Genie AI

AI Centrality

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

The models are the engine of the core capability, but they sit on a document product that would function without them. Drafting, playbook review, redlining, cross-document comparison and tabular extraction are all generative, and the Eidetic Intelligence architecture is what the vendor sells against general-purpose LLMs. Underneath that sits a conventional platform: a docx editor, PDF import, privacy-aware document sharing, department spaces and a library of more than 500 templates, all of which appear on the Free tier and all of which Genie shipped as an open-source template library in December 2018, before the assistant existed. Remove the models and a template repository with an editor remains, which is what holds this at B.

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.

Accuracy is measured and published, but not on the dimension this axis asks about. The benchmarking programme carries two dated head-to-head reports: 18 February 2026 against Claude Sonnet on a multi-jurisdiction regulatory scenario, and 22 April 2026 against Claude CoWork on a commercial supply agreement, scoring 15 and 10 dimensions respectively with per-metric rationale the vendor says a reader can reproduce or critique. Those are comparative scores on Genie's own rubric, graded by Genie, rather than a measured citation accuracy figure or a hallucination rate with the test set described. Three headline numbers used in marketing correspond to neither published report and were not located on the benchmarks index on 31 Aug 2026: 92 percent lawyer-rated accuracy, a 140 percent advantage over GPT and Claude, and 90 percent on a 65-document study against 79.3 percent for Claude CoWork and 37.3 percent for ChatGPT. The home page states that the architecture prevents hallucinations before they impact the final result, which is an unmeasured claim of exactly the kind this axis exists to test. One further discrepancy: the benchmarks page carries the meta description 'Independent head-to-head benchmarks' while the page body states that Genie's own engineering team publishes them.

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 prominently and the oversight mechanism is not described. Marketing sells agents that draft, review and negotiate contracts autonomously, invites the buyer to control when they need to and delegate when they do not, and describes agents handling complex multi-document cross-border deals. What is published about the review point sits in the agreement rather than the product: clause 3.4 disclaims any guarantee that the software comprehends what is put into it, clause 5.9 states that recommendations, suggestions or actions taken by the software are not legally binding, clause 6.10 prohibits use for automated decision-making with legal or similarly significant effects, and clause 21.7 instructs the user to have a human review each output as appropriate. Those are allocations of responsibility, not a described control structure, and no threshold, review surface or escalation path was located on the home page, the use-case pages or the model training page on 31 Aug 2026. The direction of the gap is worth noting: the marketing sells delegation while the agreement prohibits the automated decision-making that delegation implies.

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.

Named customers with named individuals and at least one figure, short of dates and method. The customer stories index carries attributed video case studies: Jude Legg, Compliance Lead at Firefish, on cutting contract review time by more than 50 percent; Jafeth Bulsink, Head of Sales at Bright Renewables, on 25 salespeople negotiating contracts with no legal department; Neil Rowe, Head of Commercial at Cambridge United, on player signings; and Nazmus Tareque of AnthroTek on closing a seed round without using a lawyer. A logo strip of 13 named organisations including hyperexponential, Blockworks and the Jamaica Stock Exchange sits alongside them, and the model training page states that Clifford Chance, Pinsent Masons and Withers ran pilots from 2018. No deployment date, scope or measurement method was located on the pages read on 31 Aug 2026, and the individual case study pages were not opened item by item, so the date and method limbs are rebuttable with a single link.

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.

The published commitment and the published agreement point in opposite directions on the question that decides this axis, and the agreement is what a customer is held to. The security centre states that customer documents, prompts and confidential data are never used to train shared or internal AI models, and the model training page is headlined on never training on your contracts. Clause 7.3 of the product terms grants Genie a non-exclusive, irrevocable, worldwide, perpetual licence to use, copy, display, distribute and reproduce a customer's Documents in order to improve the software and, in terms, its algorithms and models. Clause 6.8 adds that Genie may use Your Data and Documents to make Improvements and clause 7.5 vests ownership of those Improvements in Genie. The model training page narrows its own promise in its FAQ to contracts never being used to train models that other customers see, and states that the 200,000 businesses using Genie contribute new legal context to the corpus through normal use. Substance on the other side is why this is not lower: OpenAI and Anthropic are named and operate under zero data retention agreements, Genie is stated to be processor and the customer controller, a DPA and a subprocessor list are published, and clause 20.7 says directly that communications are not covered by legal professional privilege, which is more candid than most of this market. No tenant or matter-level segregation model is documented and no retention period is 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.

