The bot is the trade’s new apprentice

Almost every office works the same way. Information and documents arrive. Someone processes them using the firm’s procedures, their knowledge and their experience of previous cases. A result emerges, and the person who signs it takes responsibility for it. A marketing department receives the figures from a campaign and prepares the next proposal. An accountancy firm receives invoices and bank statements and prepares the accounts. A law firm receives the paperwork for a flat and delivers a contract.

Today, bots can take on some of that work in between. A person defines the task at the start and reviews the result at the end. There is more: the bot can be taught the firm’s way of working, so that today’s correction improves tomorrow’s work. An example makes this clearer. It is fictional.

A sale contract

A lawyer at a small firm in Spain prepares a private contract for the sale of a flat between two individuals. This is the agreement the buyer and seller sign before they later attend a notary’s office to sign the formal deed transferring the property. She gives the system the firm’s contract template, the terms agreed between the parties (price, date for signing before the notary, who pays what) and the documents for the flat.

A first bot drafts the contract. It is an agent, a bot with tools to carry out tasks, and it divides the work between several helpers, called subagents: one adapts the contract, another checks the Spanish Land Registry, and the third reviews dates and payments.[1] A second bot audits the draft with the same material in front of it. It should come from a different provider: Kim and his colleagues found that models from the same company tend to make the same mistakes.[2]

The auditor finds two errors. The contract still had 1 June, the date in the template, when the parties had agreed on 1 September. It also left open who would pay for work on the lift that had already been approved. Under the agreed terms, the seller pays. The auditor corrects the errors and explains them in a report that the lawyer reads alongside the contract.

What the system did not know

At eight in the evening, with the draft already reviewed, the lawyer remembers something the buyer told her: without a mortgage, he cannot afford it. The contract does not protect him if the bank says no. He would lose the money he has paid in advance.

This time, no bot failed. The information was missing from the brief. It had never made it beyond a conversation with the client. The text is consistent with the information received, yet still leaves the buyer exposed. It happens often: information arrives incomplete, or arrives later.

The lawyer stops the contract going out, speaks to the buyer and considers what condition to propose to the seller. The aim is to let the buyer withdraw from the purchase without losing that money if the mortgage is refused. The parties will have to agree to it. A bot has no authority to take their consent for granted.

This is what human review is for: checking the document against what is known about the case, including what never reached the system. Even with a drafter and an auditor, nothing goes out until a responsible person has reviewed and approved the final version. Circular 3/2026 from the Consejo General de la Abogacía Española, the General Council of Spanish Lawyers (CGAE), reminds lawyers that whoever signs takes responsibility for the content and must verify it, including when AI is involved.[3] The CCBE, which represents the European legal profession, calls on lawyers to understand the tools and protect client information.[4]

From correction to rule

The lawyer corrects the contract and writes to the bot: ‘Learn from this for next time.’ ‘Yeah, yeah,’ it replies.

A month later, on another sale, the bot again fails to ask about the mortgage.

A chat is not a memory. A correction buried among conversations is of little use unless someone turns it into an instruction consulted at the start of the next job. Some systems store notes outside the chat and retrieve them later.[5] Someone still has to decide what is worth keeping.

Here, the correction becomes a rule: ‘Before drafting a sale contract, ask whether the buyer depends on a mortgage and what they want to happen if the bank says no.’ It is added to the procedure manual, the file of instructions and examples that the bots consult before each task.[6]

A third bot, the architect, can propose the rule. At night, it reviews tasks, reports and human corrections, looking for the cause of each failure. A responsible person approves the rule, and it is recorded with a date and a reason, keeping the previous version. Without that control, the firm’s practice would change while everyone slept.

The next client is asked about the mortgage before the contract is drafted, and the auditor checks that the question was asked. That is learning. The model is still the same. What has changed is what it reads before it starts. A well-known research study, Reflexion, showed that models improve when their errors are recorded in writing and fed back to them.[7] A laboratory is not a law firm, so the new rule is tested before it is trusted: a purchase with finance, another without a mortgage and another where the answer is missing.

Experience that can be written down

Rule by rule, the manual gathers what the firm has learnt case after case: the questions to ask, the mistakes no longer repeated, the way to finish a job. This is the firm’s expertise, its know-how. Until now, it lived in the heads of those with years in the trade and passed slowly from partner to trainee. Now it can be standardised: it is written down and applied to every task, whoever handles it.

The model is rented, and any competitor can rent the same one. The manual cannot be bought anywhere. A firm with twenty years of property sales has much to teach its bot. A newly opened firm with the very same model has little. Accumulated experience is what sets one apart from the other, provided someone takes the trouble to write it down.

An internal asset needs to be safeguarded. The firm must retain its manuals, reports and corrections in readable files that it can retrieve and take to another system if it changes provider. The contract with the provider must ensure that data can be exported and will not be used to train models without permission. It must also set out confidentiality safeguards. The CCBE guides help assess those terms.[4][8] The accumulated expertise belongs to the firm. The client’s data still belongs to the client.

The office still works as it did at the start: information comes in, is processed, a result comes out and someone signs it. What changes is where the experience used to process it is held. Anyone who has trained an apprentice or a trainee lawyer will recognise the work ahead: explain the trade, review what they deliver and check, in the next case, that they have learnt to ask about the mortgage.

Sources

Accessed: 13 November 2026. Numbering for this revision. The example and organisation of the work are an illustrative proposal; the sources do not establish that the author uses this approach.

1. Anthropic, Create custom subagents, continuously updated Claude Code documentation. Helpers with their own tasks and permissions. Delegation and parallel work require tools and permissions; they are not included in every chat.
2. Elliot Kim and colleagues, Correlated Errors in Large Language Models, 09-06-2025, section 3.2. Original research. An association between a shared developer and coinciding errors; it guarantees neither independence between providers nor legal accuracy.
3. CGAE, Circular interpretativa del Código Deontológico 3/2026. Uso de sistemas de inteligencia artificial generativa y deber de verificación, 10-04-2026, sections 4.2–4.3. Institutional copy hosted by the Bizkaia Bar Association. Responsibility for the content and professional responsibility.
4. CCBE, Guide on the use of generative AI by lawyers, 02-10-2025, sections 4.1–4.2, pp. 16–19. Professional guidance. Confidentiality, competence and examination of the provider’s use of data.
5. Anthropic, Effective context engineering for AI agents, 29-09-2025, ‘Structured note-taking’ section. Persistent notes outside the active conversation. Describes how to store and retrieve notes; editorial control of the procedure forms part of the proposal.
6. Anthropic, Equipping agents for the real world with Agent Skills, 16-10-2025, updated 18-12-2025. Reusable instructions and resources. Technical basis for the procedure manual; the nightly review and its human approval belong to the proposed design.
7. Noah Shinn and colleagues, Reflexion: Language Agents with Verbal Reinforcement Learning, 20-03-2023, revised 10-10-2023. Original research. Improvements through textual memory in experimental tasks; it does not demonstrate the automation of an office.
8. CCBE, Guidelines on the use of cloud computing by Bars and lawyers, 27-02-2025, section II.6, pp. 10–12. Contractual terms and data retrieval. Limits on use and retrieval in a readable format when changing provider.