This article is about how I used AI agents to prepare a legal case without a lawyer, and how you can too, without getting burned.
A few years ago I paid £5,314 for three specialists to assess a close family member. A combined 150 pages, written by people who knew exactly what they were talking about, all reaching the same conclusion about the support that was required. The local authority dismissed it on cost grounds. What followed was years of appeals and, in the end, a tribunal.
I am not a lawyer. I run a software consultancy. But I have now fought public bodies through two different legal frameworks, special educational needs and family law, and both times I learned the same thing. The fight is rarely won in a courtroom. It is won or lost in the months before anyone gets near one: in the letters, the evidence, and the arguments you build while you can still shape them. This article is about how I used AI agents to do that groundwork myself, and how you can too, without getting burned.
That groundwork used to need a lawyer. For most people it still does, and that is the problem, because most people cannot pay for one.
Pre-litigation advocacy, sometimes called pre-action work, is the legal groundwork you do before formal proceedings: understanding your position, mapping your options, and writing correspondence that stands up to scrutiny. It is the most expensive part of the whole process, and the part that has always sat behind a fee. And it is exactly the part that an AI agent, used with discipline, can now help you do for yourself.
When the Money Runs Out Before the Fight Does
Most people meet the legal system at the worst moment of their lives, perhaps also holding the least money they will ever have to spend on it. Legal aid is narrow and getting narrower. Private fees run to hundreds of pounds an hour. By the time a matter truly matters, the budget is often already gone. More and more people end up as ’litigants in person’, representing themselves not by choice but because the alternative is nothing.
In my family’s case the running total of private fees passed £30,000, and that was before the part where we most needed advice. A single emergency appeal to a tribunal, drafted by a charity, cost £575 and was worth every penny. None of it was sustainable. We could not keep buying representation indefinitely, and we are better off than many other families.
Here is the uncomfortable truth. A well-funded institution with an in-house legal team can simply outlast you. It does not need to be right. It needs you to run out of money, energy, or time first. That is not a flaw in the system. For the institution, it is a feature.
This is the gap that AI changes. Not by winning your case, but by putting the groundwork back within reach.
The Same Pattern, Every Time
I have watched this play out three times, in three different settings.
In special educational needs, a parent with a stack of expert reports against a council that controls both the budget and the timetable. In a family law matter, a public body agreed to act but did not, with no funded way to hold it to its word. And in my day job as a software engineer expert witness, where I do this for a living, working out what really failed in a project, and why, before anyone files a claim.
The setting changes. The pattern does not. One side has resources and repetition. The other has the facts, and sometimes no way to apply pressure. Pre-litigation advocacy is how you turn facts into pressure.
What You Can Realistically Expect
Let me start by being careful, because this is where overclaiming would do real harm.
AI did not win anything for me. It will not stand up in your hearing, it will not give you regulated legal advice, and it cannot take responsibility for a decision. What it does is get you to the starting line: a clear view of your legal position, a map of the routes genuinely open to you, and correspondence that is accurate, properly grounded, and hard to brush off.
That is the phase I think of as “up to the point of representation”. You arrive at a lawyer’s door already understanding your case, with the groundwork done and the documents in order. If you can afford one hour of a specialist’s time, that hour is now worth far more, because you are not paying them to start from nothing.
And sometimes the groundwork is enough on its own. A letter before action that is accurate, well reasoned, and clearly going somewhere can change behaviour without anyone going near a court. That is the entire point of pre-action work. Most disputes are meant to settle before proceedings, and acting early, through pre-action technical assessment, is what makes that possible.
But Can You Trust AI With Something This Important?
It is the right question to ask.
Yes, AI makes things up. It will invent a case citation that looks perfect. It will cite the wrong paragraph of a regulation with total confidence. It will state a figure that is simply wrong, in fluent, professional prose. A wrong answer from an AI looks identical to a right one. That is exactly what makes it dangerous in untrained hands.
So the trick is this. You do not trust the AI. You trust the process you build around it.
In my own drafting, that process caught a string of confident mistakes. It dated a well-known human rights judgment to 1989, when the law report is 1988. It placed a legal duty under a section of the wrong Act entirely. It produced a real, well-known case but gave it a citation that does not exist, formatted perfectly for the year. And it invented a medical detail about my own family that was nowhere in the records. Every one of those would have sailed straight into a serious letter if I had trusted the first draft. I caught them the slow way: by checking each claim against the source document, and by having a different model read the draft cold.
