AI Paralegal Software: What a Small Firm Should Test Before Paying

AI paralegal software does three jobs: research assistants find and summarise law, drafting add-ins write and mark up documents inside Word, and matter or intake tools sort enquiries and summarise files. Test each with your own closed matters over a week, scoring accuracy, citations, confidentiality and time saved before you pay.

"AI paralegal" is a marketing label, not a product category with a standard meaning. No tool carries a paralegal's responsibility for deadlines, conflicts and facts, and the best-priced options are mostly quote-only, so you cannot compare them from a web page. What you can do is run the same tasks through each one and see what comes back. A firm that sets its pass marks first, such as nine of ten chronology facts right and no invented citations, ends a week's trial with a decision rather than an impression.

Follow me on Instagram@sagnikteaches

What the label covers: three kinds of tool

Research assistants

These answer legal questions from a vendor's database and draft from the results. Lexis+ with Protégé (renamed from Lexis+ AI in February 2026) covers research, drafting, summarisation and analysis of uploaded material, and its Shepard's Verify feature checks citations in AI answers and drafts and flags unsupported statements. Its uploaded documents sit in a "Vault": up to 50 Vaults of 1 to 500 documents each, and uploads of up to 10 documents outside a Vault are purged at the end of the session. Thomson Reuters' CoCounsel Legal handles research, analysis and drafting from one query, and its tabular review tool can review up to 10,000 documents against 100 questions, with each citation linked so a lawyer can click through. Both say customer data is not used to train their models.

Connect on LinkedInSagnik Bhattacharya

Drafting and review add-ins

These sit inside Microsoft Word. Spellbook's page lists Review (redlines and risk flags), Draft, Ask, Compare and Playbooks to encode your own positions, plus a multi-document agent it calls Associate. It states that it has Zero Data Retention agreements preventing use for training. These suit firms whose paralegals spend hours marking up agreements, leases or standard letters.

Subscribe on YouTube@sagnikteaches

Matter and intake tools

Clio's Manage AI (formerly Clio Duo), as Clio's help centre describes it, works on data already in the practice management system: matter summaries limited to matters you have permission to see, suggested time entries from unlogged communications, notes and tasks, and creation of notes, tasks and events. Intake is a separate job, handled by receptionist or enquiry tools such as Smith.ai, whose AI receptionist pricing page lists a free plan covering 25 calls, then $3 a call. One search result you may meet, Streamline AI, is aimed at in-house legal teams rather than small firms, so it is a different buyer's tool.

For a plain view of which tasks belong to people and which to software, how to divide the work between paralegals and AI sets out a split you can adapt.

Quote-only pricing: what to ask and how to compare

Lexis says Protégé is not sold on its own and that pricing is customised to organisation size, practice needs and subscription, and it asks you to contact sales for a quote. Spellbook says pricing is structured around the number of team members on your licence; it offers a 7-day free trial and gives personalised prices after a sales conversation. CoCounsel's page links to plans and trial requests but prints no figures. Treat that as the normal state of this market and put the same written request to every vendor.

Subject: Quote request, small firm, [number] fee-earners

We are a firm of [number] lawyers and [number] paralegals working mainly in [practice areas]. Please quote for [number] users, giving:
1. Annual price, and whether it is per user or per firm.
2. What is included (research content, citations checks, drafting, uploads) and what costs extra.
3. Contract length, renewal terms and notice to cancel.
4. Any limits on documents, matters or queries.
5. Your written position on confidentiality, retention and training on our data.
6. A trial on closed matters, and how long it lasts.

We are deciding by [date].

Then compare on scope, not on headline price. A cheaper quote that excludes citation checking or caps uploads at a low figure may cost more in staff time than a dearer one. Cost thinking follows the same pattern as in legal AI pricing for small firms: write the quote next to a number for the work it replaces.

Here is an illustrative sum. A firm of three lawyers and two paralegals spends about 60 hours a month on first-pass document summaries, chronologies and routine drafting. If a tool takes a third of that off the first pass, after checking time is allowed for, that is 20 hours a month. At a paralegal's loaded cost of $40 an hour, that is $800 a month, or $9,600 a year, as the ceiling on what a tool covering that work is worth to the firm. A quote of $6,000 a year clears it; a quote of $15,000 does not, unless the tool also wins work or reduces write-offs. The 20 hours is the number to measure in the trial, since it is the one the vendor cannot give you.