A published position on the advice line that is more specific than most of this market, held off an A by two limbs it does not reach. Clause 20 of the product terms states that Genie AI Limited is not a law firm, is not authorised or regulated by any body regulating legal services, and cannot provide reserved legal services including litigation, probate, notarisation and administering oaths, operating only in general commercial matters. It then works through the consequences for the buyer rather than leaving them implied: no requirement to hold professional indemnity insurance, no right to complain to a legal ombudsman, no access to a compensation fund such as the SRA's, and no legal professional privilege over what the user receives. Clause 6.7 excludes domestic, personal and non-commercial use, so there is no consumer-facing surface requiring disclosure. What is missing: nothing published addresses a lawyer's own competence and supervision duties, and the whole treatment is drawn from the England and Wales regulatory perimeter on a product sold across more than 150 jurisdictions including the United States, with no equivalent statement for any of them located on 31 Aug 2026.

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.

A described quality mechanism exists where a governance position should be. The model training page sets out a five-layer account in which qualified lawyers contribute training data, set prompt engineering standards, supervise retrieval logic and rate every release against a panel benchmark before it ships, and the security centre publishes a specific EU AI Act position: that Genie is not a high-risk system under the Annex III categories, that the obligations applying to it came into force on 2 August 2026, and that it maintains AI literacy measures for staff operating the system. The substance of this axis is absent. No individual or role inside the vendor is named as accountable for model behaviour, no pre-release testing regime is published beyond accuracy benchmarking, and nothing at all was located on uneven or disparate output across matter types, counterparties or populations. Searched the home page, the model training page, the benchmarks index, the security centre and the product terms on 31 Aug 2026. The ISO 27001 certification covers information security, which is a different subject and does not carry weight here.

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.

Substantive published policy across almost all of this ground, missing one specific that a buyer would need. The security centre publishes a register of 35 controls grouped as an auditor would walk them, including AES-256 at rest and TLS 1.2 or higher in transit, quarterly access reviews recorded in an Access Rights Review Log, a joiner-mover-leaver process, environment segregation, peer-reviewed code with deployment separated from development, tested encrypted backups, a business continuity plan carrying a one-hour recovery time objective for critical services, and documented security and privacy incident response with incidents tracked through an Incident and Corrective Action Log. Breach practice is stated rather than implied: notification to affected customers without undue delay and, where Genie is controller, a report to the UK Information Commissioner's Office within 72 hours. A current subprocessor list names 15 third parties with the purpose, processing location and security measures for each. The gap is retention: the register states only that formal procedures govern the retention and secure disposal of customer data, and no retention period for documents or prompts was located anywhere on 31 Aug 2026, so there is no window to hold the vendor to.

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.

An unusually complete published allocation of loss, short of anything covering the output itself. Clause 18.2 caps aggregate liability at the fees paid or payable in the 12 months before the first event giving rise to the claim, and clause 2.4 sets a separate cap of one thousand pounds for the Free Package. Clause 9 gives the customer a defence and indemnity against third-party claims that use of the software infringes intellectual property rights, with conditions in 9.2, a remedy ladder in 9.3 of procuring a licence, modifying the software or terminating, and six named carve-outs in 9.4 including Your Data and any use under a Free Package. What is absent is recourse when the output is wrong: clause 5.8 excludes liability for reliance on the software's outputs, recommendations or actions, clause 21.3 disclaims any warranty that content is accurate or complete, clause 21.4 excludes all liability for reliance, and clause 20.4 states that Genie is not required to hold professional indemnity insurance and makes no claim to hold any. This sits above a standard limitation clause because a buyer can read the whole allocation, including a real indemnity with a stated scope, before signing.

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.

Integration categories are named and no documentation an implementer could use was located. The home page lists MCP connectors, an API, and native integrations with email, CRM, cloud storage and eSign, and the Enterprise tier adds SSO authentication alongside API and MCP integration. A Word add-in has its own use-case page. Against that, every 'Learn more' link on the home page integration card resolves to an anchor rather than a destination, and no developer index, integration directory or help centre documentation was located from the site navigation or the footer on 31 Aug 2026. Nothing was found on the systems this axis asks about for legal work specifically: no document management integration such as iManage or NetDocuments, no matter management, no e-billing and no court filing. That is consistent with a product sold to commercial teams rather than to firms, but it leaves the depth unverifiable. The Word add-in page was not opened, so the depth limb is rebuttable with one link.

