Copilot for Legal vs Harvey vs Custom AI for Small Firms

By Jude Lee · · News

Lawyers in a small firm conference room comparing legal AI options on a shared screen

What was announced — and what you still have to verify

In October 2026, legal trade coverage (including Artificial Lawyer) reported a legal-focused configuration of Microsoft Copilot. Reports in the same window described Harvey shipping an intake product aimed at in-house teams, Legora expanding its third-party integration ecosystem, and eBrevia launching a contract management product. I have not independently verified any of these claims. Treat the dates, the feature scope, and any integration counts as things to check against each vendor’s own announcement page and product documentation — not as settled fact.

The clustering matters more than any single headline. Every layer of the market is converging on the same promise: an assistant that sits inside your documents and systems and does multi-step work.

Three different products wearing the same word

The fastest way for a firm under about 30 lawyers to waste a year is to evaluate these three as if they were alternatives to each other.

Suite-level legal AI (Copilot and similar)
Lives where your lawyers already are: Word, Outlook, Teams, SharePoint. Low marginal cost if you’re already on Microsoft 365. Strength is breadth — summarize this thread, draft from this precedent, find that file. Weakness is that it reasons over whatever you happen to have stored, with whatever permissions you happen to have set. It does not know your firm’s clause standards or your jurisdiction’s filing quirks unless you teach it.
Specialist legal AI (Harvey, CoCounsel, Legora)
Built for specific legal jobs — research with citations, contract review, discovery, deposition work. Deeper guardrails, legal-specific evaluation, and vendor-maintained workflows. Costs meaningfully more per seat, and you adopt the vendor’s opinion of how the work should be done. Integration with your practice-management system varies a lot; ask for a live demo against your own matter data, not a canned one.

The third option is structurally different again: connecting an assistant like Claude to the systems you already run — Clio, NetDocuments, iManage, your billing system — using MCP, the open Model Context Protocol for giving an AI governed access to tools and data. As of 2026 this is a real, usable pattern rather than a lab demo, and we’ve walked through the mechanics in connecting Claude to Clio via MCP. A custom build isn’t “better AI.” It’s the same frontier models pointed at your data, with your permission rules and your packaged instructions.

It also carries costs the Compare table above doesn’t charge the other two options. Someone has to maintain the connector as the underlying APIs and models change, and re-run evaluations when a model update quietly shifts output quality. You get no vendor indemnity and no legal-specific guardrails out of the box — you build the citation checks and the refusal behaviour yourself. There is key-person risk: if the person who wrote the server leaves, the firm owns code nobody understands. And if you serve insurer or corporate clients with security addenda, you absorb the internal review burden that a named vendor would otherwise answer for.

You are not buying intelligence. You are buying access, guardrails, and repeatability — and those three things are where the vendors actually differ.

The honest answer: a large share of what gets sold as AI in legal is deterministic automation with a language model stapled to the front — and that’s often fine. Document assembly from a questionnaire, deadline calculation from a trigger date, and conflict-check searches against a client list are rule-based jobs. They should stay rule-based, because rules are auditable and they don’t fail the way models do.

And models do fail, in recognizable ways. They fabricate or mis-cite authority; they surface a precedent that has since been overruled or superseded with exactly the same confidence as good law; and — the quiet one — they silently skip a document they lacked permission to open, then summarize the record as if it were complete. None of those announce themselves in the output. That’s the whole argument for a named human reviewer.

What’s genuinely new is the agentic layer: a system that plans several steps, calls tools, reads results, and decides what to do next. That’s what separates “summarize this PDF” from “pull the three newest medical records in this matter, extract treatment dates, flag gaps over 30 days, and draft a chronology into the matter folder.” We draw that line in agentic AI vs generative AI in law firms.

The problem none of the three solves for you

Every one of these options reads from your repository. If your precedents live in six places, half your matters have inconsistent folder structures, and nobody has audited who can see which client folder, then suite AI will confidently surface the wrong template, the specialist platform will underperform its demo, and a custom MCP server will just expose the mess faster.

This is a confidentiality question, not only a quality one. Under ABA Model Rule 1.1 Comment 8, competence includes keeping abreast of the benefits and risks of relevant technology, and Model Rule 1.6(c) requires reasonable efforts to prevent unauthorized disclosure of client information — check your own state bar’s adopted language and any technology ethics opinions, since states vary, and confirm anything with real compliance stakes with qualified counsel. Practically, permission scoping comes before model selection. We cover the mechanics in read-only vs write access for law firm AI agents.

Model the money with your own inputs

Don’t accept any vendor’s ROI slide. Build the arithmetic yourself:

hours saved/week × 52 × realization rate × billing rate
Recovered capacity — plug in your own figures
(seats × monthly price × 12) + implementation + training + maintenance
True annual cost of any option
Pilot one workflow, two people, 30 days
Smallest honest test

The deciding term is the realization rate: recovered hours only become revenue if there is billable work waiting to absorb them. If your firm is capacity-constrained, savings convert cleanly. If you’re demand-constrained, the gain shows up as reduced overtime, fewer missed deadlines, and faster turnaround — real, but not the same as cash.

A 30-day way to decide

  1. Pick one bleeding workflow

    Name a single job with volume and a predictable shape — intake triage, discovery response drafting, lease abstraction. Not “use AI for research.”

  2. Audit the data that job touches

    Where do the source documents live? Who can see them? Is the format consistent? If the answer is ugly, fix that first — it’s cheaper than any license.

  3. Try the cheapest option first

    If you already pay for Microsoft 365, test the suite-level offering on that workflow for two weeks with two people. You may discover it’s good enough for most of the volume, which changes the entire budget conversation.

  4. Test the specialist against the same task

    Run the identical matter set through a specialist platform trial. Compare output quality and, crucially, how much editing each required.

  5. Only then price a custom build

    If both off-the-shelf options failed for the same reason — they couldn’t reach your systems, or couldn’t follow your firm’s specific standard — that’s the signal for an MCP-connected assistant with packaged skills. If they failed because the task was genuinely hard, custom won’t fix it either. Price the maintenance and the security review, not just the build.

  6. Write down the review rule before you scale

    Decide who verifies citations, who confirms the agent actually saw every document it should have, and who signs off on anything that leaves the firm. Record it.

The honest bottom line

For most firms under about 30 lawyers, the sequence is: suite AI for everyday drafting and triage, one specialist tool for your highest-volume practice-specific workflow, and a custom MCP layer only when your constraint is clearly your own systems. Many firms will never need the third. Some will need only the first.

One structural point belongs in that budget conversation. The work most exposed to these tools is the work that was profitable mainly because a junior billed it slowly — which means tooling spend is really a decision about your leverage model, not a software line item. Judgment, client relationships, courtroom work, and accountability for the final product are not exposed. Our AI agent vs paralegal vs outsourcing math walks through how to model that for your own staffing.

The mistake isn’t picking the wrong option. It’s picking all three at once, with no named owner and no pilot — which is exactly how firms end up paying for three overlapping subscriptions and still doing the work by hand.

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