How work changes: watch
Work with documents has changed form several times already: from the speech to the printed text, from the archive to the database. Two films show what an assistant built on a language model changes: which steps get cheaper and what the person is still responsible for.
Exercises and files are in the text version.
Part 1Work has changed form before
Seven eras, from the ancient orator to the assistant. Each one shows what became cheaper and what became more valuable.
Exercises and details for this part · The film on its own page
Part 2What the assistant changes
One legal task for a client, worked through twice: how it was done before and how it is done now. An analyst's report or a memo for a manager breaks down the same way. Watch where the time goes and where the risk appears.
Exercises and details for this part · The film on its own page
As text
The same as the films: every shot and its text. You can copy the text and give it to your own assistant along with your question.
What the assistant changes
An assistant is a program you give a task in ordinary words and that answers in words too. Today it reads all the materials for a task in a few minutes, analyzes them, proposes a solution and prepares a document. On one example the film shows what the assistant does on its own in such work, where it goes wrong and what stays with the person. The example is legal, a landlord wants to end a lease early, but the work of an analyst, an accountant or an editor is built the same way.
Without an assistant the person goes through all this work alone, step by step. They find out what the client wants, gather the lease, the payment table and the correspondence, read them, find the right article of the law and compute the debt. Then they draw up options, write a memo, check it and sign it. Reading and writing take the most time, and checking is short, because the author read every source and wrote every sentence.
With an assistant the day starts not with reading but with the task, and the person sets it. The assistant knows only what it has received, so along with the task it gets the context, everything the answer needs: who the client is and what they want, the lease, the payment table, the correspondence with the tenant. The assistant can find some materials itself if it has access to mail or folders, but the person knows what actually matters to the case.
The assistant reads all the materials in full and within minutes builds a picture of the case from them. In the lease it finds the clause on early termination, in the table it sees that July, August and September are unpaid, and in the correspondence it notices the tenant’s request for a deferral that nobody answered. It catches most typical problems this way, because it keeps all the documents in view at once. At the end it draws up a plan: what still has to be found and computed.
The plan calls for the text of the law and the amount of the debt, and here the assistant turns not to its memory but to tools, separate programs that return an exact result. The legislation database gives the verbatim text of Article 782 of the Civil Code of Ukraine: the landlord may withdraw from the lease if no rent has been paid for three months in a row. The calculator multiplies UAH 18,000 by three and returns 54,000. This assistant has no access to the register of court decisions, so it has no way to check case law.
With the facts gathered, the assistant proposes not one answer but several options, and for each it writes what is good about it and what is risky. Withdrawing from the lease at once is fastest, but the tenant may point to the deferral request. Sending a written demand to pay first takes longer, but the landlord’s position gets stronger. The third option is to agree on ending the lease by mutual consent. It advises the second and drafts the documents right away: a memo for the client and a demand to the tenant.
The draft looks like a finished document, but the assistant errs differently from a person. Its text is written by a language model: the model picks a plausible continuation, so it does not tire or skip pages, but it can confidently invent a court ruling that does not exist. Sometimes it carries a number over wrongly: the lease gives 30 days to vacate the premises, the draft says 10. And it does not know what it was not given: the client agreed on a deferral with the tenant by phone.
So checking stays with the person, and the draft is checked not against memory but against sources. The cited ruling is looked up in the register of court decisions, and there is no case with that number. The deadline to vacate is checked against clause 8.4 of the lease, which says 30 days. The client is asked about the deferral, because it was in no file. The more a mistake would cost, the more claims are worth checking.
A checked memo is still only a proposal. Which of the three options to choose depends on things outside the documents: whether the client wants to keep the tenant, whether there is anyone else to rent to, what was promised by phone. The person weighs this and decides, and the assistant fixes the text to match the decision. The person also signs the memo and answers for it to the client, and sometimes to a court: the assistant signs nothing and is not responsible for the consequences.
This divides the working day differently. Reading, finding the rule, calculations, options and the draft move to the assistant and its tools and take minutes instead of hours. The person keeps the task and the context at the start, and checking, the decision and responsibility at the end, and checking now takes longer, because the person did not write the text. How much time this frees depends on the task, so it is worth measuring on your own work.