Best AI for a Law Office: Filing, Diary Dates and Billing Entries

A small law office can use Microsoft Copilot, ChatGPT or the AI add-on inside its practice-management system for filing, diary dates and billing entries, provided no client file reaches a tool that trains on it. Pick by job: Copilot for files already in Microsoft 365, the practice system's own assistant for billing, ChatGPT for generic templates only.

Office administration is a different question from legal drafting or research, which carry separate risks covered in which legal tasks to keep away from AI. Admin is the safer place to start because the output is checked against a document you already hold, and because a mistake shows up quickly. It is not risk-free, though, because file names, matter descriptions and time narratives all contain client information.

Follow me on Instagram@sagnikteaches

Start with the confidentiality check, before any tool

Every tool below is judged on one question first: where does the client's information go once you paste it, and is it used to train a model? Answer it in writing, once per tool, and keep the answer with your office procedures. A five-line check is enough.

Connect on LinkedInSagnik Bhattacharya

1. Which plan will staff actually use? The free and individual plans of ChatGPT and Claude let a user switch off use of their chats for training, but it is a setting each person must remember. Business plans (ChatGPT Business, Claude Team, Microsoft 365 Copilot, Gemini in Workspace) do not train on business content by default. Both ChatGPT Business and Claude Team need at least two seats.
2. Is the tool inside a system you already trust? A Copilot that reads files in your existing Microsoft 365 tenant keeps data within the permissions and settings you already administer. A separate website does not.
3. Who can see the output? Shared Projects and shared chats expose content to colleagues. If a trainee should not see a conflicted matter, do not paste it anywhere shared.
4. Does the client's engagement letter or your regulator's guidance restrict it? Some corporate clients prohibit any AI processing of their material. Ask, and record the answer on the file.
5. Can you turn history off, delete, or export? You need to be able to remove a conversation containing client details.

If you want the long version of this question for a solicitor's practice, can solicitors use ChatGPT without breaching confidentiality goes through it. The rest of this tutorial assumes you have chosen a plan that passes those five lines.

Subscribe on YouTube@sagnikteaches

The three admin jobs and the tools that fit each

Copilot Chat, the free web-grounded assistant, is included with Microsoft 365 business plans at no extra cost and now works on Outlook mail and open files. The paid Microsoft 365 Copilot licence adds reasoning across your emails, meetings and files together; on Microsoft's page on 4 November 2026, the Microsoft 365 Copilot Business add-on was $21 per user per month on annual billing ($25.20 billed monthly) for organisations up to 300 users, with a first-year promotional price of $18 on annual plans bought by 31 December 2026, and some advanced features billed by usage on top. A practice-management assistant, such as Manage AI in Clio, is priced and packaged by the vendor, and Clio's pricing page would not load for this check, so ask for the current terms before you plan around one. Prices move, so confirm all of these on the vendor pages the week you decide.

Clio's help centre gives a useful picture of what a practice-system assistant does for admin. Its AI features can extract events from documents and suggest them for a person to review, suggest time entries, help prepare invoices and draft client communications. They follow each user's existing permissions, so nobody can surface a matter they are not allowed to open. Clio says the data is not used to train external AI models, and in the unified Clio interface an administrator can switch AI off for the whole firm under Manage settings, Security & Compliance, AI features. If your firm uses a different system, ask its vendor the same four questions: what it extracts, what it suggests, who can see the result and whether it trains on your data.

Filing: let AI propose, let the system file

Filing errors in a law office are quiet and expensive: a letter saved to the wrong matter is hard to find months later and can be sent to the wrong person. AI helps most with three things: consistent names, a first sort of incoming mail, and a one-line summary on each document.

Step 1: agree a naming convention, with AI as the sounding board

You can do this without a single client detail. Ask:

I run a small law office. Propose a file-naming convention for documents saved to a matter folder. Requirements: sorts chronologically, shows document type, shows author or sender, works in a folder listing without opening the file, under 80 characters. Give five example names using made-up details, and note any ambiguities.

