Microsoft Copilot for Lawyers: How a Small Firm Should Trial It in Word and Outlook

Microsoft Copilot can summarise a long email thread in Outlook, draft a routine client letter in Word from your own notes, and condense a lengthy document into issues and dates. It falls short on legal accuracy, authority and anything needing judgement. Test it on five of your own tasks, with a stopwatch, before buying licences.

One premise to settle first: the question here is the general Copilot in the Microsoft 365 apps a firm already uses. On 15 October 2026 Microsoft launched Copilot for Legal, a set of legal skills delivered as two plugins, one for Copilot in Word and one for Copilot Cowork, and Microsoft Learn says the Word plugin is switched on by default in Word desktop for Microsoft 365 Copilot commercial subscribers. So a paid licence may bring legal skills into your test whether you planned it or not; what they add is covered in Copilot for Legal and whether you need a paid licence. Here the question is simpler: does the ordinary Copilot earn its place in a small firm's Word and Outlook?

Follow me on Instagram@sagnikteaches

Which Copilot you actually have: free Chat or the paid licence

Most small firms on a Microsoft 365 business plan already have Copilot Chat (Basic) at no extra cost. It is grounded in web data and can work on an email you have open in Outlook or Teams, or a file you upload, but Microsoft Learn says it does not include Copilot in Word, Excel, PowerPoint or OneNote. Learn also describes a middle experience, Microsoft 365 Copilot (Basic), for users without the add-on who still get standard access to Copilot inside Word and Outlook, such as drafting, rewriting and thread summaries, subject to service capacity. The paid add-on (Microsoft 365 Copilot Premium on Learn, Microsoft 365 Copilot Business on the price page) adds priority access and grounding across your emails, meetings and files. Microsoft shows an in-product label for each experience, so check what your own Word and Outlook display before deciding what to test. The apps have also been renamed (the Microsoft 365 Copilot app is now Microsoft Copilot), so expect both names.

Connect on LinkedInSagnik Bhattacharya

The $21 price is an annual subscription; $25.20 buys a month-to-month one. The one-month free trial converts to paid automatically, so put the cancellation date in the diary on day one; once a paid subscription starts, Microsoft gives a 7-day window to cancel for a prorated refund. A promotional $18 applies to the first year of an annual add-on bought between 1 July and 31 December 2026, though not to trials, and some advanced features carry usage-based charges on top of the licence. Newly assigned licences may not show up in the apps straight away, so don't plan the first test for the same afternoon. For the wider pricing picture, see Copilot pricing for business versus enterprise.

Subscribe on YouTube@sagnikteaches

Tenant and data settings to check before anyone types a prompt

Microsoft's data page says prompts, responses and data accessed through Microsoft Graph are not used to train the foundation models, and that Copilot only shows organisational data the individual user has at least view permission for. For a law firm the second sentence matters more than the first. It means Copilot is as careful as your permissions, no more.

• Permissions: check who can open each matter folder in SharePoint or OneDrive. Copilot makes over-broad sharing visible quickly, and "Everyone except external users" on a document library is the usual culprit.
• Sensitivity labels: Microsoft says Copilot honours usage rights on content encrypted with Microsoft Purview sensitivity labels or restricted permissions. Use labels on your most confidential material.
• Retention and history: Copilot interactions are stored; admins can use Purview content search and set retention policies, and users can delete their own activity history from their account page.
• Web search: Copilot can send a search query to Bing for web grounding. Decide whether staff may use it on client matters.
• Connected experiences: if your organisation turns off connected experiences that analyse content, Copilot features will not be available in Word, Outlook, Excel, PowerPoint or OneNote, so a locked-down setup can switch it off by accident.
• Third-party models: admins can decide whether Anthropic or OpenAI models are used in Copilot experiences, and Microsoft notes that Anthropic models are currently excluded from the EU Data Boundary, which matters if you have EU clients.

Microsoft's admin centre has a Copilot usage report and a readiness report under Reports, Usage, Microsoft Copilot. The readiness report checks licences and update channels, not sharing permissions, so a green tick there says nothing about oversharing. Decide before the test which matters stay off limits entirely, whatever the report says.

