Bringing Intelligence
To Your Data.

MARC follows the same defensible workflows widely accepted in the legal field today, including the sampling, validation, and iterative improvement methodology used in keyword, TAR, and CAL processes. Because every MARC decision ships with the underlying reasoning and the source passages relied on, the process is more transparent — and we believe more defensible — than alternatives that produce a bare classification with no explanation.

MARC has two distinct layers that are treated differently for disclosure purposes. The matter-specific Protocols — covering relevance, privilege, confidentiality, and the other classification types — document the background criteria that drive decision-making for your matter. While arguably privileged work product, the Protocols are transparent and disclosable, and they are the appropriate disclosure surface for cooperation discussions with opposing counsel. The underlying MARC system prompts — the hard-coded instructions that govern the platform's reasoning engine — are Altorney's proprietary intellectual property and are disclosed only if compelled, and only with appropriate protective-order protections in place. In most matters, defense counsel can and should resist disclosure of the prompts themselves; the Protocols are the right disclosure surface.

The MARC Protocols are the foundation MARC uses to apply review decisions across a matter or investigation. The Relevance Protocol is automatically generated by a proprietary engine that ingests whatever inputs counsel makes available — complaints, counter-complaints, requests for production, subpoenas, review training protocols, sample documents, or even just a summary memo of what makes data responsive. None of these inputs is individually required; the process works from a summary memo alone if that is all that is available. The output is a Review Protocol used to assess every document in the corpus. Other Protocols are template-driven, covering specific instructions for privilege, confidentiality and trade secrets, personal vs. business communications, and similar classification types.

Yes. While a core advantage of MARC is that an attorney does not need to author the Relevance Protocol from scratch, attorney oversight, adjustment, and approval are strongly recommended before any production-scale processing. After sampling and validation, the Protocol can be iteratively refined to improve performance on the specific matter.

MARC exports fully coded decisions to CSV or directly into Relativity via API. Validation sets can also be reviewed inside MARC's built-in document viewer for sample-scale QC, or exported to your preferred review platform and imported back into MARC for calculation and analysis at scale.

GenAI models can hallucinate; MARC is built to prevent that from affecting outputs. The MARC Protocol acts as a grounding mechanism so that relevance decisions are tied to actual source documents and structured system instructions, and every output cites the specific passages relied on. MARC also applies internal guardrails between the raw LLM response and the final classification — checking that each output is grounded in the source document, conforms to the protocol, and contains the required citations. Outputs that fail these checks are automatically re-processed before they are committed. Finally, sampling and validation let teams verify the accuracy and defensibility of the system's decisions before any data proceeds.

That is ultimately a question for counsel. MARC results support easy and quick QC review. In situations where speed or budget demands it (second requests, internal investigations), MARC results can be sampled and produced with the same defensibility as any other technology-assisted review approach.

The typical MARC deployment is inside the client's own environment — Azure, AWS, GCP, or on-premises hardware in the client's data center — using pre-configured Kubernetes or Docker scripts. Documents are analyzed in place. In this configuration, no corporate data leaves the firewall at any point unless the client has specifically routed the analysis to an authorized external LLM provider instead of the on-board open-source models that ship with MARC.

Yes. The typical and recommended deployment is in the client's own environment, because data sovereignty, regulated-industry controls, and matter-long records governance are strongest when MARC runs behind the client's firewall. For clients who prefer faster setup, who don't want to manage infrastructure, or whose matters don't require in-environment deployment, MARC is also available as a hosted SaaS service running in Altorney's SOC 2 environment. The deployment choice is made per client.

Altorney is SOC 2 certified. In a typical deployment, however, Altorney does not have access to client data — MARC runs entirely inside the client's environment, so the client's existing controls (DLP, audit logging, certifications such as ISO 27001 or HIPAA, and so on) govern data handling end-to-end.

Within MARC itself, built-in security controls include encryption at rest and in transit, and role-based access control. For SaaS deployments, SOC 2 controls apply to Altorney's hosted environment.

Yes. MARC is designed to be deployed according to your IT and Security team's standards, and the deployment environment can be isolated to whatever degree your security team requires. The pre-configured Kubernetes and Docker deployment scripts can be reviewed by your team in advance.

The MARC management portal is accessible via the internet for software updates, but no corporate data flows through that path. Corporate data can touch the internet in only two scenarios. First, if your organization has specifically approved an external LLM provider such as OpenAI in lieu of the on-board models that ship with the platform. Second, if you use the MARC Assistant chatbot, which is optional and requires a foundation-level model such as Gemini, GPT-5.x, or Claude — the chatbot can be routed through the client's own approved subscription using the client's API key.

