpus claim it. Clause 19 of the terms carries part of it contractually, describing tenanted AI services under which confidential client information is not accessible for model training or by unauthorised personnel. Three things hold the grade here. No privilege or work product treatment appears anywhere, on a product whose entire purpose is holding the factual record of contested matters. No model provider is named, so nothing addresses what any third party retains. And the isolation claim sits on a product page with no supporting documentation, since the security page defers detail to a trust centre.

Confidentiality is asserted in general terms and there is no commitment a buyer can read in advance. The security page states a deep commitment to best-in-class security and a heavy emphasis on rigorous measures for data protection, secure operations and cloud security, which is assertion rather than commitment. Nothing published addresses the questions this axis asks. There is no statement anywhere about training on customer documents, no retention or deletion position for matter material, no description of segregation between customers or between matters, and no position on what any model provider may retain. Privilege and work product are not mentioned, on a product built to ingest the entire evidential record of a dispute. The only published legal document is a website privacy policy covering names, email addresses, IP addresses and usage data, which never reaches customer documents. Searched the home page, the security page, the privacy policy and the footer on 4 September 2026; no terms of service, master agreement or DPA is published anywhere on the site.

A real published position, stated in the agreement and echoed as an editorial line across the site. Clause 11(d) provides that the software is designed to assist with legal tasks but is not a substitute for professional legal advice, analysis or judgment, that all outputs and recommendations must be reviewed and approved by a qualified legal professional before any action is taken or decision made, and that the vendor expressly disclaims liability for actions taken without appropriate legal review. It also declines to guarantee the accuracy, completeness or suitability of output for any particular purpose. The site carries the same idea in its own words on every page, that lawyers remain responsible and legal AI should make its work easy to check before they sign, and the product material repeats that Mary does not decide the case and that legal judgment remains with the lawyer. That is a coherent and consistently held account of where the product stops. What is absent is the professional layer: no bar association, rule of professional conduct or ethics opinion is named anywhere, no jurisdiction is identified, and nothing addresses a firm's supervision or competence duties.

Nothing is published on the advice line for a product that produces factual analysis relied on in filings and at hearings. No statement anywhere says what the output is and is not, no disclaimer distinguishes analysis from advice, no competence or supervision language appears, and no jurisdiction limit is stated despite the product being sold in the United Kingdom and the United States. The absence is not the result of an unread document: there is no terms of service, no customer agreement and no acceptable use policy published on the site at all, and the entire legal estate consists of one website privacy policy. The marketing does the opposite of caveating, describing output as testable at trial and citing an arbitration in which identified inconsistencies were relied on at the hearing. Checked home page, security page, privacy policy, footer and site navigation on 4 September 2026.

No governance position was located. There is no responsible AI page, no principles statement, no accountable owner or function named for model behaviour, no pre-release evaluation regime, no management system and no certification such as ISO 42001, and nothing addresses uneven output. The vendor does publish an editorial position on verification, arguing that verification in legal AI is a design problem and that legal work should be easy to check before a lawyer signs, and that position is genuinely held across the product; it is a design philosophy rather than a governance framework, naming no reviewer, no criteria and no results, and it is credited on the accuracy and autonomy rows rather than counted twice here. The gap has a specific shape on this product: a system that assigns each fact a status of supported, contradicted or unconfirmed is making a judgement on every entry, and nothing published describes how that classifier was evaluated or how it behaves across document types, jurisdictions or the poor-quality scans the product advertises handling. Navigation and footer inventoried 5 September 2026.

No governance position is published for a system whose output shapes what a party argues the evidence shows. There is no responsible AI statement, no governance framework, no named owner, no description of pre-release testing or evaluation, no published results, and nothing whatever on bias or uneven performance across document types, languages or matter kinds. The confidence score attached to answers is a product feature rather than a governance mechanism, and ISO 27001 and SOC 2 are information security standards that do not answer this axis. Searched the home page, the security page, the privacy policy and the site navigation on 4 September 2026.

Access and testing are addressed, and retention, deletion and the supplier picture are not. What is published: client data encrypted throughout its lifecycle, annual penetration tests and continuous vulnerability assessments, background checks and access reviews for staff, mobile device management, a cybersecurity training programme, matter-level data boundaries, and access governed by the firm's own permissions with single sign-on. What is missing is most of what a buyer would need next. No retention period is stated for documents, extracted facts, prompts or outputs, only that data is retained as long as necessary and securely purged in accordance with best practices. No deletion or return-of-data commitment for customer material was located. No subprocessor list exists; the only third parties named anywhere are Stripe for payments and Meta among advertising networks, neither of which touches matter content. One cross-reference is broken and is recorded because a buyer following it will find nothing: clause 5 of the terms states that the privacy policy addresses the vendor's obligations in respect of security breaches, and the privacy policy, last updated 22 August 2026, contains no breach notification provision at all.