An illustrative answer: 2026-11-04 Letter In - Other Side - Request for disclosure.pdf. Accept or adjust, then write the rule on one page and pin it in the office. The point is that the convention is human-agreed, and the tool never sees a real file.

Step 2: triage incoming post inside an approved tool

Say a four-solicitor conveyancing and family practice receives 60 documents a day. In an approved tool (for instance Copilot reading the scanned PDFs in your own SharePoint), ask for a table: document type, sender, matter reference if present, action needed, date mentioned. Do not let it file anything. A clerk reads the table, corrects it and files. Over a week, track mismatches: if one document in twenty gets the wrong matter, the problem is usually that the sender used a different reference format, and it is fixed by a reference-format rule, not a better prompt.

Step 3: a one-line summary per document

A line such as "Defendant's response to request for information, 14 pages, admits items 1 to 3, disputes 4" saves minutes every time someone opens the file. Ask for no more than 20 words, no legal opinion, and "say 'unclear' if you cannot tell". Check three in ten for the first fortnight. The usual failure is a confident summary of a scanned document the tool could not read properly, so look for the "unclear" instruction being ignored.

Diary dates: extract with AI, calculate and confirm with a person

Dates in letters, orders and notices are where the temptation to trust the tool is strongest and the downside largest. Keep one hard rule: AI never creates a diary entry that no person has checked.

The safe pattern has four steps. First, the tool extracts every date and obligation from the document, quoting the sentence it came from. Second, a fee earner or secretary checks the quote against the document. Third, a person calculates the actual deadline using the applicable rules, counting business days and any exclusions. Fourth, the date is entered in the diary system and a second person confirms it. The tool contributes only to step one.

An illustrative extraction prompt:

Read the document below. List every date, time limit and action it requires, one per row, with columns: quoted sentence, date or period stated, who must act, and whether the period runs from a date mentioned elsewhere. Do not calculate deadlines. If a period runs from an event whose date you cannot see, write "needs date of [event]".

An illustrative result for a letter saying "please respond within 21 days of service": the row reads "respond within 21 days of service" | 21 days | recipient | runs from service: needs date of service. That is the right output, because the tool has declined to guess. The failure mode to watch for is the model producing a calendar date by itself. If it does, discard the row, because it may have counted from the letter's date rather than the service date.

Reminders and the diary system

Once a person has confirmed a deadline, set three reminders: seven days before, two days before, and the day itself. A workflow tool or your practice system can send them. This is automation of a confirmed fact, which is the right place for it. If you are curious how far to take this, can AI manage your calendar shows what scheduling assistants can and cannot do.

Billing entries: from rough notes to accurate narratives

Billing is where AI saves the most fee-earner time, because narratives take effort that nobody values. The risk is just as clear: an entry the model made up, or one that pads a vague task, can embarrass the firm in a fee dispute. AI should reformat what the fee earner has recorded and never add work to it.

Say a fee earner at a six-lawyer practice jots down after a long afternoon: "call opp re disclosure 12m, review client docs 40m, email client advice re settlement 18m". A prompt that works:

Rewrite each line as a billing narrative in the form: [action] + [subject] + [outcome or purpose]. Keep the times unchanged. Use plain language a client can follow. Do not add any work, outcome or detail that is not in my note. If a line is too vague, write "[needs detail]" instead.

An illustrative output:

0.2 hrs  Telephone call with other side's solicitor regarding disclosure
0.7 hrs  Reviewing documents supplied by client
0.3 hrs  Email to client advising on settlement position

Notice the quick sum: 12 minutes is 0.2 of an hour on a six-minute unit, 40 minutes becomes 0.7 (rounded from 0.67) and 18 minutes is 0.3. Check rounding against your own billing rules before anything is posted, since the model may round differently from your unit. This is a good habit: ask it to show minutes alongside units for the first fortnight, then drop the minutes.

