Alexi

AI Centrality

How much of the product is actually AI. Whether the machine learning is the mechanism the buyer is paying for or a feature layered onto conventional software, and whether the vendor is specific about which is which.

The artificial intelligence is the product. The vendor describes itself as a private single tenant AI engine, built originally to generate evidence backed research memos, with a proprietary model layer and a retrieval first architecture. Remove the models and there is no product: memo generation, precedent finding, chronology building and workflow automation are all model driven. Return to A after three consecutive B grades on this axis in the contract cohort, which is the axis behaving as designed rather than drifting.

Citation Accuracy and Hallucination Disclosure

Whether the vendor publishes measured accuracy on citations and assertions, grounds output to primary sources, and says plainly what its system does when it does not know. Legal has a documented public record of fabricated citations reaching filed briefs, so an untested claim of accuracy is not evidence.

Independently measured accuracy with a published methodology, the first such record on the index. Alexi submitted to the 2025 Vals Legal AI Report legal research study, run by an outside evaluator that published its rubric and weights in advance: accuracy at 50 percent, authoritativeness at 40 percent, appropriateness at 10 percent, across 200 United States legal research questions sourced from attorneys at named firms. Per the evaluator's own published results, Alexi scored a 77 percent weighted aggregate, placing second of the four systems measured behind Counsel Stack at 78 percent, and ahead of Midpage at 76 percent and a generalist model at 74 percent, against a lawyer baseline of 69 percent aggregate. Architecture is described rather than asserted, as retrieval first specifically to ground output in primary law rather than model memory. Two qualifications recorded, neither of which removes the A, because the underlying figures are independently produced and checkable by an outsider. First, the vendor's own published account of these results is selective: it leads with 80 percent overall accuracy against a 71 percent lawyer baseline and 79 percent single jurisdiction accuracy described as tied for the highest, which are favourable component scores, and does not state that it placed second on the evaluator's weighted aggregate. A reader relying on the vendor's page would form a different impression from one reading the evaluator's. Second, the evaluator and trade press both record that several major vendors declined to participate and at least one participated but withheld permission to publish, so the comparative field is self selected. Also recorded from the evaluator: all systems measured, this one included, struggled on multi jurisdictional questions and underperformed a generalist model on fifty state statutory surveys.

Autonomy and Oversight Model

What the system decides on its own, what a lawyer must approve, and whether the vendor documents where the review point sits. A tool that drafts under review and a tool that files without one are different products and different risks.

Autonomy is claimed and oversight is asserted without a described mechanism. The product runs multi step workflows that carry legal principles and citations through successive stages, and the vendor states that consistent verifiable accuracy is what makes that automation safe, which is an argument for trusting the system rather than a control structure around it. The CEO has spoken publicly about human oversight in the context of a long term vision for AI assisted arbitration. Searched the vendor site, the FAQ, the workflow and product pages and the blog on 29 Aug 2026 and located no description of what a workflow does unaided, no threshold at which it stops or escalates, no review surface a lawyer is given, and no statement of what the vendor commits to when an output is wrong. For a product whose stated advantage is that accuracy carries through each stage of an automated chain, the absence of a published stopping rule is the notable gap.

Operational and Outcome Evidence

Named, dated evidence that the product works in production at real firms or legal departments. Case studies with figures and identified customers count. Unattributed testimonials and launch announcements do not.

Customer claims stand in for evidence. The vendor states it is trusted by leading law firms and describes buyer roles precisely as CIOs and CTOs, knowledge management leaders, innovation teams, practice group leaders and managing partners, which is useful for a reader but is a market description rather than a deployment record. Searched the vendor site, the blog, the FAQ and the news announcements on 29 Aug 2026 and located no named customer, no case study pairing an organisation with figures and a date, and no method a reader could assess. The strongest quantified material on this property is benchmark performance rather than production outcome, and the two are different things: the Vals figures show how the system performs on a test set, not what changed at a firm that deployed it.

Privilege and Confidentiality Posture

How client confidences are handled: attorney client privilege and work product treatment, segregation of one client matter from another, whether client data trains any model, and what the vendor commits to in writing rather than in marketing.