A generic website privacy policy covers the company without addressing what happens to documents and prompts after processing. The policy names contact information, IP address, browser and device data and usage data, and its retention clause says only that data is kept as long as necessary for the purposes set out; none of it reaches uploaded evidence, extracted facts or the assistant's outputs. No deletion commitment, no retention period for matter material, no subprocessor list and no breach notification practice was located. What is published and genuinely specific sits elsewhere: single sign-on, regular penetration testing, private cloud deployment, data stored in the UK, the US or elsewhere on request, and a responsible disclosure programme committing to respond to security reports within two business days with a safe harbour for good-faith researchers and credit on resolution. A Drata-hosted trust centre exists but returned a loader with no body on fetch.

The allocation of loss is published, readable before signing, and leaves almost nothing behind it. Clause 12(a) limits total liability, however arising and under any theory, to the resupply of the services, which is a re-performance remedy with no monetary ceiling behind it rather than a capped one, and clause 12(b) excludes direct, indirect, incidental, special, consequential and exemplary damages alike. Clause 11(c) provides everything as is and as available without warranty of any kind, with a long enumerated list of losses disclaimed including loss of data and unauthorised access to records. Clause 14 runs the indemnity one way from the user, covering their content and any use of the service. There is no vendor-side indemnity of any kind, no warranty on output, no service level commitment and no insurance position. One real floor exists and is recorded: clause 11(a) preserves the non-excludable guarantees of the Australian Consumer Law, which is a genuine statutory backstop for an Australian buyer and reaches a US litigation customer not at all. This is the middle band because the exposure is squarely addressed rather than left unstated.

Nothing is published on who bears the loss when the system is wrong. There is no terms of service, no master subscription agreement, no customer agreement and no order form template published anywhere on the site; the footer's only legal link is a website privacy policy, and the navigation carries no legal section at all. No indemnity, no liability cap, no warranty of any kind and no insurance position was located on any surface. A buyer cannot read the allocation of loss before signing because nothing exists to read, and this is a product whose output is described as testable at trial and was relied on at an arbitration hearing. Searched the home page, the platform navigation, the security page, the privacy policy and the site footer on 4 September 2026.

Integration is described by category throughout and no counterparty is named anywhere. The product material states that a team can upload or connect the documents it already has, shows a document system and a supported AI tool feeding the same factual record, and offers to let approved AI tools and drafting workflows build from that record, and the privacy policy confirms the mechanism, recording that data is collected from third-party applications connected to the platform, such as document management systems, when authorised by the user, and shared back with those integration partners. That is a real and reciprocal connection, and it is described only as a class. No document management system, review platform, matter management system or e-discovery tool is named on any surface read, no API or developer documentation was located, and nothing states what data moves, in which direction or on what trigger. For a product whose value proposition is being the single factual record several tools build from, the absence of a named counterparty is the gap a buyer would notice first.

No integration into practice systems was located. No document management system, e-discovery platform, review tool, transcript system or matter management product is named anywhere on the surfaces read, and there is no integrations page or developer documentation in the navigation or footer. The only connectivity evidence is single sign-on, listed among the security features, which is authentication rather than integration. The workflow described is upload a document set and interrogate it, with no statement about how documents arrive from an existing review platform. The platform page in the main navigation was not opened in this pass and is named here so the limit is visible rather than presented as settled; it is the cheapest available upgrade on this record. Checked 4 September 2026.

The isolation model is stated with unusual precision and the region is not stated at all. Customer data is described as isolated by matter, with separate access and data boundaries for each matter, and clause 19 of the terms describes the platform as operating on tenanted AI services under which confidential client information is not accessible by unauthorised personnel. Matter-level boundaries are a finer-grained statement than the tenant-level separation most records in this corpus offer, and they imply customer separation a fortiori, which is what lifts this off the band below. Against it, no region is named for storage or processing, no cloud provider is identified, and no residency commitment appears in the terms, the privacy policy or the security page, which is a live question for a product operating across Australian and United States entities with an application host on a country-specific subdomain and a trust centre on an Australian one. No on-premises or self-hosted option is offered or mentioned. This is the documented band gap where one limb is published cleanly and the other not at all.

Residency is offered as a real choice and the processing side is not addressed. The home page and the security page both state that data is stored in the UK, the US or elsewhere on request, which is a residency offering rather than a single fixed location, and private cloud deployment is listed as available for organisations that need it, which publishes a tenancy option beyond the default. That combination clears the band below comfortably. What is missing is the rest of the top band: nothing states where processing or model inference happens as distinct from where data is stored, no region list is enumerated beyond the two named plus on request, and nothing describes what changes between the standard and private cloud arrangements in terms of controls, isolation or price.

A real trust portal, a claim of recognised controls, and no standard named in text. The security page states that the security programme is built around internationally recognized controls and independently tested safeguards, and links to a trust centre hosted on a third-party compliance platform. Neither the standard nor the assessor is identified anywhere in the page text: no SOC 2 of either type, no ISO certification, no examination period, no scope statement, no certificate and no report summary appears, and annual penetration testing is asserted without naming the testing firm or publishing a scorecard. An image on the page appears to carry compliance marks and its contents could not be read, so nothing is inferred from it in either direction. The trust centre itself was not opened in this pass, so whether it fulfils on an email address or requires a sales conversation was not established; where the tier cannot be established the lower one is graded and the reason stated, which is what this note does. It is the cheapest available upgrade on this record.

