R

Referent

The agents are what Referent sells, running on top of a CRM that would still hold records without them. The company calls the product AI native practice management and describes it as a system of action rather than a system of record. Agents build the workspace from the firm's inbox, file every email and attachment to the right matter, draft replies, letters and documents from the matter's context, watch court dates and limitation periods, capture time, prepare bills, run follow ups, and screen new clients at an intake link.

A Model Context Protocol server extends the same work into ChatGPT, Claude and other assistants. Underneath sits a legal CRM for leads, clients, matters, tasks and documents, which a firm could still use by hand if the agents were switched off.

For a practice management agent, accuracy means whether a drafted reply, a filed email, an extracted deadline or a prepared bill is right, since each reaches a client or a court. Referent claims grounding and measures nothing. Its pages say drafts and answers are prepared from the matter's full context and that answers through its protocol server are grounded in the firm's practice. The architecture it describes, with an isolated workspace per firm and agents reading the matter file, is plausible grounding.

But no accuracy figure, test, error rate or evaluation is published for any agent, nothing describes how a draft cites back to the email or document it came from, and nothing says what happens when a deadline is extracted wrongly. Section 6.2 of the Terms says Referent does not review output for accuracy, completeness or currency and that output may be incomplete, inaccurate or unsuitable. Section 6.9 warns that transcripts may misattribute speakers or omit passages. The product is not presented as a case law research tool.

What runs alone, where it must stop and how the lawyer checks it are all stated. Agents do the inward work on their own: filing email and attachments to matters, tracking deadlines, capturing time and preparing drafts. Everything that faces a client or carries risk waits in an approval queue. The company states the limit categorically: nothing is sent, filed or billed without the lawyer's sign off, and drafts sit in the queue to be approved, edited or rejected.

Every agent action goes to an audit trail recording what was done, when and on whose approval, and the protocol server is described as working under the same permissions, approvals and trail. The Terms carry this into the agreement. Section 6.1 says the AI features assist users and make no decision with legal or similarly significant effect on a solely automated basis, and section 6.4 makes human review of all output mandatory.

The legal harness page ties the audit trail to the supervision a lawyer owes a nonlawyer assistant under Model Rule 5.3.

No named firm, customer quote, case study, usage figure or dated result is published. The company's own comparison page describes Referent's maturity in 2026 as invitation only with hands on onboarding, set against a competitor's 150,000 professionals and public reviews, which candidly concedes there is no track record yet. The about page points to the team's earlier consumer product, AI Lawyer, as one of the most widely used consumer legal AI products.

That is a different product sold to a different buyer and says nothing about Referent running a firm's practice. The site offers a free plan and open registration, so production use may exist; it is simply not published.

Written commitments cover training, model providers and deletion, privilege is acknowledged, and walls inside a firm are not addressed. The Terms say Referent does not use customer content to train generalized language models and contractually bars its AI providers from using it to train or improve their general purpose models. Section 14.3 of the Data Processing Agreement repeats both, and Referent also excludes model providers established in China.

The security page says every model provider works under zero retention and no training agreements, and that each firm has its own encrypted workspace never mixed with another firm's. Section 14.4 of the Terms recognizes that content may be subject to legal professional privilege or professional secrecy and says it is treated accordingly. Section 13 of the Data Processing Agreement deletes or returns data on termination, with backups overwritten within 60 days.

Access inside a workspace is described only as role based for authorized firm members, with no walls or ethical screens at matter level, and the zero retention claim on the security page is broader than the provider terms the agreement sets out.

A clear line is drawn on advice and on the lawyer's review duty, and disclosure to the firm's clients is left to the firm. Section 6.3 of the Terms says Referent is a software provider, not a law firm, does not practice law or give legal advice, and creates no lawyer client relationship with the firm or its clients. Section 6.4 requires the firm to review all output for accuracy and for compliance with professional conduct and fiduciary duties before relying on or disclosing it, and section 4.4 leaves confidentiality and privilege obligations with the firm.

The legal harness page frames the approval queue and audit trail as the record Model Rule 5.3 asks a lawyer to keep over nonlawyer assistance. The one public surface is the intake link, where a prospective client talks to an AI. Section 6.6 makes the firm responsible for telling people AI is in use, including under Article 50 of the EU AI Act, with Referent supplying disclosure text. Whether the intake link identifies itself as AI by default is not stated, and apart from recording consent, which section 6.9 says varies by US state, no jurisdiction limits are named.

Principles are published, and the machinery behind them is not. The about page sets out three commitments: AI multiplies lawyers rather than replacing them and the company refuses to sell replacement, the lawyer approves every critical step, and confidentiality is a design constraint rather than a feature. The Terms reinforce the second with a clause barring solely automated decisions with legal effect, and the company says it is building under EU law, where the AI Act's transparency duties apply.

What is missing is any account of how agent behavior is governed inside the company. No one is named as accountable for what the agents draft and send for approval, and nothing describes what is evaluated before a new agent, model route or prompt reaches firms. No finding is published on how drafts or intake screening perform across matter types, clients or languages. Security certification work is described separately and does not answer these questions.

Every part of data stewardship is published in the agreement, with dates and numbers attached. The Data Processing Agreement, effective 5 August 2026, names ten subprocessors in Annex 3 with each one's function, processing location and transfer safeguard, from Google Cloud hosting and Cloudflare to the AI providers. Section 7 commits to notify the firm in advance of any addition or replacement, with the name, location, function and certifications, and gives a 14 day right to object, shortened to 48 hours' notice where a subprocessor fails suddenly.