CORRECTED 29 Aug 2026 during the trust portal sweep. Previously graded B because no training statement had been located, which was the limb holding it down. The statement exists and was missed: the vendor's FAQ states that nothing leaves the firm's private environment and that a firm's prompts, documents and workflow data never train global models and never mix with other firms, and a document management partner's listing independently records no training on customer data. That answers both the training question and the pooling question in the negative and in plain terms. Segregation was already the strongest element and is now better evidenced: every firm on the private cloud model receives a fully isolated single tenant instance, and the vendor enumerates what is isolated rather than leaving it general, covering retrieval, workflows, memory, security protocols and governance. Memory being named in that list matters, because an AI platform that learns from firm usage could otherwise carry one firm's accumulated intelligence toward another. Supporting controls: AES-256 encryption, TLS 1.2 or higher, single sign on, role based access control and audit logs, under SOC 2 certification. Two limbs are therefore fully answered and the third is not: no retention period or deletion commitment was located, which is recorded on the retention signal row. Graded A on the same basis as Noxtua, where training and segregation were both strong and no retention period was published.

UPL and Professional Responsibility Posture

Whether the vendor is clear that it supplies a tool rather than legal advice, who its audience is, and how it addresses unauthorized practice of law, competence and supervision duties, and jurisdiction limits. ABA Formal Opinion 512 is the reference point.

The audience is unambiguously lawyers and law firms, with named buyer roles inside firms and no consumer or non lawyer surface located, which is cleaner than most of this index. The vendor engages publicly with the hallucination problem in filings and positions accuracy as a professional risk question. Searched the vendor site, the FAQ, the blog and the product pages on 29 Aug 2026 and located no published position on the advice line, no treatment of competence or supervision duties, and no statement of jurisdiction limits, despite the product spanning Canadian and US law where the applicable professional rules differ. Recorded at C because the position is inferable from who the product is sold to rather than published.

AI Governance and Bias Disclosure

Published governance over model behaviour: who owns it inside the vendor, what is tested before release, and what is disclosed about disparate output across matter types, parties, or populations.

Principles and evaluation results are published without a governance mechanism. The vendor's substantive contribution here is real and rare: it submitted to an independent third party evaluation and published the results, and it argues publicly that transparency and accountability in legal AI require exactly that rather than marketing claims. Submitting to outside measurement is a governance act. But it is a point in time evaluation rather than a management system. Searched the vendor site, the FAQ, the blog and the news announcements on 29 Aug 2026 and located no AI governance framework, no AI management certification such as ISO 42001, no named owner of model governance, no pre release testing gate, and nothing on uneven output across matter types, parties or populations.

AI Safety and Data Stewardship

Retention, deletion, access control, and what happens to prompts and documents after they are processed. Whether the vendor states its subprocessors and its incident practice, or leaves the buyer to assume.

CORRECTED 29 Aug 2026 during the trust portal sweep. Previously graded C because the architecture was doing the work a published policy would normally do, with no encryption specifics, access control detail or attestation located. The vendor's security page and FAQ, not reached in the original pass, supply that detail. Published: AES-256 encryption at rest, TLS 1.2 or higher in transit, single sign on, role based access control, audit logs, support for private VPC deployment, and region specific deployments for firms operating across offices. Architectural isolation was already recorded and stands: a fully isolated single tenant instance per firm covering retrieval, workflows, memory, security protocols and governance. SOC 2 certification underpins the control set, and a Trust Center provides a request route to a security addendum, architecture whitepaper, data processing agreement and audit reports. Not located as of 29 Aug 2026, and why this is B rather than A: a stated retention period or deletion control for research queries, uploaded case files and generated memos, a named subprocessor list, and an incident or breach notification practice.

AI Liability and Recourse

What the vendor stands behind contractually when its output is wrong. Indemnities, caps, carve outs, insurance, and whether any of it is published or only reachable through a negotiated agreement.

Searched the vendor site navigation, the FAQ, the product and workflow pages, the blog and the news announcements on 29 Aug 2026. No published indemnity, liability cap, carve out, warranty on output or insurance position was located, and no customer terms of service or master agreement was located as published on the property. Recorded as a pure absence on the surfaces reached. Worth noting the shape of the gap: this vendor stakes its positioning on measured accuracy and on reducing the risk of hallucinated citations reaching a filing, and publishes nothing about who bears the loss if one does.

Practice Systems Integration Depth

How deeply the product reaches into the systems legal work already lives in: document management such as iManage and NetDocuments, Word and Outlook, contract lifecycle management, matter management, e-billing, and court filing systems.

