A UK law firm helped a client recover £7,000 in debt at a cost of roughly £400. The one doing the heavy lifting wasn’t a lawyer—it was AI. The firm is called Garfield AI, the world’s first AI law firm to receive authorization from the UK’s Solicitors Regulation Authority (SRA)[Source: https://garfield.law/]. Almost everything up front—evidence organization, legal research, drafting collection letters and complaints—was handed off to AI. Their specialty is exactly the kind of small-debt recovery where hiring a lawyer never used to be worth it[Source: https://garfield.law/].
And in this actual winning case, where the plaintiff recovered the full amount, AI was in charge of automatically uploading contracts and invoices and precisely generating legal correspondence, bringing the total cost down to around £400[Source: https://garfield.law/]. AI took over the most labor-intensive paperwork at the front end, while humans stayed at the back end to make the judgments—and that dividing line is really the point of this whole article.
It’s Not “Law” That Gets Eaten—It’s the “Repetitive, Structured, Low-Ambiguity” Part
Let me start with a piece of logic that a lot of people get backwards: AI isn’t going after “law”—it’s going after “work patterns.” To judge whether a piece of document work will be automated first, the observation from Judy AI Lab is to look at three criteria—repetitiveness (you need a hundred copies of the same format in a single day), structure (fixed fields, standard templates), and low ambiguity (the correct answer doesn’t vary much from person to person). When all three are high, AI eats it fast and clean.
This isn’t an isolated trend. Deloitte has observed that the application of generative AI in the legal field goes far beyond simply replacing digital tools—it can transform tedious, time-consuming legal work into efficient, precise, intelligent workflows[Source: https://www.deloitte.com/tw/tc/services/tax/perspectives/2025-outlook-lmc.html]. The key distinction to draw here is that this shift is about “task-level automation,” not “job replacement rate”—one legal-tech practitioner estimated based on trends that around 60% of legal work could be taken over by AI or automation within the next two years, but emphasized that “it’s not job positions being replaced, it’s job content being replaced by AI or automation”[Source: https://www.threads.com/@lawbot_tw/post/DVGE_lKE0vY/3-ai-%E5%8F%96%E4%BB%A3%E5%BE%8B%E5%B8%ABai-%E6%98%AF%E5%90%A6%E6%9C%83%E5%8F%96%E4%BB%A3%E5%BE%8B%E5%B8%AB%E4%B8%80%E7%9B%B4%E6%98%AF%E5%BE%88%E5%A4%9A%E4%BA%BA%E9%97%9C%E6%B3%A8%E7%9A%84%E8%AD%B0%E9%A1%8C%E4%BA%8B%E5%AF%A6%E6%98%AFai-%E8%83%BD%E5%90%A6%E5%8F%96%E4%BB%A3%E5%BE%8B%E5%B8%AB%E4%B8%8D%E6%98%AF%E6%8A%80%E8%A1%93%E5%B1%A4%E9%9D%A2%E7%9A%84%E5%95%8F%E9%A1%8C%E4%BB%A5%E7%99%BC%E5%B1%95%E8%B6%A8%E5%8B%A2%E4%BE%86%E7%9C%8B%E5%85%A9%E5%B9%B4%E5%85%A7%E8%87%B3%E5%B0%91%E6%9C%83%E6%9C%89-60-%E5%B7%A6%E5%8F%B3%E7%9A%84%E6%B3%95%E5%BE%8B%E5%B7%A5%E4%BD%9C%E8%A2%AB-ai]. What AI actually takes over first is high-volume case-file retrieval, repetitive document drafting, and tedious procedural follow-up, freeing up humans for higher-value work like legal analysis and strategy[Source: https://www.deloitte.com/tw/tc/services/tax/perspectives/2025-outlook-lmc.html]. In other words, what disappears first is the “work content,” not “the person.”
Debt collection letters hit all three. They follow a fixed playbook: header, facts of the debt, amount, payment deadline, consequences of non-payment, legal basis. The only variables are names, amounts, and dates. Hand this to a human and it’s pure drudgery; hand it to AI and it’s the kind of fill-in-the-blank work it’s best at. That’s how Garfield can compress the cost of drafting a collection letter so low—so low that small businesses, who used to give up on recovery because the amount was too small to justify a lawyer, finally have a usable tool. And that’s exactly why an AI law firm getting formal SRA authorization matters[Source: https://garfield.law/].
