Paralegals and AI: How to Divide the Work in Your Firm

A small law firm should give AI bounded preparation work, give paralegals the evidence and accuracy checks, and keep legal judgement and final approval with a supervising lawyer. Paralegals must verify every material fact, quotation, citation, date and recipient against the matter file or an authoritative source before anyone relies on an AI-assisted draft.

The useful unit of delegation is a task, not a job title. “Draft a chronology from these approved documents” can be controlled and checked. “Handle this matter” cannot. The managing partner must decide which information may enter a tool, which outputs need a second review and who can send or file the finished work.

Follow me on Instagram@sagnikteaches

Give each stage an owner before anyone opens a tool

AI can rearrange text and surface patterns quickly. It does not owe duties to a client, know what has been withheld from its input, or accept responsibility when a persuasive sentence is wrong. A paralegal knows the file and can reconcile sources, but must not be made the unspoken final approver of legal advice. The lawyer assigns the task, states its limits and signs off where professional judgement is required.

Connect on LinkedInSagnik Bhattacharya

Formal Opinion 512 on generative AI tools, issued by a national bar association's ethics committee in July 2024, works through competence, protecting client information, client communication, supervising staff, advancing only meritorious claims, candour to the tribunal and reasonable fees. It is useful management guidance, although the firm's qualified lawyer must check the rules that actually govern each matter. NALA, the paralegal association, has published a piece for litigation paralegals that flags fabricated case citations, weak output from broad prompts and firms restricting public chat tools over privacy: the same failures this split is built to catch.

Subscribe on YouTube@sagnikteaches

The table is a starting operating policy, not a rule that every jurisdiction or practice area shares. A firm doing transactional work may use more document comparison; a disputes team may need tighter controls on pleadings and authorities. Put the authorised task, input boundary and approval point in the matter record so that the next person does not have to infer them from a chat history.

Start with material that can be traced back to a file

For the first pilot, choose a task whose output can be checked line by line: a document index, a chronology, or a neutral list of questions for a witness interview. These tasks still carry risk, but the paralegal can point to the actual document and page for each entry. Set a rule that the output has a source column. An unreferenced fact is a question to investigate, not a fact to paste into a letter.

For example, a paralegal receives twelve approved correspondence files and asks an authorised AI system for a dated list of communications. The illustrative output says “agreement accepted on 14 March”. The actual email says “we could accept subject to revised payment terms”. The paralegal changes the entry to “conditional response sent on 14 March” and links the email. That small verb difference may change the advice, so the lawyer sees both the first draft and corrected source.

A second example is a 46-page contract comparison. AI marks a changed termination notice period, but overlooks that the definition of “business day” also changed elsewhere. The paralegal checks the complete clause, definitions and schedules, then reports both changes together. A highlighted difference is only a search aid; the lawyer decides the commercial and legal effect. For a deeper drafting workflow, see AI-assisted routine letters in a small law firm.

Before any upload, the firm should classify the input. Public template text, anonymised training material and live client documents are different risk categories. Approval for one category does not imply approval for another. Check the vendor terms, retention, access controls and matter permissions; decide whether client instruction or consent is needed. A document can be confidential even when the name has been removed, because a rare transaction detail can identify the client.

A one-matter budget shows where review time belongs

Consider an illustrative matter with 18 source documents, 64 pages in total and a four-day internal target for a first advice draft. The managing lawyer allocates 20 minutes to define the question and authorised sources. A paralegal spends 35 minutes cleaning the file names and removing duplicates, 25 minutes getting an AI-generated chronology and issue list, and 95 minutes checking each proposed item against originals. The lawyer spends 70 minutes evaluating the relevant law, client objectives and omissions, then 25 minutes reviewing the final letter.

That is 270 minutes of recorded human work, plus tool processing. The example is not a promised saving. It shows why a pilot must measure the full chain rather than only the 25-minute generation step. If manual checking takes longer than preparing a chronology from scratch, the firm may still gain a more complete source map, but it should not claim a time benefit. If the AI invents several events, stop and redesign the input or abandon that use.

The paralegal logs 18 source files, 21 suggested chronology entries, three corrected dates, two unsupported assertions removed and one question escalated. The lawyer can see exactly which content changed. After five similar matters, compare actual time, correction count and rework with a baseline prepared without AI. Keep matter complexity visible; a short file and a disputed bundle should not be averaged into a misleading productivity figure. Measuring time saved in a small firm goes deeper on that calculation.

The paralegal's review is an evidence exercise

Give the paralegal a checklist that requires affirmative marks rather than “looks fine”. The first check is input provenance: were all the files authorised, complete and from the correct matter? The second is output coverage: did the model omit an important document or page? The third is assertion checking: can each named party, sum, date and event be located in an original? The fourth is authority checking: is every law, quotation and citation found in a trusted source and still relevant? The fifth is communication checking: does the draft match the client's instructions and the lawyer's scope?