I should be honest about the cost of that process. It is real work. Skilled, patient, and at times exhausting work. This is not a prompt you fire off before lunch. My own letter went through three legal research agents, then nine drafts and several rounds of reviews. It took several days and a small swarm of agents running in Claude Cowork. It is closer to running a small case review, several times over.
The Harness: How to Make AI Output Trustworthy
Here is the method, stripped down to the parts that matter. It is the same loop I use for agentic software development, adapted for legal work.
Anchor everything to a primary source. Never let the AI’s summary stand in for the actual document. Work from the order, the statute, the report, or the regulation itself. And check that you genuinely hold the final version of each, because a confident summary of a document you have never actually seen is exactly how people go wrong.
Make the AI generate options, and you make the calls. The AI is good at widening your view: six ways to frame an argument where you would have thought of one, the route you had not considered, the risk you had missed. In my own case it reframed the whole objective, showing me a smarter thing to ask for than the one I had set out to ask for. But which option fits your situation is your decision, not its. That decision stays with you.
Run adversarial review, not just one pass. Take your draft and have a fresh AI session attack it, briefed cold, with no idea what the last one concluded. I tell it to assume the letter is wrong, and to find the weakest claim and the citation most likely to be invented. Then do it again with a different model entirely. When two independent reviews agree, that is corroboration. When they disagree, chase it. Twice, the second model caught a problem the first had declared clean.
Verify every citation and number against the source, by hand. This is the unglamorous one, and the one you cannot skip. Every case, every section number, every date, checked against the real thing. If you cannot verify it, it does not go in the letter.
Know where the boundary is. The most useful thing the process gave me was not the letter. It was a clear list of the points where I genuinely needed a regulated specialist to confirm the position before I relied on it. Knowing what you do not know is the whole difference between empowered and reckless.
Protect what you feed it. If you are dealing with a child, a medical history, or a live legal matter, be careful what you paste into a public chatbot. Use tools that do not train on what you give them, and strip out names and identifying details where you can. The point is to help your case, not to leak it.
If you want to see exactly how that was set up — the roles, and the shape of the prompt I gave each one — I have written it up separately.
Some Mistakes I Made
A few ways this goes wrong. I made each of these at least once, but caught each one before it reached the final letter. That is rather the point: the aim is not a flawless first draft, it is a process that catches the mistakes before they matter. You will trip on some of these too, and that is completely normal.
- I trusted the review over what I knew. More than once an independent review told me a strong point was weak, or pushed a change that sounded clever and was wrong for my situation. I had witnessed the events; the model was only reading about them. A cold review is one of the most valuable things you can run, but it is an input, not a verdict. When it contradicts something you know first-hand, weigh it, do not just obey it.
- I assumed the newest draft was the cleanest. By the ninth version I trusted it more than the first. I should not have. Fixes I had already made had quietly crept back out, and new errors had crept in, because every edit is a fresh chance to break something. Re-check each version, the citations above all, as if it were the first.
- I let it rewrite the whole thing at once. When I asked the AI to regenerate a full draft, it silently changed wording I had deliberately chosen and reshaped passages I had not asked it to touch. For anything where the exact words matter, and in law they always do, make small targeted edits and read the difference each time, rather than handing back the whole document.
What This Actually Puts Within Reach
I am not telling you to fire your lawyer. I am telling you that the months before you need one are no longer dead ground you can only cross with money.
If you are a parent with a stack of reports and a council that keeps saying no, or anyone facing an institution that is quietly betting you will give up, the groundwork is now within reach. Done with discipline, an AI agent can help you understand your legal position, build your evidence, and write a letter that is accurate, grounded, and hard to dismiss. It will not replace the expert. It will get you ready for the expert, and sometimes it will get you what you need without one.
If you want to start somewhere, start with the documents. Get every order, report, and letter that matters into one place, and read them yourself before you read anyone’s summary of them. That is the ground everything else is built on.
But I should be straight about where I write this from. Pre-action discovery, establishing what went wrong and building the evidence before proceedings, is what I do professionally as a software expert witness, under CPR Part 35 and trained with Bond Solon. If you are a solicitor or a business facing a software or IT dispute, that groundwork is exactly the work I take on. For everyone else, the same discipline is now yours to borrow. This is what access to justice can look like when you cannot pay for it.
The field is still tilted. It is just a little less steep than it used to be.