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.

Residency is published in more detail than most of this market and the tenancy model is not stated at all. The security centre states that the platform runs on Amazon Web Services with customer documents stored in UK and EU data centres, and the subprocessor table goes further by giving a processing location for each of 15 third parties, which lets a reader separate where data sits from where it is processed: AWS and MongoDB in UK and EU, CloudAMQP in the EU, and OpenAI and Anthropic in US and EU data centres. Two things hold this at B. Nothing published states whether the platform is multi-tenant or single-tenant, and no private, single-tenant or on-premises option and no customer-selectable region was located on 31 Aug 2026. There is also a tension a buyer should read alongside the storage claim: the same subprocessor table lists Google Cloud and Stripe at global data centres, which is broader than the UK and EU statement on the security page.

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.

Certification is real, current and documented to an unusual level of detail, with the report itself the one thing held back. The security centre names the standard as ISO/IEC 27001:2022, gives certificate number 310012019, names the issuing body as Citation ISO Certification, and publishes the original approval date of 7 February 2019, a current certificate dated 7 February 2026, expiry on 6 February 2029, a last external audit on 29 April 2026 and a last management review in June 2026, with the page marked last reviewed 30 July 2026. The vendor also states plainly that it does not hold a SOC 2 report, which is more useful to a buyer than a silent omission. The certificate, the ISMS scope and the Statement of Applicability are available on request under NDA through a named security mailbox with a stated two working day response, which is a promise of follow-up rather than a self-serve download, so the attestation earns no access credit and the ISMS scope, the part a buyer most needs, is what is withheld. The control register refers to independent penetration testing following significant changes but names no testing partner and publishes no summary.

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 providers are named and the change commitment runs the wrong way. OpenAI and Anthropic are both listed on the published subprocessor table with their purpose, processing locations in US and EU data centres and stated security measures, and the security centre adds that both operate under Zero Data Retention agreements so that inputs and outputs are not retained by the provider after processing. The architecture is described rather than gestured at: the model training page states that Genie passes legal context at inference time instead of fine-tuning, specifically so newer foundation models can be swapped in as they land. Two gaps, checked 31 Aug 2026. No specific model or version is identified for either provider, so a buyer knows whose models but not which. And on notification the subprocessor page states the opposite of a commitment, saying changes will be posted on the page and asking the reader to check back regularly, which puts the burden of noticing a supply chain change on the customer.

Commercial Transparency

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

Real pricing across the whole range including a floor for the enterprise tier, which is more than most of this index publishes, held off an A by implementation. The pricing block on the home page carries a figure, a unit and a term for all three tiers: Free at $0 for one user with limited tokens and no document exports, Pro from $75 per month for one user with enough usage for roughly five documents, and Enterprise from $600 per month for unlimited users and unlimited tokens. Each tier's contents are itemised rather than summarised, and signup is self-serve with no credit card required. Two things hold it at B. Every figure carries a 'From', so each is a floor rather than a rate. And nothing published states what onboarding, the dedicated customer success manager or the bespoke AI training programme bundled into Enterprise actually add. A buyer should also note a currency mismatch: prices display in US dollars while clause 13.3 of the product terms states that payment will be made in pounds sterling unless stated otherwise.

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.

Coverage is documented precisely on the dimensions the vendor sells, and the limits are stated in the agreement rather than left open, which is rare on this axis. Dedicated pages exist for six user groups including in-house lawyers, procurement, sales, finance and founders, for three company types across startup, mid-market and enterprise, and for six industries including energy, construction, technology, real estate, mining and sport. The Enterprise tier states coverage of more than 150 jurisdictions and ten languages. The boundary statement is unusually explicit: clause 20.2 of the product terms names what Genie cannot do, being reserved legal services including litigation, probate, notarisation and administering oaths, and confines the product to general commercial matters, while clause 6.7 excludes domestic, personal and non-commercial use. What is absent is the firm dimension this axis also asks about. No law firm segmentation was located, nothing addresses government or court use, and the practice dimension is expressed as industry verticals rather than areas of law, so a reader learns which sectors are served but not which law is covered.