• Identity: compare client, counterparty, matter number, document version and recipient with the case-management record.
• Numbers: recalculate totals and date intervals independently; inspect currency, units and rounding.
• Source: record the original file and page beside each substantive claim, not just a model-generated footnote.
• Authority: open every cited decision or provision in an approved source; read the passage around the quotation.
• Missing material: ask what was excluded from the tool and whether that exclusion could reverse the draft's conclusion.
• Escalation: mark legal interpretation, disputed facts, privilege questions and any adverse authority for the lawyer.
• Release: confirm the required lawyer approval, final version, attachments and delivery channel before sending.

Suppose the AI converts “payment within thirty days after acceptance” into “payment within thirty days of delivery”. The paralegal catches it by comparing the signed agreement, not by asking the same model to check itself. The discrepancy is escalated because “acceptance” may be contested. This illustrates why a second AI pass can help locate an error but cannot replace an independent source check.

A citation deserves more than a search-result snippet. Check that the authority exists, the quotation is exact, the proposition is supported in context and the authority applies to the matter. Also look for a relevant later change or contrary source. If the research question is outside the paralegal's competence or the source is ambiguous, mark it for the lawyer instead of resolving it with a confident paraphrase. The related case-law verification routine covers this specific failure in more depth.

Make supervision visible at three hand-off points

At assignment, the lawyer names the question, permitted material, prohibited uses and expected output. “Identify clauses changed between these two approved drafts; do not interpret enforceability” is a workable instruction. “Review the contract” is too broad. The lawyer also sets the review tier: clerical summary, client-facing draft or high-stakes legal position. This prevents a paralegal from having to decide after the fact how much checking was intended.

During the work, the paralegal flags uncertainty in a short exception note. For example: “Clause 7 changed from 14 to 30 days; the earlier signed schedule may override it; no conclusion drafted.” The lawyer can then decide whether to ask for more documents or write the interpretation personally. An exception note is more useful than a long AI transcript because it identifies the source and the decision needed.

At release, the lawyer reviews the substantive position, not just a clean PDF. The approval record should identify the final version and any unresolved qualifications. If a client asks a new question in reply, that is a new assignment rather than an invitation for the AI or paralegal to extend the original advice. Set a rule that urgent deadlines, adverse facts, settlement terms and filing language always return to the responsible lawyer.

Keep the hand-off note short enough that people complete it. A filled illustrative record could read: “Matter 24; task: chronology from approved bundle version 3; AI output: 19 entries; paralegal checked 19 against originals, corrected two dates and removed one unsupported event; open issue: whether the later email replaced the signed schedule; lawyer to decide; no client message sent.” Add the checker and approver initials, document version and approval time. This is a process record, not a substitute for the underlying evidence. If a later correction arrives, the team can identify which version was relied on and which documents need rechecking.

Do not make the paralegal responsible for proving the absence of every possible omission. Give them an agreed source list and a scope statement. For example, if only the client's correspondence was supplied, the note should say “other party's correspondence not reviewed”. The lawyer can then decide whether the limited chronology is fit for the current purpose. This avoids a common failure in which a tidy output looks complete because nobody recorded what it did not contain.

A practical weekly review can sample low-risk work as well as all high-risk outputs. If the partner finds a recurring error, such as AI confusing draft and executed versions, change the intake step and prompt. Do not merely tell paralegals to be more careful. A supervision system should make the failure less likely and easier to detect next time. The firm-wide boundary for legal tasks that stay human is a useful companion when setting those red lines.

Use a source-bound prompt and require an exception list

For a document-index task, the managing lawyer can approve a reusable instruction. It should say what the AI may use, what to output and what to do when the answer is missing. The prompt below is illustrative and must be used only in an approved environment with authorised material. It deliberately asks for uncertainty rather than a polished narrative.

Task: prepare a chronology from the attached approved matter documents.
Use only the supplied files. For each entry give the date, neutral event,
file name and page or message identifier. Mark a disputed or inferred date
as UNCERTAIN. Do not infer legal effect, cite law, or draft advice.
Finish with a separate list of missing documents and conflicting accounts.

An illustrative output might read: “14 March — revised terms proposed — email-07, message 3 — UNCERTAIN whether accepted.” That is useful because it preserves the uncertainty. A defective output would say “14 March — agreement concluded”, citing the same message. The paralegal should correct the entry, save the source and flag the legal question. If the model cannot provide reliable file identifiers, use manual indexing instead.

A different task needs a different constraint. For a client-status message, provide only lawyer-approved milestones and ask for a neutral draft that does not predict the outcome. An illustrative result, “Your documents are being reviewed; we expect to update you after the scheduled call,” may still be wrong if no call is scheduled. The paralegal checks the diary and the lawyer approves any sentence that discusses prospects or next legal steps. Client status updates with AI examines that narrow workflow.

