C

CaseGen

Take the agents away and there is nothing left to buy. CaseGen sells three named voice agents, each doing a job a person used to do. Justina answers inbound calls, runs the intake questions, handles existing clients and third parties, screens out spam and alerts the firm to urgent or high value cases. Justin calls back leads who did not sign and finishes their intake, and Maya chases medical providers and clients on treatment for personal injury matters.

The dashboard, transcripts, summaries and integrations all exist to carry what the agents produce into the firm. The company positions itself against human answering services and call centers rather than intake software. Its engineering write up describes a speech to text layer, a language model layer and a speech output layer, each with its own fallback, which is the working core of the product. There is no conventional workflow product underneath that a firm could keep using if the models were switched off.

For an intake agent, accuracy means whether what the agent hears, writes down and repeats back is right: the caller's facts, the case type, the injury, the dates, and the summary a lawyer later relies on. CaseGen publishes real platform data, but not on that question. Its intake benchmark draws on 13,280 calls over six months of 2026 and reports that 97.6 percent of callers never asked for a person. It states its own limits plainly: no denominators per account, silent hangups invisible, and a handful of greeting configurations that could be confounded.

That measures caller behavior, not the accuracy of what was captured. The engineering article on fallback systems is the closest thing to a grounding method. A low confidence transcript switches to an alternate speech engine, and an unstructured, empty or hallucinated model response triggers a retry, a rephrase or a backup model. No transcription or extraction accuracy figure is published, and section 15.2 of the Subscription Agreement warns that AI output may contain inaccuracies or hallucinations and should not be relied on without human review. The product does not cite legal authority.

The agents run whole calls unattended, and the firm has real ways to watch and step in, but no published rule for when an agent must stop. On autonomy, Justina answers, routes and responds "on autopilot" and collects complete intake with "no staff required". Justin places outbound calls to finish intakes, and Maya reschedules missed medical appointments without staff involvement. On oversight, the intake agent flags high value leads in real time for live transfer, and a live transcript lets staff read a call as it happens.

A live call intercept, released as one of the most requested features, lets them take the call over, and every call can be replayed afterwards. Section 7.3 of the Subscription Agreement limits the agents' authority to named tasks and states they will not give legal advice, and section 7.2 says human review of calls the AI handles may be necessary. The threshold is missing: what makes a lead high value, what a caller can say that forces a handoff, and what the agent may not do all sit in a configuration CaseGen operates for the firm and does not publish.

Plenty of volume, and nothing traced to a named firm. The home page quotes attorneys by first name and practice only, such as "Narimon, PI Attorney" on switching from an after hours answering service. The about page says CaseGen has handled thousands of calls for firms across the country, built by a founder with fifteen years of legal marketing behind the company. The most substantial evidence is the company's own intake benchmark of 13,280 calls across personal injury, family, criminal and immigration practices, which is real production data with a stated method and stated limits.

It is platform data pooled across unnamed accounts, so it shows the agents are in use rather than what changed for any firm that adopted them. No named firm, case study, before and after figure or dated outcome is published, and the partnership page invites applicants to call the agent rather than pointing to a customer.

The agreement speaks to privilege directly and also reserves a right to learn from the calls, and the two sit side by side. The Subscription Agreement commits CaseGen to keep all client communications strictly confidential and not to use or disclose attorney client privileged information except to provide the service. It will add safeguards the firm reasonably requests for privileged or highly sensitive material, train its systems to recognize sensitive legal information, and notify the firm immediately of any actual or suspected breach of confidentiality.

It acknowledges that, as the firm's agent, it may carry the same confidentiality obligations the firm does. Against that, section 7.1 lets CaseGen use the firm's data and call recordings to train, develop and improve its machine learning models, calling that use essential to its competitive advantage. Section 15.1 narrows the right to anonymized and aggregated data with no identifiable client information. The Privacy Policy adds that identifiable data is not shared across customers or used to train third party foundation models. Separation inside a firm is not described, and the third party AI platforms that hear the calls are not named.

The advice line is drawn clearly in the agreement, and whether to tell callers they are talking to AI is answered two ways. Section 7.3 of the Subscription Agreement appoints CaseGen as the firm's limited agent for answering calls, collecting intake, texting and transferring. It commits that the agent will clearly identify itself as an AI assistant working for the firm and will not give legal advice or practice law. Section 8.2 says CaseGen is not a law firm and that the firm remains responsible for its rules of professional conduct, including supervision of nonlawyer assistance, and section 7.9 addresses voice and biometric consent.

The company's own writing is softer on disclosure. An article on whether to tell callers they are speaking to AI says conversation design matters more than disclosure and recommends an opening built around purpose rather than a technical label. Its benchmark says an agent that confirms it is an AI the moment anyone asks is being honest. Jurisdiction limits beyond recording consent, which the agreement says vary by state, are not stated.

CaseGen publishes a good deal about reliability, including a layered fallback design with monitored flows, health checks and circuit breaking, and its benchmark is candid about its own limits. None of that is governance of the agents' behavior. No one inside the company is named as accountable for how the agents speak to callers, nothing describes what is tested before a new agent script, voice or model reaches a firm's phone line, and no finding is published on how the agents perform across accents, languages or callers in distress, which matters for a product that advertises more than 20 languages and answers people after accidents.

The Privacy Policy's statement that CaseGen evaluates prompt logic and agent behavior for service quality describes a use of data, not a governance process.

Most of the ground is published with numbers attached, without naming who else touches the data. The Privacy Policy (last updated 27 August 2026) sets default retention at 90 days from the call for recordings, transcripts and intake data and for outbound text and email content, adjustable on request. Billing records are kept two years and website analytics twelve months. Deletion requests are typically processed within 30 days, and residual copies in encrypted backups are overwritten under backup practice.

