Luminance vs Robin AI: how they compare in 2026

The artificial intelligence is the product, and the vendor owns much of the stack. Founded by mathematicians in 2015, it runs a multi model architecture it calls a Panel of Judges, combining foundation, fine tuned and proprietary models. It ships its own legal intelligence model, Luna Crescent, trained in house and deployed in its own environment. Remove the models and there is no product.

The models are the product. Review, redlining, the Word add in's Ask, Draft, Edit and Research modes, clause level risk analysis and repository search are all generative or machine learning capabilities. The vendor states a training base of more than 4.5 million legal documents and 100 million legal clauses. One qualification applies: the company also sells Managed Services combining the software with human legal review, so part of what it sells is people. The software product itself is model driven throughout.

Substantive disclosure with a real gap. The vendor publishes ContractIQ Bench, a proprietary benchmark of 189,000 manually annotated and reviewed data points assessing interpretation of named provision types, including liability caps, termination for convenience and confidentiality obligations. It is tested on held out documents and concepts excluded from training, with blind evaluations by legal experts alongside. It publishes a result, 5 percent higher accuracy than leading general purpose models on contract understanding, and a speed figure of 200 to 400 tokens per second. It also states that the model is trained to prioritize faithful extraction and to identify absence rather than invent an answer. Two gaps remain. The published figure is a relative difference with no absolute accuracy rate and no named comparator models, so a buyer cannot tell what the 5 percent is measured against. And the benchmark is proprietary, with no sample tasks or rubric published, so an outsider cannot inspect or run it again.

Accuracy is asserted without measurement. Grounding exists in outline. The Word add in's Research mode is described as enhancing answers with authoritative, current information from a curated set of trusted legal sources on the web, and review output is clause level with risk flags tied to company standards. No accuracy figure, hallucination rate, test set or evaluation method is published, and nothing describes how retrieval reaches those trusted sources or how a reader opens and verifies one. The headline claim across vendor material is a speed figure, 80 to 85 percent faster review, rather than an accuracy figure.

A real published commitment with a described control mechanism, short of the full structure. The Panel of Judges architecture is itself an oversight design and is documented: multiple models analyze each clause independently and reach consensus, which the vendor states reduces hallucination risk. Outputs are described as traceable, and Traffic Light Analysis ranks deviation risk visually so a reviewer sees where to look. The vendor publishes a position piece arguing that human in the loop alone is insufficient and that systems must be designed for accuracy and transparency. Not located: where the review point sits when the product negotiates with a counterparty directly, the threshold at which it escalates to a lawyer, and what the vendor commits to when an output is wrong. The first gap matters because the product sends agreements to counterparties and negotiates on the customer's behalf.

A real published commitment has the models working alongside a supervising lawyer, with documented review surfaces and no full control structure. The Word add in surfaces clause level recommendations as tracked changes, comments or highlighted text based on company standards, which a lawyer accepts or rejects in place. The four modes separate asking from drafting from editing, so the user chooses the level of intervention. The Managed Services line puts the company's own legal professionals in the loop as a paid option, a literal form of oversight. Not located: any threshold at which the system defers, what it does on its own, and what the vendor commits to when an output is wrong.

Real deployment evidence with substance, short of attribution and method. Named customers appear in vendor and trade material including Hitachi, AMD, BBC Studios, Yokogawa and Koch, alongside a stated base of more than 700 organizations across 70 plus countries and all four of the Big Four consultancies. A customers page is published. The recurring figure, negotiation time reduced by up to 90 percent, is a vendor claim carrying a hedge and is not tied to any named customer, dated, or accompanied by a method. Searched the site, the customers page, the press releases and the resources index and located no case study pairing a named organization with figures and a date.

Customer logos and press mentions stand in for evidence, with results quoted from an internal source. A customers page exists and press logos including the Financial Times, Bloomberg, Forbes and CNBC are displayed prominently, alongside investor logos including Google, Temasek and PayPal. The most specific published figure, review completed over 85 percent faster, is attributed to the vendor's own internal legal team testing its own add in rather than to a customer. Searched the site, the customers page, the news and blog indexes and the guides and reports index and located no named customer deployment carrying figures, a date and an assessable method. Press coverage and investor backing are not deployment evidence.

Substantive published commitments, short of the full picture. Segregation is the strongest element and is documented precisely: each customer receives a dedicated single tenant instance with complete isolation and no mixing of data. Access control is documented in depth, including that vendor staff cannot view customer documents without explicit authorization given through the user interface, with all access tracked and audited. Encryption is specified to the key management service, cipher and rotation practice. Two gaps remain. No prohibition on training with customer content is published anywhere in vendor material, which stands out for a vendor that trains its own models and publishes a figure of 220 million legal documents. Attorney client privilege and work product handling is not addressed directly.

