Evaluating AI Legal Assistants for Law Firms

You are staring at a pile of intake files on a Tuesday afternoon. Reconstructing the timeline from three months of email threads, scanned PDFs, and handwritten notes will take four hours you do not have. You have tried running a preliminary summary through a consumer chatbot.

The output looked clean, until you realised it cited a jurisdiction that did not exist in the document. That single error meant re-reading the entire file from scratch. The tool did not save time. It created liability.

This is the triage trap. And the way out is not faster reading – it is a different category of tool.

The Triage Trap: Why the Obvious Solution Creates a Bigger Problem

Legal AI is not a new idea. What is new is that the tools have reached a level where the output is genuinely useful on real, complex matters, and that has created a dangerous moment for law firms. The instinct is to reach for whatever AI tool is already installed. ChatGPT is open in another tab. Copilot is in the Microsoft 365 sidebar. Why not use it?

Because consumer-grade AI is unsafe by default for privileged data. These models train on user inputs, fabricate facts, and invent case law. A practitioner who uploaded a commercial agreement to a generic AI to extract governing law clauses found that the tool had hallucinated a jurisdiction not present anywhere in the document. The error was only caught on manual review. It permanently broke trust in the tool for privileged work.

The structural problem is not the model. It is the verification surface. When a tool produces an output with no link back to the source, the attorney has to re-read the document to check it. That is not a time saving. It is a time transfer – from reading to checking, with added malpractice risk.

The Three Tiers of Legal AI: Where Most Firms Get Stuck

The AI Legal Assistant Software market splits into three distinct tiers. Understanding where each one sits explains why most firms are stuck using tools that are either unsafe or inaccessible.

Market Tier	Security Posture	Access Model	Output Verifiability	Cost
Consumer AI (ChatGPT, Gemini)	Unsafe by default. May train on inputs.	Public. No procurement.	None. High hallucination risk.	Low upfront. High risk cost.
Enterprise Legal AI (Harvey, CoCounsel)	Legal-grade privacy.	Demo gate. Long procurement cycles.	High.	Five to six figure annual commitments.
Self-Serve Legal AI (Lexity)	ISO 27001. Swiss Tier IV. Zero LLM retention.	Pay-as-you-go. No contract.	Every claim cited to exact line, exact paragraph.	Pay-as-you-go/pay based on what you use.

Solo practitioners and boutique firms are structurally disadvantaged by this market. Consumer tools are unsafe for client files. Enterprise tools are priced and procured for firms ten times their size. The self-serve tier – secure, attorney-vetted, and accessible without a procurement cycle – is where boutique firms can compete at the level of much larger practices.

Core Capabilities: What Legal AI Actually Does Well

Legal AI replaces unbillable administrative work with structured, verifiable output. It does not draft final submissions. It does not apply legal judgment. What it does is eliminate the structural work that has to happen before the real work can begin – timeline reconstruction, document triage, clause extraction, gap identification – and it does it in minutes rather than hours.

Two Workflows You Can Run Today

Running a Closing Checklist Extraction

Extracting deliverables from a 150-page SPA is a three-to-four-hour task for a junior associate. A boutique M&A partner can run the same extraction in under five minutes.

1. Upload the signed or negotiated SPA into the isolated Lexity workspace. The document never leaves a protected environment.
2. Select the Closing Checklist Clickflow™ from the library. No prompt engineering required – the workflow is pre-configured by M&A practitioners.
3. Run the Clickflow™. The output is a responsibility matrix identifying every signatory, condition precedent, and deadline in the document.
4. Click any row to verify the citation – each item links directly to the exact paragraph in the SPA where the obligation appears.
5. Download and share with the transaction team. Every row is sourced, auditable, and ready for execution.

What you get: A complete closing checklist with source citations, in under five minutes, with no manual reading required.

Executing a Claim Challenger Workflow

A litigator needs to identify inconsistencies in an opposing statement of claim before the first hearing. Manual cross-referencing of a complex commercial dispute can normally take days.

Here’s what you can do instead:

1. Upload both the opposing counsel's statement of claim and your client's evidentiary documents.
2. Select the Claim Challenger Clickflow™. The workflow cross-references the opposing claims against the actual evidence in your file.
3. Run the Clickflow™. The output identifies factual inconsistencies, evidential gaps, and logical weaknesses in the opposing argument, each one cited to the exact document and paragraph that contradicts it.
4. Use the cited findings to develop targeted lines of inquiry and a defensible rebuttal. Every point is grounded in the file, not in the model's general knowledge.