Put confidentiality and fees into the operating rule

A firm should maintain an approved-tool list and prohibit copying client information into unapproved personal accounts. For each approved tool, record who has access, whether prompts or uploaded files are retained, how data is used, whether a matter can be deleted and how access is revoked when a staff member leaves. Test the controls with a dummy matter before permitting live data. Check the governing professional and client obligations with a qualified lawyer; a generic software privacy statement is not enough.

For example, a paralegal preparing a summary discovers that a document contains details from another client's negotiation in a forwarded chain. The right action is to stop, separate the material and ask the supervising lawyer what may be processed. Removing the visible client name is not a sufficient control when the quoted transaction terms are distinctive. Record the decision so the team does not repeat the mistake on a later upload.

Time recording also needs a rule. Record the real minutes spent preparing inputs, checking outputs and obtaining sign-off. Do not bill the time AI appeared to save as if a person had spent it, and do not hide substantial review time behind an “AI task” label. The lawyer decides what is reasonable to charge under the engagement terms and applicable rules. A team member should be able to reconstruct the work from the matter record without exposing unnecessary prompt content in an invoice.

Decide whether the split is working after a small pilot

Choose one repeatable task and five matters with comparable documents. Before the pilot, record the ordinary process time and typical error types. For each AI-assisted matter, capture input preparation, generation, paralegal checking, lawyer review, rework and any client correction. Use the same output standard in both groups. The primary measure is a reliable final work product; elapsed time alone is a poor measure if the lawyer must later repair hidden errors.

A simple stop condition is useful. If an output includes an invented authority, crosses a matter permission boundary, or repeats a material factual error after the workflow has been corrected, suspend that use and review the tool and task definition. For minor format errors, revise the template and test again. The partner, not the software vendor, decides which errors are tolerable for the firm's work.

Consider a final illustrative check: five chronology pilots produce 104 entries. Paralegals correct eight dates, remove four unsupported events and flag six genuinely disputed entries. The responsible lawyer accepts the corrected chronologies but finds that two source links point to the wrong version. The firm revises file naming and requires version IDs in the output before expanding to drafting. That is a sounder decision than announcing “AI saved hours” on the strength of a fast first pass.

The task split should be written on one page and revisited when the tool, client instructions or matter type changes. Make it easy for a paralegal to say “this needs a lawyer” without being treated as a failed automation. The strongest use of AI here is to bring a well-organised, source-backed question to the professional who is accountable for the answer.

Questions managing partners ask about the split

Can a paralegal send an AI-drafted client letter?

Only under the firm's ordinary delegation and supervision rules. The paralegal should check the matter facts, sources, recipient, attachments and tone, then route advice or a material legal position to the responsible lawyer. A polished draft is not evidence that its law or facts are right. Record who approved the final version before it leaves the firm.

Should every AI-assisted document receive the same level of review?

No. Use a risk tier set by the responsible lawyer. A meeting agenda compiled from approved notes may need a factual check; a pleading, settlement proposal or advice letter needs independent legal review and formal approval. Confidentiality and access checks apply to every tier, including seemingly harmless summaries.

What if the AI tool gives a citation without a source link?

Treat it as an unverified lead, not an authority. Find the primary source in an approved legal research service, read the relevant passage and check that it remains applicable to the matter. If the source cannot be found or supported, remove the point and tell the supervising lawyer. Never insert a citation merely because it looks plausible.

Further reads

• AI Implementation Plan for a Small Law Firm: The First 90 Days — Plan the wider firm rollout after the task split.
• AI Client Intake for Law Firms: Qualify Enquiries Out of Hours — Set boundaries for the first client enquiry.
• AI Time Capture for Solicitors: Recover Hours You Never Billed — Record actual human work accurately.
• AI Acceptable Use Policy for a Small Professional Firm — Turn approval rules into a written policy.
• Per-Seat or Pay-As-You-Go? Legal AI Pricing for Small Firms — Compare licences after defining the workflow.
• AI Paralegal Software: What a Small Firm Should Test Before Paying — Three kinds of AI paralegal software, a six-test trial run on your closed matters, and the questions to ask when pricing is quote-only.
• AI Legal Brief Writing: A Draft and Cite-Check Routine — A routine for small litigation firms: build an authority table, draft only from it, cite-check each source line by line, and meet court AI-disclosure rules.
• AI Tools for Freelance Paralegals: Drafting, Summaries and Billing — Three AI workflows for an independent paralegal, with a permission note for the supervising lawyer, confidentiality tiers and disclosure wording.
• AI for Personal Injury Lawyers: Seven Jobs to Hand Over First — Seven injury-matter workflows with sample outputs, review gates and a clear boundary around legal judgement and confidential records.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Formal Opinion 512, Generative Artificial Intelligence Tools (bar association ethics committee, 29 July 2024; archived copy read 4 November 2026); NALA Facts and Findings, AI for Litigation Paralegals: Managing Risk to Increase Your Value (Q4 2024, read 4 November 2026).