Harvey vs CoCounsel vs a Custom AI Agent for Small Firms

By Jude Lee · · Comparison

Three colleagues at a small law firm comparing AI tools on laptops in a conference room

The three things people mean by “law firm AI software”

Search for law firm AI solutions and you get a category that isn’t one category. Three architecturally different products share the label, and they fail in different places.

Vertical legal AI platforms. CoCounsel is Thomson Reuters’ legal AI assistant, built on the Casetext product Thomson Reuters acquired in 2023, and sits alongside Westlaw content. Harvey is an AI platform aimed primarily at large firms and in-house legal departments. Lexis+ AI is LexisNexis’s equivalent. Clio Duo is AI folded into a practice-management platform many small firms already run. These arrive with legal-specific workflows already defined — research memos, deposition prep, contract review — and, in the research products, links back to primary law.

A general-purpose assistant connected to your systems. Claude, ChatGPT, Copilot, and Gemini are not legal products. What changed is that they can be given governed access to your firm’s actual systems through MCP (the Model Context Protocol), an open standard for exposing data and tools to an AI assistant. That turns “a chatbot that writes decent prose” into “an assistant that can read matter 2024-0117’s file, draft the letter, and save it back.” MCP originated at Anthropic, but adoption is broader now: OpenAI, Microsoft, and Google have each announced MCP or MCP-compatible connector support in their agent and developer tooling, so a firm already paying for Copilot or Gemini is not automatically shut out. Verify current support in each vendor’s own documentation before you plan around it. The mechanics of one such connection are in connecting Claude to Clio with MCP.

A custom agent. Software your firm commissions: a multi-step process that runs against your data, with your rules, usually calling a general model underneath. An intake agent that qualifies a lead, runs a conflicts search, and opens a matter is the canonical example.

The option that isn’t AI at all

Before any of this: a lot of legal document work is deterministic and should stay that way. Document assembly, clause libraries, matter-type checklists, and the workflow automation already sitting unused inside your practice-management system produce the same output every time, need no verification pass, and cost a fraction of an AI seat. For engagement letters, standard pleadings, routine discovery requests, and intake forms — several of the workflows firms most often shop AI for — templates are cheaper and more reliable than any of the three products above. Use AI where the input varies unpredictably. Use templates where it doesn’t.

What each is genuinely good at

Vertical platforms are strongest where the hard part is legal content you don’t own: case law, statutes, secondary sources, standard-form review against market norms. They also carry a real governance advantage — the vendor has thought about confidentiality, retention, and audit logs, and will sign the paperwork. For a two-lawyer litigation shop that mostly needs faster, better-cited research, this is the shortest path.

Assistant-plus-MCP is strongest where the hard part is your own material: the 400-page production in the Henderson matter, the fee agreement you always send, the client email thread nobody has read since March. No vendor knows your file naming conventions, your local judge’s standing order, or the way your firm phrases a demand letter. MCP is how you hand that context over without pasting documents into a chat window. It also composes across systems more cheaply than one-off integrations, because a single connection can serve drafting, summarizing, and status questions. The caveat: MCP is a young standard, and connector security, authentication, and scoping are your responsibility, not the protocol’s.

Custom agents are strongest where a workflow is repetitive enough that you want it to run the same way every time, and specific enough that no vendor will build it for you.

Off-the-shelf legal AI
Live in weeks. Vendor handles security review, model updates, and legal content licensing. Workflows are pre-defined — fast if they match how you work, frustrating if they don’t. Per-seat cost is predictable, but watch for seat minimums and annual commitments that price a five-lawyer firm like a much larger one. Ask what leaving costs: exporting saved workflows, prompt libraries, and generated work product is rarely free or quick. And vertical does not mean immune — legal-specific tools still mis-summarize and still produce output whose citations need checking.
Assistant + MCP (or custom agent)
Weeks to months, and needs someone technical (in-house or contracted). Fits your file structure, your templates, your matter types. Runs across systems the vendor never integrated with. You own the maintenance, the prompt quality, the connector security, and the failure modes. Costs shift from subscription to build plus usage.

Pick the workflow first, then the product

The reliable way to avoid buying three overlapping tools is to refuse to evaluate tools until you’ve named the job.

  1. Name one workflow and count it

    Not “drafting.” Something like “first-pass review of incoming commercial leases” or “summarizing depositions before mediation.” Count roughly how often it happens per month and who does it today.

  2. Ask whether the input actually varies

    If the document is the same every time with fields swapped, stop and build a template. AI is for variable input.

  3. Ask whether the hard part is legal content or firm content

    If the work depends on case law, statutes, or market-standard clauses you don’t hold, lean vertical platform. If it depends on documents already sitting in Clio, NetDocuments, SharePoint, or email, lean assistant-plus-MCP.

  4. Pilot on real files, under a real agreement

    Run the same three actual matters through each candidate. Have the responsible lawyer grade output on a fixed rubric. Confirm data-handling terms before any client material goes in.

  5. Check the last mile

    If the tool produces good output but a human still retypes it into your system, you’ve bought a demo. Ask specifically: can it write back?
Almost nothing in this category fails because the model can’t write. It fails at the seams — permissions, file structure, and who owns the thing in month six.

Modeling the money without inventing numbers

Ignore every headline savings figure, including any you read here, and do your own arithmetic. The three inputs below are placeholder values chosen to show the shape of the calculation — substitute your own before deciding anything.

6 hrs/mo
Placeholder: hours saved on one named workflow
2 hrs/mo
Placeholder: lawyer time spent verifying output
$X/mo
Your actual quoted subscription plus build cost

The formula: (hours saved − verification hours) × your loaded hourly cost, minus subscription and build cost, gives net capacity. To turn capacity into revenue, multiply recovered hours by your realization rate and billing rate — and only if someone actually reallocates those hours to billable work. If nobody does, you’ve bought a better workday, which is legitimate but shouldn’t be counted as revenue. A fuller version of this arithmetic is in the automation ROI walkthrough.

The disadvantages you should price in

Fabricated citations remain the headline risk, and the sanctions record is public — Mata v. Avianca (S.D.N.Y. 2023) is the case every practitioner should read before letting unverified AI output near a filing. Vertical research tools reduce but do not eliminate this; verify every cite against the primary source.

On ethics, the ABA’s Formal Opinion 512 (July 2024) addresses generative AI directly — competence, confidentiality, client communication, supervision, and how fees should reflect AI use. Read it in full rather than a summary. For state guidance, most state bars publish a searchable ethics opinion index on their own website, and many have a member ethics hotline; start there rather than with a general web search, since secondary summaries lag. Your malpractice carrier’s risk-management line is the other free resource firms underuse. Where client confidences, privilege, or fee arrangements are involved, confirm your approach with a qualified ethics professional before rollout.

The quieter disadvantage is drift. A workflow that worked in January degrades when your templates change, the model updates, or the person who tuned it leaves. Someone has to own it, by name.

What most small firms should actually do

My opinion, stated as opinion: for a firm under roughly twenty lawyers, the default stack is templates and document assembly for anything repetitive, one vertical tool where legal content is the bottleneck, and a general assistant connected to your own systems for everything else — with no custom build until a specific workflow has proven itself painful for two consecutive quarters. Harvey and CoCounsel are not competing with an MCP connection to your document system; they solve a different half of the problem. Buying both deliberately is reasonable. Buying both by accident, for the same job, is how firms end up paying twice and using neither.

Where is your firm losing billable hours?

Get a free automation audit: we map your intake-to-invoice workflow and show you exactly what's worth automating — before you spend a dollar.

Get a free automation audit

Related guides