Three legal workflows worth testing

Workflow one: the thread summary in Outlook

A client conversation has run to 28 messages. Open the thread and ask for a summary of decisions, open questions and dates. What good looks like: a short list that matches what you would have written. What goes wrong: it flattens a conditional ("subject to the lender's approval") into a statement, and it can misattribute who proposed what. Always scan the original for dates and conditions.

Workflow two: the first draft of a routine letter in Word

You give Copilot your attendance note and ask for a client update letter in your firm's tone, under 250 words, plain language. Point it at a letter you wrote before as a model. What good looks like: structure and tone are right in one pass. What goes wrong: it adds reassuring sentences you did not authorise, such as "we expect completion shortly". Strike every forward-looking promise. The drafting side is covered more generally in how small firms use AI to draft letters.

Workflow three: the document digest

Open a 40-page agreement and ask for a table of parties, term, payment obligations, termination triggers and anything unusual. What good looks like: a navigable map that sends you to the right page. What goes wrong: a missed clause or a wrongly confident reading of a defined term. Treat the digest as an index, never as a substitute for reading the operative clauses.

The small-firm task test: five tasks, a stopwatch, a score

Run the same five tasks twice, once by hand, once with Copilot, using real but non-sensitive or consented material. Time each, then score the Copilot output from 0 to 3 for accuracy and 0 to 3 for how much you had to edit. Add the minutes and subtract the minutes it took to check.

Task test sheet (illustrative)
Task                          | By hand (min) | With Copilot (min, incl. check) | Accuracy 0-3 | Edit load 0-3
1 Summarise 28-message thread | 15            | 6                               | 2            | 1
2 Client update letter        | 20            | 11                              | 3            | 2
3 40-page agreement digest    | 45            | 25                              | 2            | 2
4 Reply to a routine enquiry  | 8             | 4                               | 3            | 1
5 Meeting note to action list | 12            | 7                               | 3            | 1
Total minutes saved: 47. Any task scoring 0 or 1 on accuracy is a stop.

The main worked example: a four-lawyer firm

An illustration with invented figures. A four-lawyer firm giving property and wills advice gives licences to two lawyers and the practice manager for four weeks. They run the five tasks above, each repeated across real work: about 20 thread summaries, 12 routine letters, 3 agreement digests, 40 routine replies and 10 meeting notes across the month.

Using the test-sheet times, they estimate savings per item of 9 minutes for a thread, 9 for a letter, 20 for a digest, 4 for a reply and 5 for a note. That is 180 plus 108 plus 60 plus 160 plus 50, a total of 558 minutes, or about 9.3 hours across three people in a month. At a blended cost of $60 an hour for time that would otherwise be spent, the value is about $558. Three licences at $21 cost $63 a month. Even if they only trust half the saving, the licence cost is small beside the time.

Now the check that matters. In the same month, they log accuracy slips: Copilot misread a condition in two thread summaries and added a promise to one letter. All three were caught on review. The honest conclusion is that the licences pay for themselves only if the review step is non-negotiable, so they write it into the firm's rules. How to measure the payback properly is in how to measure whether Copilot is paying for itself.

Where Copilot falls short for lawyers

Several limits show up in every test. It is not a research tool: it does not replace a legal database and it can state a rule or a case with confidence it has not earned, so any authority it mentions must be checked at source, as in how to stop AI inventing case law. It does not know your matter, only what is in the documents and mail it can reach, so an incomplete file gives an incomplete answer. It does not exercise judgement about privilege, conflicts or limitation periods. And it cannot tell which version of a precedent is current unless you point it at the right one.

A subtler limit is the permissions mirror. If an assistant in the firm can open a file, Copilot can summarise it for them. That is correct behaviour and still a problem where an ethical wall relies on people simply not looking. Move the wall into permissions before the test starts.