Encryption in transit (TLS 1.2 or later) and at rest, access controls, monitoring and logging are listed. The Subscription Agreement commits to notify the firm of a data breach within 48 hours and to give it what it needs for its own reporting duties. On termination, data may be deleted at CaseGen's discretion, so a firm has to export before it leaves. Subprocessors are described by category only (hosting, telephony, payments, analytics and support tools), and retention changes go through an account representative because customers cannot change configuration themselves.

The agreement does more than disclaim the product's risk; it hands it to the firm. Section 7.2 of the Subscription Agreement disclaims any guarantee that information collected on a call is accurate or complete, that the AI will interpret callers correctly, or that it will make appropriate judgments. Section 7.4 excludes liability for miscommunications and errors in calls the AI handles, for decisions the firm makes on what was collected, for delays in passing on urgent client information, and for anything the agents say while acting as the firm's agent.

The same section requires the firm to defend and indemnify CaseGen against claims arising from the agents' actions as its authorized agent, and section 7.9 adds an indemnity for biometric consent failures. The website Terms cap liability at fees paid in the three months before a claim. No indemnity, insurance or output warranty runs from CaseGen to the firm. The allocation is published and can be read before signing, and nothing in it gives the firm recourse when an agent mishandles a caller.

The integrations that matter for an intake product are named and described by what they move, in marketing prose rather than documentation. The Clio page says each completed intake appears in Clio Manage in real time as a lead with contact details, case type, intake notes and transcript, ready to review or convert to a matter, and that CaseGen's team sets the connection up. The Filevine page describes potential client information, intake details and documents flowing into Filevine automatically, with appointment notifications and reminders handled on the CaseGen side.

Lawmatics, 8am, Google Sheets and Zapier have their own pages, and the FAQ adds Calendly, LawPay, PracticePanther, Smokeball and MyCase. CaseGen's FAQ states a limit plainly: CasePeer cannot be connected because it offers no API. Depth an implementer could use is missing. There is no field mapping, no statement of whether any integration writes back into CaseGen, and nothing on the permissions the connection holds inside the case management system.

Where the data lives is stated in one sentence, and nothing about deployment is offered as a choice. Section 13 of the Privacy Policy says CaseGen's services are hosted in the United States and that personal information and customer data are processed and stored there, which answers both the storage and the processing question for a US firm. The policy describes a managed service rather than self service. CaseGen configures and operates the agents and the call handling on each firm's instructions, and customers cannot change configuration directly.

No cloud provider is named, and no single tenant, private or regional option is described. Nothing addresses where the third party AI platforms named in the Subscription Agreement process call audio, which may differ from where CaseGen stores the result. For a firm outside the United States or one with client data location requirements, the only published answer is that everything is in the US.

No SOC 2 report, ISO certification, HIPAA attestation, penetration test or trust center is published, and no route to request one is offered. What is published is a list of controls. The Privacy Policy names encryption in transit and at rest, access controls, monitoring and logging, and the Subscription Agreement commits to regular security assessments and vulnerability tests without a named assessor or date. That matters more here than for a typical intake form.

The medical coordination agent calls providers about treatment and the product carries health information for personal injury firms, and no business associate agreement is offered or referred to.

The architecture is described in detail, and the companies inside it are not identified. CaseGen's engineering article lays out four layers: speech to text, a language model, call routing and speech output. It says each layer has multiple providers and a fallback. An alternate speech engine takes over when confidence drops (with Whisper given as an example), a retry, rephrase or backup model handles a response that comes back empty or hallucinated, and an alternate voice engine steps in when output fails.

The Subscription Agreement says calls are processed by CaseGen's own systems and by affiliated third party AI platforms, and the Privacy Policy says identifiable data is not used to train third party foundation models. A pricing article says its agents run on premium models and dismisses cheaper ones as unfit for legal intake, without naming either. A firm learns that several outside models hear its callers and that one can be swapped for another during a call. It does not learn whose they are, where they run, or whether it will be told when the set changes.

The shape of the bill is published and the numbers are not. A CaseGen pricing article says charges are set to the firm's call volume and billed per call, plans run month to month with no long term contract, and an implementation fee applies depending on the setup. The Subscription Agreement fills in the mechanics. Fees are quoted and payable in US dollars, a monthly recurring fee is due regardless of usage, and subscribers who pay per use are also billed for actual call usage at the rates in their sales order.

CaseGen's pricing page shows no figure. It invites the firm to talk to Justin, an AI discovery agent, for about ten minutes to receive an estimate, and is restricted to law firms, with consultants and agencies sent elsewhere. CaseGen's comparison page quotes a competitor's published plan prices to the dollar while withholding its own. A firm can understand what it will be charged for before a sales conversation, but not what it will pay.

Built for law firms only, across a named list of practices, with firm size and the edges of the service left open. CaseGen has dedicated pages for personal injury, workers' compensation, lemon law, employment, family, criminal, immigration, real estate and business law, and a set of language pages covering Spanish and around 20 others, each describing intake in that language. The pricing page is restricted to law firms, the medical coordination agent is specific to personal injury, and the integration FAQ names one system it cannot serve (CasePeer).

What is not stated is which firms it fits: nothing says whether it is built for a solo practice fielding a few calls a day or for a high volume advertiser, what call volumes it has been run at beyond the pooled benchmark, or what an agent will not handle, such as existing client matters of a certain kind or calls that should always go to a lawyer. Practice area coverage is clear; fit for a given firm is not.