MARC is model-agnostic. We have tested it on OpenAI, Azure AI, and local open-source models including Llama and Gemma. Model choice affects speed and cost, and the best choice is whichever your security team has vetted and is comfortable with. For most clients we recommend Gemma running locally on your own cloud or on-prem resources — fastest, most cost-effective, fully behind the firewall, and stable: once deployed, the model remains exactly as validated, with no surprise version changes or vendor deprecations during a multi-year matter.

The local LLMs (such as Llama or Gemma) running in your environment retain nothing — the model servers exist only during document analysis and are destroyed when the job completes. Any traffic routed by client choice to major external providers (e.g., OpenAI) is subject to the client's enterprise terms with that provider, which typically prohibit retention or training on client data.

MARC is designed for fast IT/Security approval. Installation and operation happen entirely in your approved cloud or on-prem environment using deployment scripts your team can audit. LLM traffic is either contained within the MARC environment or routed to an external provider that has already been approved under your AI policy. No data leaves the firewall without explicit authorization. Many corporate IT and security teams have approved MARC within days rather than weeks. Installation and operation do not require a dedicated cloud-ops team — once up and running, the system requires very little ongoing interaction from IT.

No. Your team has full access to the environment, but the MARC application manages the underlying infrastructure without manual intervention. The LLM servers come pre-configured with the MARC environment and are entirely managed by the system.

Because MARC runs inside the client's environment, all matter data — source documents, classification outputs, Protocols, and audit logs — stays under the client's existing records-retention and deletion policies. When a matter concludes, the client controls data destruction directly.

Yes. Altorney is a developer partner of Relativity. MARC has a direct API integration with Relativity for both ingestion and output. MARC automatically creates the required fields in Relativity, and results flow in along with the source documents (if those documents aren't already loaded in the workspace). Results can also be exported as CSV overlay, TXT, or native files for any other review platform.

MARC operates on extracted text and images. Native document processing — extracting text from PDFs, Word documents, emails, spreadsheets, and other formats — is typically handled by an upstream processing tool like Nuix or Relativity, and the extracted text is then fed to MARC. Audio and video analysis are currently in beta. For smaller matters and contained investigations, MARC can also ingest PST files and loose Microsoft 365 documents directly, and can pull data from the Microsoft Graph API with appropriate permissions; these direct-ingestion paths are not intended to replace processing tools at production-scale volumes.

Foreign-language documents are not a constraint. MARC analyzes content in its source language across most major languages, with extensive testing in Chinese, Japanese, and Korean (CJK).

MARC is purpose-built for in-house teams. The MARC Protocol automates the complex prompt-engineering work so internal teams aren't burdened with composing detailed instructions to get usable results. If you want your outside counsel or service-provider advisor involved in sampling and validation, MARC supports that workflow too.

MARC scales horizontally across as many GPU servers as the client allocates. We routinely see 500,000–1,000,000 classification decisions in a 24-hour span, with the exact throughput depending on the number of GPU servers available.

No. Your IT team runs a deployment script that spins up MARC as a Kubernetes cluster (or Docker containers, whichever fits your environment) in your cloud tenant or on on-prem hardware. We can provide the script in advance for your team to review exactly what MARC does. The full platform is installed and ready for processing in about one hour.

MARC is licensed at $6,500 per month for unlimited use — a flat fee regardless of document volume or how many classifications are run on each document. At typical matter scale, that works out to a fraction of a cent per document. The only consumption costs above the license are either (a) the underlying cloud machine time — typically $0.01–$0.04 per document when running local models like Gemma on your own cloud resources — or (b) token fees you pay directly to public LLM providers like OpenAI or Google under your own enterprise agreements, if your team has elected to use an external provider. On-prem deployments have no per-document compute charge beyond your existing infrastructure (typically a few hundred dollars per month in operating cost). MARC does not upcharge for LLM usage. For reference, traditional first-pass contract review typically runs $0.50–$2.00+ per document.

No. MARC does not upcharge for LLM usage. Your only consumption cost is either the Azure machine time for local models running in your environment or token fees you pay directly to a public provider like OpenAI or Google under your own enterprise arrangements — depending on which LLM path your team has approved.

No. MARC is a stand-alone solution from Altorney and completely separate from the Altorney Platform marketplace for sourcing and managing legal talent. You do not need to engage Altorney's review resources to use MARC, and you do not need to use MARC to engage Altorney reviewers. However, using both unlocks bundled discounts.