Before and after: a vague entry

Before: "Work on file, 1.5 hrs." After, once the fee earner has added three words of detail and the prompt has run: "Preparing chronology of events from client's emails (1.0 hr); updating issues list (0.5 hr)." The tool did not invent anything; it asked the fee earner for detail by returning "[needs detail]". That is the behaviour to keep.

Time capture from emails and documents

Some practice systems and Microsoft 365 tools can suggest entries from calendar items, emails and documents opened. These suggestions are a prompt to remember, not a record. The fee earner must confirm each. For more on recovering unbilled work, see AI time capture for solicitors.

Which tool to choose, by office size

Illustrations only. A sole practitioner who works in Microsoft 365 already can start with Copilot Chat on the existing plan at no extra cost, and add the paid Copilot licence only if file and email reasoning is wanted. A three-solicitor firm that uses a practice-management system should ask its vendor what the built-in assistant does for billing and filing before buying anything else, since the integration saves copying and pasting. A ten-person firm with a mix of tools may need a policy more than a product: one approved assistant, one written rule on what may be entered, and one person who owns the settings. Pricing models differ, and per-seat or pay-as-you-go legal AI pricing explains how to compare them.

A word on ChatGPT and Claude. Both are good general assistants, with business plans that do not train on your content by default. They sit outside your case files, so they suit generic work: designing naming conventions, writing a checklist for opening a matter, producing a template for billing phrases. Give them dummy examples, never live ones, unless your firm has an approved business plan and a policy to match.

One office day, walked through with numbers

An illustration, using a four-solicitor practice with two administrators. At 8.30 the post and scanned documents arrive: 45 items. One administrator runs them through the approved tool, which returns a table in about six minutes. She spends 25 minutes checking it against the documents, corrects four matter references, and files all 45. Before the tool, sorting and naming took her about 70 minutes. At 11.00 a solicitor dictates a note after a client meeting; the practice system's assistant (or Copilot, if the note is in Word) produces a two-line file note, which she edits in two minutes. At 16.30 each fee earner pastes rough time notes, around 8 lines each, into the narrative prompt. Four earners times about 4 minutes is 16 minutes, against roughly 10 minutes each when they wrote narratives by hand with the rest of the day's interruptions. The saving there is small per person, but entries are posted the same day instead of the following month, which is what improves cash flow.

The totals are illustrative: roughly 45 minutes saved on filing, small savings on narratives, and a same-day posting habit. Judge it on your own figures after a month, not on those.

Opening a new matter: a checklist AI can help you build

Admin errors cluster at the start of a matter. Use the assistant, with no client details, to turn your habits into a one-page checklist, then have the senior partner mark it up. A typical result for a small litigation and property practice:

1. Conflict check completed and recorded, with the name of the person who ran it.
2. Engagement letter sent and signed copy filed; any client restriction on AI tools noted on the front of the file.
3. Client identification checks recorded.
4. Matter reference created in the format agreed; folder structure created from the template.
5. Key dates diary created, with the "who confirmed" box blank until a second person signs.
6. Billing arrangement entered: hourly rates, fixed fee or stages, and the billing unit.
7. Client told when and how they will receive updates and bills.

The tool contributes the structure and the wording; the partner supplies the content that makes it specific to your practice. Re-run the prompt each year with the changes you have made, so the list stays alive.

Prompting tips specific to office documents

Three habits improve every prompt in this tutorial. First, name the output format: a table with named columns, or a numbered list with a word limit. Second, tell the model what to do when it cannot tell: "write [unclear]" or "write [needs detail]". Without that line, models fill gaps. Third, give a worked sample line in your own style before asking for fifty more. For billing, one example narrative in the house style is worth a paragraph of instructions, because it fixes tense, order and length at once.

Handle scans carefully. Poor scans, handwriting and stamps are where text extraction fails quietly. If the tool says it cannot read a page, believe it; if it reads everything fluently from a faint fax, spot-check the numbers, because digits are the most common misreads. Keep the original image beside the extracted text for any date, amount or reference number that will drive an action.