Confidentiality is asserted in general terms, and the one commitment that matters carries a qualifier that changes it. The security page states that customer data will not be used for model training, fine tuning or other feature development without express consent. That is permission gated by consent rather than a prohibition. Real controls are published: AES-256 at rest, current TLS in transit, and a statement that data never leaves the vendor's AWS environment. Not located: handling of attorney client privilege or work product, any segregation model between customers, users or matters, and any retention or deletion terms. For a product whose core use is taking in counterparty contracts at volume, the absence of a segregation statement is material.

A boilerplate structure sits in the terms while the product is sold well beyond lawyers. Dedicated solution pages target compliance, executive, sales, procurement, finance, human resources and marketing teams alongside legal, and the product negotiates contracts on a customer's behalf. Searched the site, the solution pages, the published terms and conditions, the privacy policy and the resources index on 29 Aug 2026. No position on advice versus tooling, no treatment of competence or supervision duties, and no statement of jurisdiction limits is published, despite operation in more than 70 countries. The tool is marketed to marketing and HR departments.

The audience is lawyers and legal teams, and the terms carry a standard structure, but no position on the advice line is published. The site, the platform and services pages, the published terms and the privacy policy, searched, showed no statement on advice versus tooling, no treatment of competence or supervision duties, and no jurisdiction limits. That is despite the company operating across the UK, US and Singapore and marketing to multiple jurisdictions. The Managed Services line, in which the vendor's own legal professionals review work for customers, raises the advice question more sharply than a pure software product would, and nothing published says how that service is scoped or supervised.

Principles and architecture are published without a governance mechanism a buyer could audit. What exists is real and substantial: a published white paper on how the AI is built, a named Director of AI who writes technical material under his own name, and a research team based in Cambridge. There is also a described validation regime through ContractIQ Bench, and a security advisory board of named external experts. That board governs security rather than model behavior, and no equivalent exists for AI governance. Not located: a named owner of model governance, a testing gate before release as distinct from benchmark results, an AI management certification such as ISO 42001, or anything on uneven output across matter types, parties or populations.

No governance position for model behavior is published: no named internal owner, no pre release testing regime, no published responsible AI or AI governance framework, no AI specific certification such as ISO 42001, and nothing on uneven output across matter types, parties or populations. The vendor publishes real security governance, which is a different subject.

Substantive published policy covers most of the ground in specific detail. The security FAQ describes AWS Key Management Service encryption at S3 and EC2 level with AES-256 keys rotated regularly, TLS 1.2 or higher in transit, and dedicated single tenant instances. It describes role based and division level permissions configured by the customer under least privilege, configurable password and session timeout policy, mandatory staff security training, threat detection through Darktrace's Enterprise Immune System and Juniper firewalls, and an incident management process covering detection, mitigation and communication. Backups are stated precisely: nightly to a secondary AWS data center in the same region, encrypted, and kept at least 14 days. Not located: a retention period or deletion control for customer documents in normal operation as distinct from backups, and a named subprocessor list.

A generic security posture covers the product without addressing what happens to documents and prompts after processing. Published and real: AES-256 at rest, current TLS in transit, security built into the development lifecycle, data replication across geographic locations with disaster recovery exercises and automated failover, and named monitoring tools in AWS CloudWatch, CloudTrail and Datadog. That is operational security described in useful detail. Not located: any retention period for documents or prompts, any deletion control, a named subprocessor list, and an incident or breach notification practice. Anthropic and AWS are named as partners on the security page, which partly discloses the supply chain but is not a subprocessor list.

Liability is addressed only through published terms carrying a standard structure. Terms and conditions are published openly alongside a privacy policy, cookie policy and anti slavery statement, so a buyer can read the allocation of loss before entering a sales process, which keeps this above a pure absence. Searched those documents, the security page and the security standards white paper entry point and located no indemnity running to the customer for third party claims arising from output, no warranty on output, no stated liability cap figure and no insurance position. Worth noting the product negotiates with counterparties on a customer's behalf, which raises the recourse question more sharply than a review only tool.

Liability is addressed only through published terms carrying a standard limitation structure. Terms and a privacy policy are published openly, so a buyer can read the allocation of loss before entering a sales process, which keeps this above a pure absence. Searched those documents, the security page and the trust center entry point and located no indemnity running to the customer for third party claims arising from output, no warranty on output, no stated liability cap figure and no insurance position. Worth noting that the Managed Services line involves the vendor performing legal review work, which raises a distinct recourse question that published material does not address.

Integrations are asserted at platform level with no documentation an implementer could use. The vendor describes connecting intake, negotiation, workflow and repository intelligence in one platform, and a Collaborate product exists for working with counterparties, and third party sources reference Microsoft Word and Outlook working surfaces. What is not published on the vendor's own property, after checking the platform pages, the technology page, the security page and the resources index, is any integrations page, any named connector for document management, contract lifecycle, e signature, CRM or ERP, and any description of what an integration moves or what an administrator configures. For an enterprise platform sold to procurement and finance functions, that absence is notable.

Integrations are claimed without documentation an implementer could use. The Microsoft Word add in is real, documented at user level in Robin University, and is the primary surface, alongside a browser application with a searchable executed contract repository. Beyond Word, third party sources describe CRM and CLM integration but no vendor integrations page is published, and no named connector for document management, contract lifecycle or e signature is published on the property. A third party review notes limited file format support beyond Word documents. Nothing published describes what any integration moves, in which direction, or what an administrator configures.