Section 11 commits to breach notification within 72 hours with the details a firm needs for its own reporting. Section 13 deletes or returns data on termination, overwrites encrypted backups within 60 days and confirms deletion in writing on request. The Privacy Policy sets retention by category, such as account data for the contract term plus six months, and the security page limits production access to a small number of engineers, logged and reviewed. One reservation is that the security page's summary does not always match the annex beneath it.

The agreement is direct about AI output, and what it says is that the firm carries the risk. Section 6.2 of the Terms provides the AI features and their output as is. Section 6.5 says Referent is not responsible for any decision, advice, filing, omission, deadline or outcome based on or influenced by output, or for any resulting loss, professional liability or regulatory consequence. Section 6.10 has the firm defend and indemnify Referent against claims arising from its use of the AI features, their use toward its clients and any failure to make required disclosures.

Section 12.3 caps each party's liability at fees paid in the twelve months before the claim, section 12.4 lifts the cap only for the firm's payment obligations and indemnities, and section 12.5 excludes loss the firm could have avoided with backups. Referent's own indemnity in section 13.1 covers intellectual property claims against the service and expressly excludes output. For a product that tracks court deadlines and drafts client correspondence, a missed deadline is the firm's alone.

The Google side of a small firm is connected in depth, and most of a legal stack is not connected yet. With Gmail access Referent files incoming client correspondence to the right matter and sends replies the lawyer has approved. Google Calendar syncs hearings, meetings and deadlines in both directions, and Google Drive is connected. Referent says it connects only to the Google services a firm chooses, and its Data Processing Agreement names Nango as the service that manages the access tokens.

Outlook and Microsoft 365 are marked as coming soon. A Model Context Protocol server lets a firm work with its matters from ChatGPT, Claude, Perplexity, Gemini and other assistants under the same permissions, approvals and audit trail, and the company says it migrates clients, matters, tasks and files from existing tools. No document management system, accounting or trust accounting package, electronic billing format or court filing service is connected, and its own comparison page concedes that Referent is not an accounting system.

Regions and processing locations are published in detail, and two of Referent's own documents disagree about where US data is stored. The security page says Referent runs on Google Cloud with regional separation, so US firms' data is stored in US data centers and EU firms' data in EU data centers, and each firm has its own encrypted, isolated workspace. The Terms say AI processing for customers in the EEA and the UK happens within the EEA.

The Privacy Policy notes that text to speech runs through xAI in the US. Annex 3 of the Data Processing Agreement gives a processing location for each subprocessor, with OpenAI running in an EU region for EU and UK customers and otherwise in the US. The same annex lists Google Cloud hosting only in europe-west1 in Belgium and names no US region. A US firm reading the agreement it signs would conclude its data is hosted in the EU, and one reading the security page would conclude the opposite. No single tenant or private deployment option is offered.

No attestation exists yet, and Referent says so plainly. The security page has a section on certifications, headed as where the company stands, honestly: SOC 2 is in preparation with the control set mapped, ISO 27001 is in preparation with requirements adopted before the certificate, penetration tests are conducted and findings fixed, and single sign on is on the roadmap. Its FAQ answers whether it is SOC 2 certified with "Not yet" and promises updates as certifications complete.

Section 12 of the Data Processing Agreement commits to provide third party audit reports and certifications where available and to answer security information requests within seven business days, and security questionnaires are answered by email. No tester or date is named for the penetration tests.

The providers, what each does and where it runs are named in the agreement, with a commitment to notice of change, and the published lists do not match. Annex 3 of the Data Processing Agreement names OpenAI for generation, transcription and embeddings, in an EU region for EU and UK customers and otherwise in the US. It also names Mistral AI for character recognition in France, xAI for text to speech in the US, Turbopuffer for the vector index and E2B for the isolated environments agents run in.

Section 7.3 commits Referent to notify customers in advance before any subprocessor is added or replaced, and the Terms exclude model providers based in China. The security page, however, says models come from providers including OpenAI and Anthropic, routed by task, and Anthropic appears nowhere in the annex that governs. No model family or version is named for any task.

The entry point is free and stated, and what a firm pays after that is not published. Referent says registration is open to solo lawyers and small firms with no credit card, that the product starts free, and that paid plans include AI usage rather than billing it as an add on. It makes that a point of contrast with competitors whose per user prices it quotes, such as a comparison page citing about $39 per user per month for Clio.

The Terms describe subscriptions that renew automatically, seats that can be added or removed, no refund for canceling during a term, and price changes on 30 days' notice before renewal, so the mechanics of a paid plan are published. The plan itself is not. No pricing page exists, since the address returns a not found page, and no plan name, figure, unit of charge or AI usage limit appears on any page read. A firm can start without talking to anyone and cannot learn what it will pay to stay.

The buyer is defined precisely and the practice is not. Registration is open to solo lawyers and small firms, which the company describes as roughly two to ten lawyers needing matters, clients, documents, deadlines and billing in one place, plus shared context and permissions. Boutique firms are addressed, and a tier for large firms is marked as coming. The company says it is based in Poland and building for US firms, while its data terms also serve EU and UK firms with EU residency.

It states some limits plainly: it is not an accounting system, Outlook and Microsoft 365 are not yet connected, and access in 2026 is described as invitation only with onboarding help. No practice area is stated. No page says which kinds of matters its intake screening, deadline tracking and drafting are built for, whether it handles litigation calendaring rules by jurisdiction, or where it should not be used. The only matter types visible are examples in a product screenshot.