Certification is real and stated across two surfaces, short of accessible evidence. ISO 27001 certification and SOC 2 Type II accreditation are both claimed on the home page and repeated on a dedicated security page alongside single sign-on and regular penetration testing. A trust centre operates at trust.wexler.ai and is publicly linked from the security page, described as highlighting the security systems with support for all customers. Two things hold this at B. No certifying body or auditor is named, no certificate number, scope statement or observation period is published, and no report date appears, so on the third-party verifiability test a buyer cannot check either claim against a register without contacting Wexler. And the trust centre itself is Drata-hosted and returned only a loading frame with no body on fetch, so its contents and its access tier could not be established; under the gated-is-not-absent rule the lower tier is assumed for that reason and stated here. The responsible disclosure programme, with a two-business-day response commitment and researcher safe harbour, is a real published security practice and is credited in the note rather than as certification.

The vendor describes its AI architecture and identifies nothing underneath it. The terms speak of a platform powered by industry leading AI processes and of tenanted AI services, and the product material makes an architectural claim in some detail, contrasting per-fact extraction linked to source with systems that store documents as embeddings in vector databases. No model is named, no version, no provider and no hosting arrangement, and no commitment to notify customers when any of that changes was located. The only third parties named anywhere on the estate are Stripe as payment processor and Meta among advertising networks, neither of which processes matter content. The question carries real weight here because the material is the entire factual record of contested litigation, including privileged work product, and because the tenanted AI services claim in clause 19 is precisely the kind of assurance a buyer would want to trace to a named provider and cannot. This is the middle band because a distinct architecture is described rather than merely labelled.

Nothing is published about the model supply chain a customer inherits. No model is named, no model provider is named, no subprocessor list exists on any surface, no location is given for inference as distinct from storage, and no commitment to notify customers of a change to any of it was located. The assistant is given a product name, Kim, which identifies the interface rather than what sits underneath it. The security page's mention of cloud security and the residency statement identify where data rests without saying whose models touch it. This is the single largest disclosure gap on the record and it sits against a customer base of international disputes practices whose clients routinely ask exactly this question. Searched the home page, the security page, the privacy policy and the site navigation on 4 September 2026.

No pricing information is published at any level, including the unit of charge. There is no pricing page in the navigation, which carries security, resources and company only, and every route on the site is a demo request. The terms describe payment mechanics without ever reaching a charging model: clause 7 establishes that access requires purchasing a subscription and paying a subscription fee for a subscription period, clause 8 provides that the fee may be paid by whatever methods are available and that the vendor can vary it at any time with the change taking effect at the end of the current subscription, and clause 9 offers refunds only under the Australian Consumer Law or at the manager's absolute discretion. Nothing states whether the subscription is charged per user, per matter, per document or per firm, no rate, band, tier or term length appears, and no implementation cost is described. A self-serve entry point exists and is advertised as free to start, with nothing published about what it includes or what the paid tier costs. No pricing row is owed on this record.

No pricing information is published at any level, including the unit of charge. There is no pricing page in the navigation or the footer, no tier names, no feature split, no indication of whether charging is per user, per matter, per page or per document, and no free trial or self-serve route. Every call to action on the site resolves to booking a demo or a contact form requesting name, email, company, title and team size. No VendorPricing row is written, because nothing lifted this axis off the floor. Checked home page, security page, privacy policy, site navigation and footer on 4 September 2026.

Four practice areas are described with substance rather than listed, and the boundary is left open. Commercial litigation is framed as reconstructing the sequence across contracts, emails, board materials, witness accounts and the production; family law as tracing finances, communications and parenting events across disclosures, bank statements, messages and filings; personal injury as reconstructing treatment, symptoms, diagnoses, work capacity and prior history across the medical file; and in-house disputes and investigations as maintaining one reviewable record across investigations, business teams, outside counsel and changing tools. Each names the material that practice actually generates, which is a more useful coverage statement than a list of sectors. Roles are addressed too, with partners, associates and litigation support named as working from the same record. What is absent is the limit: no jurisdiction is stated anywhere on the estate, no firm size band is given, nothing says which matter types or document sets the product handles poorly, and nothing addresses the languages it supports.

Segment and practice coverage are described with real substance and the boundaries are left open. The buyer is identified consistently and narrowly as disputes and investigations teams, with the product built for complex disputes, and the named customers place that at the top of the market: Clifford Chance, Goodwin, HSF Kramer and Burges Salmon, plus case studies at an AmLaw 100 firm and a global firm. Practice coverage is evidenced rather than claimed, spanning early case assessment, international arbitration and cyber work through the roles quoted. What is missing is the edge: no firm size floor or ceiling is stated, in-house and government use are not addressed, no jurisdiction of substantive coverage is named despite offices in London and New York, and nothing states which matter types or document types the product does not handle.