Checklist before any client material goes in

Do this before the trial, not after. Each item says why it matters and how to verify it.

1. Is client data used to train models? Why: confidentiality and privilege. Verify: the vendor's written terms, not a sales email. The three research and drafting vendors above each state that they do not train on customer data; confirm the wording in your contract.
2. Where is data stored and for how long? Why: retention duties and deletion on request. Verify: ask for retention periods for uploads, queries and outputs. Lexis, for example, states a 90-day history for Vault results and session purge for small uploads.
3. Who at the vendor can see our data? Why: staff access and sub-processors. Verify: the data processing terms and the vendor's security attestations such as SOC 2 reports.
4. Does our professional regulator have guidance on AI? Why: supervision and competence duties. Verify: your regulator's published guidance, and ask your insurer whether AI use changes cover.
5. Do clients' engagement terms permit this? Why: some corporate clients restrict AI use on their matters. Verify: check outside counsel guidelines before testing on those files.
6. Who is the responsible lawyer for AI output? Why: someone must sign off. Verify: name the person and the review step in the firm's written AI policy, and check each quarter that the name is still current.

Use closed matters, anonymised if possible, for the trial. If a vendor's terms are unclear on any item, stop and ask rather than proceeding on a free trial.

Six tests for a week-long trial

These tests use material you already know the answers to. Set the pass mark first.

Test 1: the citation trap

Take a closed matter with a clear legal question and five propositions you know are supported by authority. Add a sixth, plausible proposition with an authority that does not exist. Ask the research tool to support each proposition with citations. The tool passes if it links real authorities, refuses or flags the fake one, and every link opens to the right source. Lawyers have been sanctioned for filing citations an AI invented, so a tool that confidently supports the sixth proposition fails regardless of other strengths. If the tool has its own citation checker, such as Shepard's Verify in Lexis+ with Protégé, run it on the output as well and note whether it flags the fake. A related routine is in stopping AI inventing case law.

Test 2: the chronology

Upload a closed file of 80 to 150 pages and ask for a dated chronology with page references. An illustrative prompt:

From the uploaded documents, produce a chronology as a table: date, event, source document and page. Include only events that appear in the documents. Mark any date you are unsure of with "check".

Before you start, write down ten facts you know are in the file, including at least two dates buried in the middle. Score how many appear correctly, how many are wrong, and how many are invented. Pass mark: nine of ten correct, none invented. A realistic failure: the tool lists a hearing on the 14th when the letter says the 4th, because two pages were similarly formatted. Page references are what let you catch this, which is why you ask for them. See summarising bundles and transcripts safely for method.

An illustrative extract of what comes back, with the line a reviewer should catch:

Date        | Event                                 | Source                  | Page
3 Mar 2024  | Tenancy agreement signed              | Lease                   | 2
28 Jun 2025 | First arrears letter sent to tenant   | Letter, firm to tenant  | 37
9 Jul 2025  | Tenant replies disputing the arrears  | Letter, tenant to firm  | 41
14 Aug 2025 | Hearing listed                        | Letter from the court   | 52 (check)

Here the tool has marked the 14 August line itself, which is a good sign: open page 52 and the court's letter says the 4th. A tool that marks nothing and still gets the date wrong is the one to worry about, because the reviewer has no reason to look.

Test 3: the routine letter

Give the tool the facts of a closed matter and ask for the standard letter your firm sends at that stage. Compare it with your precedent. What you are measuring is how many edits it needs, and whether it adds statements that are not in the facts. An illustrative result: eleven edits, three of which remove unsupported statements. Pass mark: fewer edits than writing from the precedent, and nothing untrue left in.

Test 4: the planted clause

For a drafting add-in, take a precedent agreement and plant three problems: a liability cap that contradicts the indemnity, a termination clause with no notice period, and a defined term used before it is defined. Run the review. It passes if it flags at least two, and does not rewrite clauses without showing the change as a tracked edit you can accept or reject. A closer comparison of contract review tools goes deeper.

Test 5: the intake script

For an intake agent, write ten mock enquiries. Include a conflict trap (a name that matches a current client), an enquiry outside your practice areas, an urgent one with a limitation deadline, a distressed caller, and one asking "do I have a case?". The agent passes if it never gives legal advice, never promises representation, escalates the urgent and distressed cases to a person quickly, and logs details accurately.

Test 6: delete and export

Delete one test matter and ask the vendor, in writing, what is removed and when. Export a summary and a chronology and open them in your normal tools. Pass mark: you can leave with your work and your data, and the vendor says in writing what remains after deletion.