On the flip side, the more a task demands improvisation, reading people, and strategic bets, the less AI can touch it. A judgment table might look something like this:
| Document Type | Repetitiveness | Structure | Ambiguity | Suitable for AI Drafting? |
|---|---|---|---|---|
| Debt collection letters / payment reminders | High | High | Low | Yes (AI draft + light human review) |
| Standard NDAs / boilerplate lease contracts | High | High | Medium | Mostly yes (custom clauses need human review) |
| Small-claims filings, court forms | High | High | Low to medium | Yes (mostly procedural) |
| Contract negotiation, custom clauses | Low | Medium | High | No (strategic) |
| Courtroom defense, witness cross-examination | Low | Low | High | Definitely human |
| Factual disputes, evidence judgment | Low | Low | High | Definitely human |
Contract documents follow the same logic: they’re highly structured from generation through review to downstream use, which is why they’ve long been one of the most active battlegrounds for generative AI in the legal world—and again confirms that rule of “the more structured, the sooner it gets eaten.”
Three-Layer Architecture: Template Generation, Case-Law Retrieval, and Compliance Comparison—Each Owns a Piece
If you want to build your own debt collection letter pipeline, don’t think “just throw a prompt at AI and have it spit out a letter”—large models hallucinate, and they’re also sycophantic, easily led by how the user frames the question into saying what you want to hear. That’s especially dangerous in a legal context. Split it into three layers to stay safe. This kind of division of labor is a common multi-model pattern in document-heavy professional settings: first use an OCR model to batch-recognize text from paper documents, then use a Named Entity Recognition (NER) model to extract key info like names, company names, and dates, and only then hand it off to a large model to analyze relationships and generate a report—each module does its own job. A debt collection letter pipeline can borrow the same idea.
Layer 1: Templated document generation. Lock down the skeleton of the debt collection letter as a template, and let AI only fill in the variables (debtor, amount, due date, late interest). With the skeleton fixed, the room AI has to wander is cut in half. This “structured case-template generation” approach works equally well for small-claims complaints—hand the model the parties’ info, loan amount, interest rate, collateral, etc., fill them into a fixed format to produce a draft, because the fields in these documents are already highly standardized.
Layer 2: RAG retrieval of statutes and case law. RAG (Retrieval-Augmented Generation—plain English: let AI look through a database you’ve given it before answering, instead of making stuff up from memory). Compile the relevant civil code articles and the local court’s handling conventions for small debts into a knowledge base. When AI needs to cite legal authority, it can only pull from there—if it can’t find anything, it’s not allowed to write. One of the big criticisms of large models is that they hallucinate and are sycophantic, easily led by how the user frames the question into saying what you want to hear; this layer is exactly the key to blocking those “sounds professional but the article doesn’t exist” hallucinated clauses at the door.
Layer 3: Compliance and fact comparison. Generation isn’t the end. This layer takes the produced draft and cross-checks two things: whether the facts—amounts, dates—match your source data, and whether the tone and content cross any unfair-debt-collection red lines (threatening or harassing language is illegal in many jurisdictions). You can call on AI again here for the cross-check, but a human absolutely must do the final review.
Hands-On: Build a Debt Collection Letter Draft Generator with Claude in About 30 Minutes
You don’t need to code. Get a knowledge base file ready (just a folder with your compiled key regulations, collection letter templates, and your shop’s payment terms), then feed Claude a system prompt like this:
| |
Two small things in the retrieval settings most affect quality: don’t make your knowledge base chunks (segments) too large—one article per chunk is ideal for retrieval; and turn on the “only cite retrieved content” mode. Better to let it reply “not found” than to let it fabricate a fake legal article just to make the answer sound pretty. This is the same principle we wrote about in the IBKR MCP Brokerage Account Secure Setup Guide—lock down its ability to run wild first, then talk about convenience.
Red Line: AI Drafting ≠ AI Accountability
The last section is the most important, because the more useful everything above gets, the easier it is to slack off here.
No matter how smoothly Garfield’s pipeline runs, what that winning case saved on was repetitive labor and document costs, not responsibility for review and consequences. Translated into something a small business owner can use: AI can help you generate a debt collection letter draft in minutes, but before you send it out, you need a human to verify that the amounts are correct, the legal citations actually exist, there are no jurisdictional issues you’re unfamiliar with, and whether to actually go to court if the other party ignores you—these are judgments involving consequences and strategy that can’t be left to AI alone. Documents carry positions, and large models hallucinate and can be led; that’s exactly why the review step has to be left to a human. When the amount is large, when the other side might bite back, or when there are factual disputes in the case, hire a lawyer. AI saves you the cost of repetitive labor, not the responsibility of carrying your judgment.
The line is actually pretty clear: put AI on “drafting” and “retrieval,” keep humans on “deciding” and “reviewing.”
A debt collection letter being automated looks like a small thing. But it marks exactly that line—which tasks are repetitive enough that machines should take them, and which ones are expensive but still have to stay with humans. The rest is up to which side you stand on.