The cost sums for a four-lawyer firm

Put the licence options side by side for the same illustrative firm: four lawyers and one practice manager, five people. Everyone keeps free Copilot Chat on the existing plan, at no extra cost. Buying paid Copilot Business for all five at $21 a month on an annual commitment is $105 a month, or $1,260 a year. Buying for the three heaviest users is $63 a month, or $756 a year. If the firm is moving from Business Basic ($7) to Business Standard ($14) anyway, the bundle at $23.50 per user is cheaper than buying the plan and the add-on separately, but the bundle only makes sense for people who need the desktop apps.

The break-even is easy to state. At $60 an hour of internal time, a $21 licence pays for itself if it saves 21 minutes a month. The firm in the example estimated about 3 hours per user per month, which is far above the break-even, so cost is not the real risk. The real risks are accuracy and confidentiality, which is why this tutorial spends more time on the test and the settings than on the price. A cheap tool used carelessly costs more than an expensive tool used well.

A realistic mistake and how it showed up

Here is the kind of slip a four-week test is designed to catch. A lawyer asks Copilot in Outlook to summarise the thread with a client about a boundary dispute, and the summary says the neighbour "has agreed to remove the fence". In the original thread the neighbour's solicitor wrote that the neighbour "may be willing to consider removing the fence if costs are shared". The summary turned a possibility into a decision. It was caught because the lawyer's rule was to check every verb of agreement against the source; without that rule, the next letter would have told the client the fence was coming down.

The fix was in the prompt, not in the lawyer's vigilance alone. The firm changed the Outlook prompt to say "Quote any statement of agreement or intention exactly, and say whether it is conditional". Output quality improved and the review step got faster, because the quotes sat beside the summary and could be compared in seconds. It is a good example of the loop to run all month: find a failure, add a rule to the prompt, re-run, record whether the failure recurs.

Before and after: a routine reply in Outlook

Before, a lawyer types a four-line reply to a client asking when documents will be ready, then spends two minutes making it polite and complete. After, the prompt is "Draft a reply to this email confirming we need the signed form and ID copies by Friday, and that we will send the draft contract on Monday. Polite, three sentences." Copilot returns a version with a friendly opening and the two dates. The lawyer corrects "ID copies" to the firm's usual phrase and adds the client's name. Total time falls from about 8 minutes to about 4, which matches the test sheet, and the output scored 3 for accuracy because every fact came from the prompt.

A four-week rollout that keeps the test honest

Run the trial in four short steps. In week one, fix permissions on matter folders, assign licences, and wait for Copilot to appear in each tester's apps, which is not always immediate. Give each tester the three prompts above and ask them not to improvise yet. In week two, run the five-task test and fill the sheet in as they go. In week three, add one task each that the tester chooses, so you see what people reach for when nobody tells them. In week four, review: total minutes saved, accuracy slips, anything surprising Copilot revealed from a shared folder, and whether anyone stopped using it.

Do not skip the last item. Usage that falls to nothing by week three is information, usually that the prompts were awkward or the task was not worth the habit. Pull the usage report from the admin centre and compare it with what testers say; the two often disagree, and the report is the more reliable of the two.

How the answer differs for a one-lawyer practice

A sole practitioner has no team to coordinate, which makes the test quicker, but also removes the safety net of a second reader. The sole lawyer should lean harder on the review rules and keep a short written note of what Copilot is used for, which doubles as evidence of care if a client or insurer asks. Permissions are simpler, because one person owns every folder, though the same tidy-up applies to any shared folders with an outside bookkeeper or locum. The licence maths is also kinder: one licence at $21 a month is a small experiment. The $21 price needs an annual commitment, while $25.20 buys a month-to-month subscription, which suits a one-person experiment; or use the one-month trial and set a reminder to cancel or convert before it rolls into a paid term.

Prompts that work for a first week

Outlook thread: "Summarise this conversation as: decisions made, open questions, dates mentioned. Quote any condition exactly as written. Do not add anything not in the thread."

Word letter: "Using my attendance note below and the style of the attached letter, draft a client update under 250 words. Plain language. No predictions about timing. Mark anything you are unsure of with [CHECK]."

Word digest: "List the parties, term, payment obligations, termination rights and any unusual clauses, each with the clause number. Say 'not found' rather than guessing."

