Ertas + OpenClaw for Law Firms

The Challenge

OpenClaw's ability to triage emails, review documents, extract contract terms, and draft correspondence makes it a natural fit for legal workflows. Law firms, though, work under confidentiality obligations that make the default cloud API setup a hard sell to their own risk committees.

Attorney-client privilege protects communications between lawyers and their clients from disclosure. When a lawyer uses OpenClaw with a cloud API to review a privileged document, the document's contents are sent to the model provider's servers as prompt input. Whether that puts privilege at risk depends on the provider's terms, the jurisdiction and the firm's own safeguards, and many firms would rather not have to argue the point.

Beyond privilege, law firms carry obligations around client data. The Australian Privacy Act, GDPR (for firms with EU clients), and bar association rules all bear on how client data is stored, processed, and transmitted. Every commercial AI API that touches client documents is another vendor to assess under each of them.

The tasks OpenClaw handles well (document review, email classification, clause extraction, memo drafting) are the tasks that consume the most paralegal and junior associate time. The productivity gain is real, and the open question is where the model runs.

The Solution

Ertas lets law firms run OpenClaw on a custom model that the firm hosts itself. The setup is straightforward: fine-tune a model on the firm's legal domain data in Ertas Studio, export it as GGUF, serve it with Ollama on a firm-controlled server, and point OpenClaw at that local endpoint.

Training runs on Ertas's cloud GPUs, so the training set is uploaded to Ertas. Firms should anonymise or minimise it first and confirm that using the material for training fits their engagement terms. After export, the model runs on the firm's server, so the documents, emails and matter details OpenClaw reads during daily work stay inside the firm's network.

A model trained on the firm's actual work product learns its specific criteria: what counts as an unfavourable indemnification clause, which liability caps are acceptable for different deal sizes, which jurisdiction-specific terms need flagging. A general-purpose model applies general legal knowledge, while a trained model applies the firm's standards, and the firm measures the difference on its own held-out examples before relying on it.

Multi-practice firms can train a separate model for each practice group. The corporate team's model learns from M&A documents and due diligence checklists, the litigation team's from discovery responses and motion drafts, the IP team's from patent specifications and prosecution history. Each model is a file the firm owns and can host behind its own access rules.

Key Features

Example Workflow

A mid-size commercial law firm in Sydney deploys OpenClaw to speed up contract review for its corporate advisory practice. The firm's paralegals currently spend 8 to 12 hours per week reviewing vendor agreements, flagging unfavourable terms, and extracting key dates and obligations into a matter management system. The firm exports 400 annotated contracts from its document management system, each with paralegal-marked clauses (flagged or acceptable) and extracted data fields, and removes client names and identifying deal terms. The dataset is uploaded to Ertas Studio, where a Llama 3.1 8B base model is fine-tuned using LoRA (rank 16, 3 epochs) alongside a Qwen 2.5 7B run for comparison. The team scores both on a held-out set of contracts, and against the prompt-engineered frontier model it had been testing, using the same criteria. The chosen model is exported as GGUF (Q5_K_M quantisation) and served with Ollama on a Mac Studio in the firm's server room, and OpenClaw is configured to use that local model. The contract review agent now reads incoming contracts via email, extracts key terms into structured JSON, flags unfavourable clauses with explanations referencing the firm's criteria, and drafts a review memo for the supervising partner. Paralegals verify the flagged items, and their corrections become training examples for the next run. Running costs are the firm's own hardware plus the Ertas subscription, with no per-token API bill for the agent's day-to-day work.