For people who work in a professional firm: solicitors, accountants, consultants. Annual accounts, judgments, contracts that come back from the other side, council circulars, clauses across several contracts, time reports, the documents to ask the client for, the current text of a statute, the meeting with the client. Nine cases, each with the request to copy and adapt and with what to check before the work leaves the firm.
The examples use fictional documents from a small firm and its clients. Each case lists the stages of the manual that teach the move. The cases were tested in Claude in the Italian edition of this guide, and the cases that report what the model did in a test were tested again in English on these documents, in November 2026. Where a test showed a typical mistake, the case says so.
Where the test used a fictional document, the case offers it for download: you can redo the test and compare the answer with what the case describes.
Before you start: client confidentiality and the regulators
People who work in a firm handle papers that the client entrusted to the firm alone. Four rules apply to every case below.
- With a free or personal tool, upload only documents without names, National Insurance numbers, tax references and details that lead back to the client. Replace them with labels ("company A", "client B") before uploading, not after.
- If the firm has a professional subscription, check what it says about data: whether it is used to train the models, how long it is kept, where it is stored (fact sheet "If the tool comes from your office").
- The UK has no AI law for the professions: the regulators apply the rules that already exist. For solicitors, the SRA's warning notice "Misuse of AI" (17 August 2026) says that client information should go into an AI system only with appropriate contractual, technical and organisational safeguards, and that the use of AI does not diminish or transfer your professional responsibilities. For accountants, the ICAEW guide on generative AI and ethics advises against loading confidential information into public tools and asks for transparency when AI supports work delivered to a client. Client personal data is also covered by the UK GDPR. If your engagement letter tells clients how you use AI, the line must be true.
- An opinion, a submission or a return signed by you stays yours. Every statute, judgment and figure cited is checked on the source before it leaves the firm.
1. From the annual accounts to the analysis report
The situation. Every matter has its accounts to read, reclassify and comment on: the same work on documents that are always different, and a ratio wrong by a decimal point changes a decision.
The request. In three steps, each with your go-ahead. First only the transcription of the tables, line by line with the page number, with no calculations. Then the reclassification and the ratios, calculated with code and with the formulas shown. Finally ten lines of assessment, each statement with the figure and the page that support it. Once the work runs smoothly, the three steps become a custom assistant that does it again on the next set of accounts.
What to check. That total assets equal total liabilities and equity, before going on. That the operating profit rebuilt by the model matches the one in the accounts. Three figures redone by hand, one of them a ratio. The questions the model asks you during the reclassification (where does this line go? do lease liabilities count in net debt?) are the point of the work: you give the answers, and they become rules for the assistant.
The full walkthrough. Actually run on the 2025 consolidated annual report of Brunello Cucinelli, 344 pages, with the assistant's instructions to adapt and the spreadsheet from the example: The analyst: from the accounts to the report. Two hours, in two sittings.
Stages. 3 (a document of your own), 7 (Excel: formulas, data and charts), 14 (a custom assistant).
2. Summarising a judgment
The situation. A ten-page judgment to read to see whether it bears on the matter on your desk.
The request. Upload the judgment, then:
"Summarise this judgment in half a page: the parties, what the claimant asked for, what the judge decided and why, the order on costs. List the statutes and rules cited with the exact section. Do not add cases or statutes that are not in the text. Then tell me in three lines whether the reasoning also applies to [your situation, without names]."
To redo the test. The fictional document used: case2_judgment.docx.
What to check. The statutes and rules cited, one by one, against the text of the judgment. If the summary contains a Court of Appeal or Supreme Court decision that is not in the document, it is invented until you find it in a law report or database: this is the mistake that has already brought lawyers before the courts, in England and elsewhere. The order on costs: it is the line a summary gets wrong or skips most often. The last question (does it apply to my case?) asks for an opinion: the model gives you a hypothesis to check. If it adds a rule the judgment does not cite, it should say so and mark it "to be checked": that is the right behaviour, and the check is yours.
3. The draft that comes back from the other side
The situation. You sent your draft contract. It comes back "with a few tweaks", with no tracked changes.
The request. Upload the two versions in Word, then:
"Compare our draft with the one returned by the other side. List every difference in a table: clause, our text, their text, effect for our client, who is [the buyer]. Put first the clauses that are no longer in their version and the new clauses. Then prepare a Word file with their version and our counter-proposals as tracked changes, with a comment for each one."
To redo the test. The fictional documents used: case3_contract_counterparty_draft.docx, case3_contract_our_draft.docx.
What to check. The clause that has gone: missing text does not show when you read only the new version, and it is the change that matters most. The small words: "normally", "reasonable", "unless otherwise agreed" turn an obligation into an intention. The numbering: if a clause disappears, the others move up and a comparison by number goes wrong; ask for a comparison by content. In the test the model found all six changes, the vanished termination clause and the new liability cap first, and noted the shifted numbering. The Word file: open it and check that the tracked changes are real ones, which you can accept and reject one by one.