What you get: A structured inconsistency analysis with source citations, ready to inform your litigation strategy.

Security: What Legal-Grade Actually Means

Legal-grade security is not a marketing category. It has three specific, verifiable properties. Any platform that cannot demonstrate all three should not be used for privileged client data.

Security Standard	Consumer Cloud	Legal-Grade (Lexity)
Hosting environment	Shared global servers.	Isolated user workspace. Swiss Tier IV.
LLM data retention	Models may train on inputs.	Zero data retention by any LLM. Contractually enforced.
Third-party audit	Internal privacy policies.	ISO 27001 certified. Verifiable at trust.lexity.ai

Data residency. Data residency means knowing exactly where the file is stored, not just that it is "in the cloud." Platforms using Swiss Tier IV servers under Swiss law provide the baseline physical security required for sensitive client files.

• Zero data retention by LLMs. The language model layer must not retain inputs or outputs for training or cross-session context. This is distinct from the platform's own storage – it is a contractual requirement imposed on every LLM provider the platform uses. Make sure to verify this at the provider's trust portal and not in the marketing copy.

• ISO 27001 certification. Third-party audited certification confirms that the platform meets the international standard for information security management. Make sure to request the certificate directly and verify it at the public trust portal: trust.lexity.ai

Frequently Asked Questions

Is a consumer chatbot safe for legal work involving client files?

No. Consumer models are trained on user inputs by default and do not provide source-verifiable citations. For any work involving privileged client data, you need a platform with zero data retention by LLMs, isolated user workspaces, and third-party certifications. Consumer AI is useful for non-privileged, non-sensitive drafting tasks, not for client files.

What is the difference between Harvey and Lexity?

Harvey is an enterprise legal AI platform requiring a demo, a procurement cycle, and a significant annual commitment. It is built for large law firms with dedicated legal technology teams.

In contrast, Lexity is a self-serve platform – pay-as-you-go, no contract, no minimum spend, $100 in credits to start. Both offer attorney-vetted workflows, but Lexity is designed for solo practitioners and boutique firms that find it hard to navigate an enterprise procurement process.

Does zero data retention by LLMs protect attorney-client privilege?

Yes, as part of a broader security architecture. Zero data retention means the language model layer does not retain your inputs or outputs for training or cross-session context. Combined with data residency in a strong jurisdiction like Switzerland and ISO 27001 certification, this addresses the core technical requirements for privileged data. The professional obligation to maintain privilege remains with the attorney – the platform's job is to ensure the infrastructure does not undermine it.

Can a small law firm afford enterprise-grade legal AI?

Yes, through the self-serve tier. Enterprise platforms like Harvey and CoCounsel require annual commitments typically ranging from five to six figures.

Lexity operates on a pay-as-you-go model: $100 in welcome credits gets a solo practitioner or small team through a meaningful number of real-matter executions, with no subscription and no minimum spend. The same ISO 27001 infrastructure, without the enterprise gate.

What legal AI workflows work without a dedicated IT team?

Attorney-vetted Clickflows™ are designed to run without configuration, prompt engineering, or technical setup. You upload the document, select the workflow, and run it – no IT team required. Platforms like Lexity are specifically built for lean firms that cannot spare a legal technology specialist.

Does legal AI work for non-English documents?

Yes! Lexity supports translation across 40+ languages while preserving the original document's formatting and structure. The translation gives the attorney a working draft rather than a blank page. Jurisdiction-specific legal interpretation of translated content remains with the attorney – the platform handles the structural and linguistic work and not the legal judgment.

How do I verify a legal AI platform's security credentials?

Ask for the ISO 27001 certificate and verify it directly at the provider's public trust portal. Request the sub-processor list to confirm which LLMs are used and what their data retention terms are. Ask specifically whether the LLM providers are contractually prohibited from training on customer inputs, not just whether the platform has a privacy policy. For Lexity, all of this is available at trust.lexity.ai.

Start With One Workflow on One Real Matter

The right way to evaluate any legal AI platform is not a demo. It is a real document, a real matter, and a real output you can check against what you already know.

Run the Closing Checklist Clickflow™ on an SPA you recently closed. Compare the output against your own checklist. Every discrepancy is a data point about whether the tool earns a place in your practice.