The deployment model is stated clearly, with partial residency detail. Three things are published and specific: a dedicated single tenant AWS instance per customer with complete isolation, deployment within the customer's own environment as an alternative to the hosted option, and backup to a secondary AWS data center within the same region, which confirms data stays in region. The vendor states AWS global infrastructure provides a solution tailored to geographic requirements. What is missing is the list: no named available regions, no statement of which regions a customer may select, and no statement of where processing happens as distinct from where data is stored.

Cloud delivery is stated, and neither the tenancy model nor the region is. The vendor says data never leaves its AWS environment and is replicated across multiple geographic locations for disaster recovery. That is a resilience statement rather than a residency one, and it arguably cuts against a customer wanting data confined to one jurisdiction. Not located: any named region, any residency the customer can choose, any tenancy model, and any statement of where processing happens as distinct from storage. Other sources mention private cloud deployment options, which is not vendor material. For a UK founded vendor selling into the UK, US and Singapore, published residency options would be expected.

Certification is real and stated with correct nouns, short of accessible evidence. ISO 27001:2022 is named with its version and described as certification. The SOC 2 Type 2 language is precise: the vendor says it successfully completed a SOC 2 Type 2 examination assessing controls related to security, availability and confidentiality, which names the trust services criteria in scope. Regular independent third party penetration testing is stated. A named external security advisory board, including a former Director General of MI5 and two former Darktrace executives, is published with full biographies. What is missing is the evidence route: no trust portal is published, no report is downloadable or requestable through a published flow, and no coverage period, report date or auditor name is published as of 29 Aug 2026. A security standards white paper is published, the nearest thing to an evidence route.

Certification is real and stated with an open access route, without scope or evidence. ISO 27001 and SOC 2 are both named as certified and GDPR compliance is stated. A trust center is published at a stable URL as a self serve route rather than a sales gate. Not located: any coverage period, audit scope, report date, SOC 2 type or named auditing firm. The security page writes ISO without a number in one place and ISO27001 in another. The certification is stated clearly enough elsewhere that this reads as loose copy rather than a problem with the claim.

The supply chain is partly disclosed and the architecture is described in depth. The vendor publishes that it runs a multi model Panel of Judges combining foundation, fine tuned and proprietary models, that it continuously evaluates and selects the best model for each task, and that its own model Luna Crescent is deployed inside its own AWS environment. It states the consequence in the customer's terms: owning the model reduces reliance on external model providers, limits data exposure to additional subprocessors, and protects customers from third party availability, pricing and access disruption. Hosting is named as AWS. Not published is which foundation models sit in the panel, from which providers, or which tasks route to them, and no subprocessor list or change notification commitment is published. A buyer knows the shape of the chain and its own model, but not the third party links in it.

The vendor names a partner without disclosing the supply chain. Anthropic and AWS are both named as partners on the security page, and other coverage describes collaboration with Anthropic and AWS to build secure models, so a reader can infer the likely model provider. But naming a partnership is not the same as stating which model serves which task. Not located: which models are used in production, where they run, whether any other provider is involved, a subprocessor list, or any commitment to notify customers when the supply chain changes. The partner logo invites an inference that the published material does not confirm.

No pricing page exists on the property, no rate is published, no unit of charge is stated and no tier structure appears. Every commercial path terminates in a demo request. No free trial or self serve entry point is published, and no third party pricing figure is published either.

No pricing page exists, no rate is published, no unit of charge is stated and no tier structure appears. Every commercial path ends in a demo request. Vendor material refers to a free account with add in access and paid accounts with playbook customization, and tells buyers to speak to a member of the team to learn what paid accounts include, so even the split of features between free and paid is not published. Other sources describe subscription pricing without figures.

Who the product serves is documented precisely across two dimensions, each with its own published pages. Six industries: manufacturing, financial services, pharmaceutical, technology, insurance and chemical. Eight business functions: legal, compliance, executive, sales, procurement, finance, human resources and marketing. Both law firms and corporate legal departments are addressed, with a stated base of more than 700 organizations across more than 70 countries including all four Big Four consultancies and named enterprises. An academic program is published as a separate segment. The practice boundary is clear from the structure and consistent throughout: this is contract work end to end, from generation through negotiation to post execution analysis and investigation, and nothing on the property claims litigation or research capability it does not have.

Coverage is claimed broadly without evidence that the breadth is real. Vendor material addresses in house legal teams, law firms and corporate clients across multiple jurisdictions. The practice focus is clearly contract work: review, drafting, negotiation and tracking obligations after signature, with other sources naming NDAs, service agreements, procurement and employment agreements. Offices are published in London, New York and Singapore. No segmentation a buyer could use is published. There are no pages by firm size, industry or role, no statement of which practice areas or team sizes the product is built for, and no statement of what it is not built for. The customers page exists but does not resolve the segments claimed.