Stages. 3, 6 (Word and Outlook).
4. From the circular to a note for clients
The situation. A circular or a new rule comes out and clients need to be told: who has to do what, by when.
The request. Upload the circular, then:
"Write a note for the firm's clients on this circular: who it concerns, what they have to do, by when, how much it costs, what they risk if they do not do it. One page, plain language. Include all the exceptions and the different deadlines for particular categories. At the end, a table with the deadlines."
To redo the test. The fictional document used: case4_council_circular.txt.
What to check. The exceptions: a different deadline for one category of business, an exemption that applies only if a form is submitted. They are the lines a summary loses first, and losing them lands a client with a penalty. Every figure against the circular. If the note cites a rule that is not in the circular, check it or remove it.
Stages. 2 (the same request, done well), 3.
5. The clauses of several contracts in one table
The situation. A client is changing hands, or wants to know where it is exposed. All its contracts have to be read looking for the same things.
The request. Upload the contracts, then:
"Make a table of the contracts: counterparty, term and renewal, termination, early exit charges or liquidated damages, jurisdiction, assignment, change of control. For each cell give the clause it comes from. If the contract says nothing, write 'not stated' and do not infer. At the bottom, list the clauses that could be triggered if the ownership of the client changes."
To redo the test. The fictional document used: case5_three_contracts.txt.
What to check. The "not stated" cells: if the model fills them with the rule of law that applies when there is no clause, the information may be right, but it is the model's and not the contract's. The change of control clause, which often sits at the end, in a clause on "general provisions". The clauses cited: open two at random.
6. From recorded time to the time report
The situation. At the end of the month the time recorded in the practice management system has to become the report for the client.
The request. Export the time entries to Excel or CSV, then:
"Prepare the October time report for client [label]: only their billable entries, grouped by type, with the hours and the total. Do the calculations with a program. List separately the rows that look like duplicates and the non-billable ones you excluded."
To redo the test. The fictional document used: case6_activities_october.csv.
What to check. The duplicate rows: a call recorded twice is a mistake the client notices before you do. The excluded rows: the model has to list them, so that you can see what it left out. The total: redo it.
Stages. 7.
7. The list of documents to ask the client for
The situation. Every matter starts with a list of documents to request, and the client always brings half of them.
The request. Paste the firm's internal list for that type of matter and the notes of the first meeting, without names, then:
"Using the firm's list and the meeting notes, write the email to the client with the documents they need to obtain. Only those needed for their situation: remove the ones that do not apply and tell me separately which ones you removed and why. For each document, where to request it, in one line."
To redo the test. The fictional document used: case7_probate_documents.txt.
What to check. The documents removed: compare the reason with the notes. Documents added that the firm's list does not include: they may be useful, but you decide. The "where to request it": if the model names an office or a website, check it, because procedures change.
Stages. 2, 5 (the finished work).
8. The current text of a statute and its amendments
The situation. You need the text of a section as it stands today, and to know when and how it has changed.
The request. With web search switched on (stage 15):
"Find on legislation.gov.uk the current text of section [number] of the [Act]. Give it word for word. Then list the amendments it has had since it was enacted: date, amending legislation, what changed. For each amendment, the link to the source."
What to check. The text: open it on legislation.gov.uk and compare it with the one reported, at least in the subsections you need. A model that cannot open the page may rebuild the text from commentary and articles, perhaps of an earlier version, or give from memory a section that has since been replaced: ask it which page it took it from. The amendments: on legislation.gov.uk amended words are in square brackets with a note (F1, F2...) under the section naming the legislation that made the change, and at the top the page says whether the text is up to date or lists changes not yet applied; that is the quickest check. A reference the model adds on its own is checked on the current text in the same way. In the test, asked for Article 22 of the UK GDPR, the model read legislation.gov.uk directly, said that the article no longer has a text in force because it was replaced by Articles 22A-22D (Data (Use and Access) Act 2025), gave the last version word for word with its dates, and said which page it could not open.
Stages. 15 (when it searches instead of remembering), 4.
9. Preparing the meeting with the client
The situation. Tomorrow you see the client. The matter is made of emails, notes and documents built up over months.
The request. In a project dedicated to the matter (stage 12), with the documents uploaded without names:
"Tomorrow I meet the client. Prepare one page for me: where the matter stands, what we have already decided, what is still open, the deadlines in the next three months, and the questions to ask to close the open points. For each point, the document it comes from."
To redo the test. The fictional document used: case9_matter_emails.txt.
What to check. The things "already decided": the model can take a proposal written in an email for a decision. The deadlines, against the documents. The questions are the most useful part; if one that seems obvious to you is missing, a document is missing from the project. In the test the model kept the waiver as a proposal ("nothing has been decided yet") and flagged the time report promised and not found in the file. Also read what it says about the files: a spreadsheet uploaded to a project can reach the model without its rows. A spreadsheet you really need, attach it to the chat.
Cases tested on 7 November 2026.