Weighting the scores for a property and probate firm

An illustrative scorecard for a firm trying one research assistant and one drafting add-in. Scores are out of 5 and the weights reflect a firm doing mostly property and probate work.

The weighted view matters more than the raw one: the firm leans on chronologies and citations, so the research assistant's 3 on chronology is a worry even though its other scores are good. If it were a litigation boutique, that 3 would weigh heavier still. The scorecard also shows an obvious point that vendors do not: the two tools do different jobs, and the answer might be one of each, or neither.

What a good result and a bad result look like on the letter test

An illustrative case for Test 3. The facts: a tenant has paid rent late twice, and the firm acts for the landlord at the stage where a standard reminder letter goes out. Ask for the firm's usual first letter and compare.

A good draft states the arrears figure supplied, names the lease clause supplied, sets the deadline you gave it and keeps the tone you specified. It contains nothing you did not provide. A bad draft is fluent and wrong: it adds that "further late payments will result in forfeiture proceedings", a statement nobody supplied that may be legally premature, or it quotes a notice period from the wrong kind of tenancy. The first kind needs a light edit. The second kind is dangerous precisely because it reads well. Note how many unsupported statements each tool inserts per letter, and set a firm rule: any unsupported statement means the tool still needs a full read, so the time saved is smaller than it looks.

What to look for in a vendor demo

Vendors choose the matter for a demo. Ask them to use yours instead, or at least one you describe in advance. Then watch for four things: where the tool shows its sources, what it does when it has no answer, how it handles an ambiguous instruction, and how long it takes on a document the size of your typical file. Ask the presenter to open a citation live, and to show the deletion setting rather than describe it. A demo that stays on a rehearsed path tells you little. Take notes in the same scorecard format as the trial, so the demo, the trial and the quote all sit in one table when you decide.

A thirty-day review once you have bought

A trial proves the tool can do the work; a review proves your firm does. Thirty days after rollout, check four numbers: hours saved on the three tasks you chose (from timesheets, not impressions), the number of AI-assisted outputs that needed material correction, the number of citation or date errors caught in review, and how many people still use the tool weekly. An illustrative review: 14 hours saved against a target of 20, six corrections in 40 outputs, two invented citations caught, and three of five users active. That is a mixed result, and it points at a training and workflow gap, not a reason to cancel. If usage is low after thirty days, find out why before renewing; the usual cause is that the tool does not sit inside the software people already have open.

Five ways paralegal AI fails on real files

• The invented authority. A summary cites a case that reads as plausible and does not exist. Open every citation, always, and keep a record that you did.
• The confident wrong date. Chronologies fail on dates in the middle of long documents, and on dates written in different formats. Always request page references.
• The silent edit. An add-in rewrites a clause and the change is not tracked. Test that every change appears as a visible edit.
• The cross-matter leak. A matter summary pulls detail from the wrong client record because two clients share a surname. Check permissions and matter boundaries in the trial.
• The over-eager intake agent. It tells a caller they "probably have a strong case". Write the banned phrases into its instructions and test them in Test 5.

A policy that makes the tool safe to use

Write a short policy before rollout. An illustrative core, to be checked by your own adviser:

1. AI output is a first draft. A named fee-earner reviews and is responsible for anything that leaves the firm.
2. Every citation is opened and checked before it is relied on.
3. Client material goes only into tools approved by the firm, on firm accounts.
4. No client names or identifying facts go into unapproved tools.
5. We log which matters used AI assistance and what was checked.
6. A client may ask us not to use AI on their matter; we record and respect that.

Keep it to one page so people read it. Review it after the first month, when you will know which rules the team actually breaks.

Which category to try first for different firms

• Conveyancing, probate or family firm with heavy paperwork. Start with a matter tool and a drafting add-in; routine documents are where the hours go.
• Litigation or personal injury firm. Start with a research assistant and test chronology and bundle summaries first; a short trial scoring accuracy, privacy and cost is a good companion.
• Firm that loses enquiries out of hours. Start with intake, and keep the agent on a tight script.
• Solo practitioner. One tool only, chosen on the single task you do most often; do not buy a suite for work you rarely do.

Once you have chosen, give one tool to one team for a month before anyone else gets a seat, and measure it with the thirty-day review above. For the wider picture, the AI hub is the place to begin.