What to put in the office policy

A one-page policy is enough for a small office, and the tool can help you draft a template, again without client data. Include: the names of approved tools and the plan each one must be on; what may never be entered (anything under a client restriction, anything in a conflicted matter, identification documents); the rule that a person checks and owns every output; the rule that a second person confirms every diary deadline; where staff record a mistake or near miss; and who reviews the policy and when. Ask a solicitor or your regulator's guidance to check it against your professional obligations, because the exact duties differ between offices and I am not giving legal advice here.

Realistic mistakes and how they showed up

Three failures are common enough to design against.

The helpful date. An assistant turned "within 28 days of the order" into a calendar date counted from the date the order was made, not from service. Nobody noticed until the reminder fired a week early, which happened to be harmless. Prevention: the extraction prompt above, which forbids calculation.

The over-friendly narrative. A tool turned "reviewed documents" into "carefully reviewed and analysed client's documents to assess the strengths of the claim". The client queried the entry. Prevention: ask for plain, short narratives and no adjectives, and delete any outcome the fee earner did not state.

The summary of the unreadable scan. A poor-quality scan produced a summary that sounded fluent and was partly wrong. Prevention: the "say unclear" instruction, and a check on a sample each week.

How to roll this out in two weeks

Week one: write the five-line confidentiality check for one tool, agree the naming convention, and choose one person to run filing triage on one day's post. Week two: add diary-date extraction for incoming letters on a single practice area, and billing narratives for one volunteer fee earner. At the end of each week, hold a fifteen-minute review: what did the tool get wrong, what did it save, what should change in the rules. Resist adding a fourth job until the first three have run for a month without a serious miss.

Measuring whether it works

Choose three numbers and write them down before you start: the minutes per day your administrator spends on filing, the days between work done and bill entry posted, and the number of billing queries from clients each month. After six weeks, compare. If entries are posted faster and queries have not risen, keep going. If queries rise, tighten the narrative prompt. If filing errors rise, return to human filing and use the tool for summaries only.

When you are ready for a wider plan, an AI implementation plan for a small law firm sets out a ninety-day sequence that builds on these admin wins before anything touches advice or drafting.

Law office AI admin: the questions that come up first

Can a law office use free ChatGPT for admin?

Only for material with no client information in it, such as a generic file-naming convention or a template billing phrase. Anything naming a client or matter belongs in a business plan or in a tool your confidentiality review has approved, with training on your content switched off.

Is Microsoft Copilot Chat safe for client documents?

Copilot Chat is included with Microsoft 365 business plans and, signed in with a work account, is covered by Microsoft's enterprise data protection, shown by a green shield in the chat, and Microsoft says it does not use your data to train foundation models. Check your own tenant settings and your regulator's guidance, and keep a written record of the decision.

Should AI set court or limitation diary dates?

No. Let AI extract and propose dates, but have a person calculate and confirm every deadline against the rules, then enter it in the diary system with a second-person check. A wrong limitation date can turn into a professional negligence claim.

Will AI billing narratives be accepted by clients?

They are accepted when accurate and specific. Clients object to vague or padded entries, not to the tool. The fee earner must still confirm the time and the work described before the entry is posted.

Further reads

• How to Summarise Bundles and Transcripts With AI Safely — Summarising bundles safely.
• How Small Law Firms Use AI to Draft Letters and Routine Documents — Routine letters once admin is under control.
• AI Client Intake for Law Firms: Qualify Enquiries Out of Hours — Qualify enquiries out of hours.
• AI Readiness Checklist for Accountants, Solicitors, Consultants — Check the firm is ready before widening use.
• How to Stop AI Inventing Case Law: A Small-Firm Checking Routine — A checking routine for any research use.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Microsoft Copilot for Business pricing page and Microsoft Learn's Copilot Chat privacy page, ChatGPT Business and Claude pricing pages, OpenAI's data-controls help page, Google Workspace's generative AI privacy page, Clio help-centre articles on AI features (all checked 4 November 2026; Clio's pricing page blocked automated access).