Each prompt includes a rule that forbids invention and a way for the tool to admit doubt. That pairing, no additions plus an explicit [CHECK] marker, is the most effective defence against the common failure of fluent, wrong text.

Questions to put to your own IT adviser

If an outside adviser looks after the firm's Microsoft 365 tenant, send them a short list before the trial and ask for written answers. Which sharing links exist on matter folders, and which are open to the whole organisation? Are sensitivity labels in use, and on which libraries? Is web search for Copilot allowed, and is that decision recorded? Which update channel are the desktop apps on, since Copilot features depend on current versions? Who can see Copilot activity history through Purview, and for how long is it retained?

The answers do two jobs. They tell you whether the tenant is ready, and they give you a dated record of the care taken, which is useful if a client, a regulator or an insurer later asks how the firm decided to use AI. Keep the answers with the test sheet and the four-week review in one folder, and add the date of the next review so that the decision is revisited when the product or your matters change.

Deciding after four weeks

Set three conditions before the test starts. First, a minimum saving, for example 30 minutes per person per week. Second, a maximum for serious accuracy slips that reached a client, which should be zero. Third, a ceiling on the admin burden of permissions clean-up. If all three hold, buy for those who use it, review quarterly, and use the Copilot usage report to see who is active over 7, 28, 90 or 180 days; licences nobody uses are the easiest saving to find. If the second condition fails, stop and fix the process before spending more.

If the firm would rather compare Copilot with a legal-specific product, read ChatGPT or a legal AI tool, and keep in mind that legal-specific tools bring their own prices and terms. This tutorial is general information, not legal or regulatory advice; check your professional body's guidance on technology and confidentiality before using any AI tool on client work.

Copilot questions small law firms ask

Is free Copilot Chat enough for a small firm?

For some tasks, yes. Copilot Chat comes with eligible business plans and can work on an email you have open in Outlook or a file you upload, but it has no Copilot inside Word. Some users also get standard access to Copilot in Word and Outlook without the add-on. The paid licence adds grounding across mail, meetings and files. Test both before paying.

Does Copilot train on our client documents?

Microsoft states that prompts, responses and data accessed through Microsoft Graph are not used to train the foundation models behind Copilot. That is a statement about training, not about confidentiality duties, which still depend on your permissions, settings and your regulator's expectations.

Why does Copilot show documents some staff should not see?

Copilot only surfaces what the person can already open. If SharePoint permissions are loose, Copilot makes loose permissions visible fast. Tidy matter-folder access before rolling it out, and use sensitivity labels for the most confidential files.

How many licences should we start with?

Two or three, held by people who do different kinds of work, for a fixed four-week test. Copilot Business lists at $21 per user a month on an annual commitment and has a 300-seat cap, so a small trial costs little. Keep the test short and measured.

Further reads

• Per-Seat or Pay-As-You-Go? Legal AI Pricing for Small Firms — How per-seat legal tools compare with a general licence.
• AI Readiness Checklist for Accountants, Solicitors, Consultants — A wider readiness check before any AI rollout.
• AI Implementation Plan for a Small Law Firm: The First 90 Days — A 90-day plan once you have chosen a tool.
• How Small Law Firms Use AI to Answer Client Status Questions — A natural next Outlook workflow to test.
• AI Time Capture for Solicitors: Recover Hours You Never Billed — Where else recovered time shows up in a firm.
• How to Set Up Human Review for AI Work Without Slowing Down — A review routine that does not slow the firm down.
• Spellbook Pricing: How a Small Firm Should Read the Quote — Spellbook is quote-only with a 7-day free trial. How a small firm can read the quote beside Copilot and Gavel Exec, line by line, and test all three on one NDA.
• Copilot for Insurance Agents: Outlook and Word Tasks for the Agency — Email triage, client letters from policy wording and meeting recaps: which Copilot tasks an insurance agency can hand over, and at what cost per seat.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Microsoft Learn (Copilot overview and licence comparison; data, privacy and security; Manage Copilot for legal; Copilot usage and readiness reports), Microsoft Tech Community announcement of Copilot for Legal (15 October 2026), Microsoft 365 Copilot Business and Microsoft 365 Business Standard pricing pages.