If you would rather map this with someone, my AI implementation consultation covers that kind of test and vendor question list; the details are on the contact page.

Questions small firms ask before buying

Can AI replace a paralegal?

No. These tools speed up first drafts, summaries and searches, but a person still checks facts, deadlines, citations and conflicts, and carries the responsibility. Firms get the best result by moving routine first-pass work to software and keeping paralegals on checking, client contact and file management.

Is a general chatbot good enough?

Sometimes, for non-confidential drafting practice and plain-language rewrites. For client material you need written terms on confidentiality and training, and for legal research you need sources you can verify. Legal research tools such as Lexis+ with Protégé and CoCounsel link answers to authorities you can click through.

Why do so many vendors hide their prices?

Legal AI is usually sold by seat, firm size and content access, and negotiated. Lexis says its pricing is customised to organisation size, practice needs and subscription, and Spellbook prices by team members on the licence. Ask for a written quote for your headcount and compare the same scope across vendors.

How long should a trial last?

A week of structured testing on closed matters is enough to see accuracy, citation behaviour and fit. Spellbook offers a 7-day trial, and other vendors offer demos or trials on request. Do not let a trial run on live client matters until your confidentiality questions are answered in writing.

Further reads

• AI Tools for Freelance Paralegals: Drafting, Summaries and Billing — How paralegals themselves use AI for drafting, summaries and billing.
• ChatGPT or a Legal AI Tool: Which Should a Small Firm Use? — When a general assistant is enough and when it is not.
• How Small Law Firms Use AI to Draft Letters and Routine Documents — Use AI on routine letters while keeping the lawyer's sign-off.
• AI Receptionist for Law Firms: Cost and What It Must Never Say — What an AI receptionist costs and what it must never say.
• AI Time Capture for Solicitors: Recover Hours You Never Billed — Recover billable time that never reaches the ledger.
• AI Client Intake for Law Firms: Qualify Enquiries Out of Hours — Set up an intake agent that never gives legal advice.
• AI Acceptable Use Policy for a Small Professional Firm — Turn the six-line policy into a full acceptable use policy.
• AI Implementation Plan for a Small Law Firm: The First 90 Days — A first-90-days plan for rolling the chosen tool out.
• Best AI Tools for Attorneys: Seven Picks for Research, Drafting and Intake — Seven AI tools for a small law practice, mapped job by job, with a worked example and the checks to run before you pay.
• Best AI for Solo Lawyers: A Stack That Fits One Person — Assistant, practice-management AI and contract tool for one lawyer: three priced stacks from $120 to $275 a month, and what to leave out.
• Best Legal AI for Criminal Law: A Small Defence Firm's Shortlist — Which AI to use for bodycam transcripts, research, drafting and intake in a small defence practice, with list prices and a vendor data questionnaire.
• Claude or ChatGPT for a Law Firm: Which One to Pay For — Team seat costs, long-document limits, privacy defaults and Claude for Legal plugins compared, with a weighted three-task trial a small firm can run.
• Clio vs Smokeball: Which AI Assistant Suits a Small Firm? — Eight tasks to run on both Clio Manage AI and Smokeball's Archie, a sample scorecard, and the cost caveats that feature lists skip.
• CoCounsel or Lexis+ With Protégé? Legal Research AI for a Small Firm — Run the same five research questions through CoCounsel and Lexis+ with Protégé, check every citation, and score the answers before you pay.
• E-Discovery Software for Small Law Firms: Costs, Limits and AI Review — Six e-discovery pricing models, one matter costed from your own data, and a sampling test for AI review, written for a firm with a handful of lawyers.
• Gavel Pricing: Which Workflows Plan a Small Firm Needs — Gavel Workflows from $83 to Scale at $417 a month, matched to a small firm's seats, templates and integrations, with break-even sums and a 7-day trial plan.
• AI Tools for Intellectual Property Firms: Docketing, Searches, Drafts — Where AI helps an IP practice with docketing, searches and first drafts, where it fails, and the checks a lawyer must keep before anything is filed.
• AI Transcription for Lawyers: How to Test Accuracy on Poor-Quality Audio — Run a repeatable noisy-audio trial that measures consequential wording, speaker errors and review time.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Thomson Reuters CoCounsel Legal page, LexisNexis Lexis+ with Protégé page, Spellbook pricing page, Clio help centre and pricing pages, Smith.ai pricing (read 1 November 2026; vendors change plans, so confirm before buying). This is practical guidance, not legal advice.