One integration is named and none is documented. A product integration with iManage is announced, which is the legal specific document management connector that matters most for a firm facing research product and which several better resourced vendors on this index lack. Searched the vendor site, the FAQ, the product pages and the news announcements on 29 Aug 2026 and located no integrations index, no other named connector for practice management, court filing or Microsoft Word, and no documentation describing what the iManage integration moves, in which direction, or what an administrator configures. Recorded at C because the connection is announced rather than documented.

Deployment Model and Data Residency

Where the software runs and where the data sits. Multi tenant cloud, single tenant, private deployment, on premises, and whether region of residence is a published option or an enterprise conversation.

Deployment model is stated clearly with partial residency detail, which is the B band. Two models are published and the difference between them is meaningful: a private single tenant environment per firm as the default, and private cloud deployment for firms wanting complete ownership and control of their own data. Single tenancy by design is stronger than most of this index offers and it is stated as architecture rather than as an enterprise upsell. What is missing is geography: searched the vendor site, the FAQ and the product pages on 29 Aug 2026 and located no named regions, no customer selectable residency, and no statement of where processing happens as distinct from where data is stored, which matters for a vendor operating across both Canadian and US jurisdictions.

Security Certifications and Trust Center

Independent attestation a buyer can pull without a sales call: SOC 2, ISO 27001, penetration test summaries, a trust center with current reports and named scope rather than a badge image.

CORRECTED 29 Aug 2026 during the trust portal sweep. Previously graded D on the finding that no independent security attestation of any kind was located and no trust portal existed. That was wrong on both counts. The vendor publishes a dedicated security page stating SOC 2 certification, and operates a Trust Center reached through a request access flow, which it states carries a security addendum, an architecture whitepaper, a data processing agreement and audit reports. That is a self serve request route to substantive diligence material rather than a sales gate, and the document set named is broader than most on this index. SOC 2 certification is separately corroborated on the vendor's own platform pages, its about page, its FAQ and by a document management partner's technology listing. Supporting controls published on the FAQ: AES-256 encryption, TLS 1.2 or higher, single sign on, role based access control, audit logs and support for private VPC deployment. Short of an A because no coverage period, audit scope, report date, SOC 2 type designation or named auditing firm was located, and because the certification is asserted on marketing pages rather than evidenced by a certificate reached in this pass.

Model Supply Chain Disclosure

Which models sit underneath, whose they are, where they run, and whether the vendor commits to telling customers when that changes. A legal buyer inherits every dependency it cannot see.

The vendor refers to proprietary models without identifying what sits underneath them. Published: a proprietary model layer the vendor brands as Alex, described as trained specifically on law rather than general purpose, and a retrieval first architecture. That tells a buyer the shape of the stack and that a specialist layer exists. Searched the vendor site, the FAQ, the product pages and the blog on 29 Aug 2026 and located no named foundation model or provider underneath the proprietary layer, no statement of where models run, no subprocessor list, and no commitment to notify customers when the supply chain changes. A private single tenant environment narrows the exposure question without answering it.

Commercial Transparency

Whether a buyer can learn what this costs without entering a sales process: published rates, the unit being charged, what sits behind an enterprise tier, and what implementation adds.

Checked the vendor site navigation, the FAQ, the product and workflow pages and the blog on 29 Aug 2026. No pricing page was located, no rate is published, no unit of charge is stated and no tier structure appears. Every commercial path located terminates in a demo or contact request. No free trial or self serve entry point was located, and no third party pricing figure was located either.

Firm and Practice Coverage

Who the product is actually built for. AmLaw, midlaw, small firm and solo, in house departments, government and courts, and which practice areas are supported rather than merely claimed.

Segment and practice coverage is described with substance, short of the boundaries. The buyer is law firms specifically, and unusually the vendor names the roles inside them: CIOs and CTOs, knowledge management leaders, innovation teams, practice group leaders and managing partners. Practice coverage spans litigation, transactional, operational and administrative work, with a published practice area treatment for personal injury covering case file summarisation, chronology building and memo drafting. Jurisdictional scope is Canada and the United States. Not located as of 29 Aug 2026: any statement of which firm sizes or practice areas the product is not built for, and any enumeration of jurisdictional coverage at state or provincial level, which for a research product is the boundary a buyer needs most.