An AI Law Firm Just Won Its First Trial. It Charged £400. The Losing Side's Lawyers Cost Thousands.
The Economist calls AI lawyering's results "mixed." But a regulated AI firm just won a contested three-hour trial in the UK, the average US attorney charges $353/hour, and the World Justice Project says the legal needs of 5.1 billion people worldwide go unmet. An original cost-per-outcome analysis suggests AI legal tools don't need to be perfect. They need to be better than nothing.
Four hundred pounds is what Tamires Camal Taquidir paid Garfield AI, the world's first regulator-approved AI law firm, to recover £7,000 in unpaid fees from a hospitality business through three hours of contested trial at Wandsworth County Court on May 14, 2026. Her opponent hired a solicitor and a barrister; Dominic Li argued the case for her in court. They won. When the Economist reported this case, it called AI lawyering's results "mixed," burying the courtroom win in the final paragraph beneath five paragraphs of hallucinated case law and judicial sanctions, inverting what actually happened.
Legal representation is so expensive that most people cannot afford it, and the first AI system to clear a regulatory bar just beat a traditionally lawyered opponent in open court at roughly 1/20th the cost. That is not a mixed result.
The Cost Nobody Compares
Most of the Economist's word count goes to what happens when unguided people paste legal questions into ChatGPT, and the failures are genuine. Omar Rafique filed fake case law in a British tax dispute. A student at the University of Ottawa was ordered to pay C$10,000 in costs after using AI improperly to challenge her removal from a PhD programme. Sullivan & Cromwell, a top New York firm, apologized for errors caused by AI hallucinations. Real problems, all of them. But the article never asks the question that matters: compared to what?
According to Clio's 2026 Legal Trends Report, the average US attorney hourly rate for civil litigation is $353. Collections and debt recovery runs $320 per hour. A small claims dispute that might take 10-15 hours of attorney time costs $3,200 to $5,300 before anyone sets foot in a courtroom. For a £7,000 ($9,300) debt like Taquidir's, traditional legal fees would consume 34 to 57 percent of the recovery. Her AI fees consumed 5.7 percent.
At the top end, rates are accelerating away from ordinary people. Reuters reported in January 2026 that Susman Godfrey now charges $4,000 per hour for senior partners. Average partner rates across large US firms hit $1,114 per hour. These numbers have risen 83 percent in a decade.
Meanwhile, the people who cannot pay those rates simply go without. Anand Shah and Joshua Levy, the MIT and USC researchers cited in the Economist piece, found that 17 percent of federal civil litigants represented themselves in 2025, up from a longstanding 11 percent. What gets framed as AI "emboldening" people to go it alone could just as accurately be described as AI giving people access to legal reasoning they previously could not afford at any price.
The Hallucination Problem Is Already Solved
Every failure case in the Economist's article shares a common architecture: a person pasting a legal question into a general-purpose chatbot and treating the output as verified legal work product. ChatGPT is not a legal tool; it is an autocomplete engine that happens to know some law, and treating its hallucinated citations as the state of AI lawyering is like judging the automobile industry by the performance of a riding lawnmower on the highway.
Garfield solved this problem before it launched by eliminating the risk entirely. As the Solicitors Regulation Authority noted explicitly, Garfield's system "will not be able to propose relevant case law, which is a high-risk area for large language model machine learning." Instead it handles procedural heavy lifting: letters before action, claims forms, default judgment applications, settlement negotiations, trial preparation. A human barrister argues in court while named regulated solicitors remain accountable, and every client action requires explicit approval.
Not a chatbot, but a regulated legal workflow engine with AI at its core and human oversight at its edges, and it just won a contested trial.
Original Analysis: Cost Per Legal Outcome
No reporting on AI lawyering has compared cost-per-outcome across delivery models. Using the Garfield trial as the first data point for regulated AI, and Clio's rate data for traditional representation, we can construct a rough comparison for a £7,000 small claims recovery:
| Model | Estimated Cost | Net Recovery | Cost as % of Claim |
|---|---|---|---|
| Garfield AI (actual) | ~£400 | £6,600 | 5.7% |
| Traditional solicitor (Clio avg) | ~£2,500-£4,200 | £2,800-£4,500 | 36-60% |
| Self-representation (no AI) | £0 | £0-£7,000 | 0% (but high failure rate) |
| Self-representation + ChatGPT | ~£20 | Variable, risk of sanctions | Negligible + risk |
One data point is not a trend, but the ratio is striking enough to demand attention. Garfield's model delivers professional-grade legal process at consumer-app pricing, with regulatory accountability that raw ChatGPT usage completely lacks. What matters is not whether AI lawyering works in the abstract, but whether it works better than the status quo for the billions of people currently priced out of the system entirely.
The 158% Statistic Cuts Both Ways
Shah and Levy found that AI-assisted self-represented litigants file 158 percent more documents in a case's first 180 days, a statistic the Economist presents as a clogging problem for courts. Antony Sendall, a British employment barrister quoted in the article, complains that "a grievance that would have been a few sentences is now more like 10 to 12 pages." But consider what that complaint actually says: people who previously lacked the legal vocabulary to articulate discrimination they experienced can now produce comprehensive documentation that resembles what a lawyer would write. Employment discrimination law in the UK requires claimants to identify protected characteristics, distinguish direct from indirect discrimination, and demonstrate shifting burdens of proof, and a "few sentences" from an unrepresented worker almost certainly undersells the actual harm suffered rather than reflecting a proportionate claim. Courts may not like the volume, but the question is whether a person's access to justice should depend on their ability to pay $353 an hour for someone to write properly formatted motions on their behalf.
For 5.1 billion people whose legal needs go unmet according to the World Justice Project's 2024 Rule of Law Index, that answer is obviously no.
What You Can Do
If you have a small claims dispute under £10,000 in England and Wales, Garfield AI is live and SRA-regulated right now. For disputes elsewhere, general-purpose chatbots can help you understand your legal position, but never cite a case they generate without independently verifying it exists in an official law database. An emerging legal consensus holds that your AI conversations may be discoverable in court proceedings, so treat them like notes to your own lawyer rather than a private diary. If you run a small business with outstanding invoices, the regulated AI model is already cheaper and more effective than ignoring the debt or paying a solicitor to chase it.
Limitations
This analysis relies on a single trial outcome from one regulated AI provider operating in one narrow domain: UK small claims debt recovery up to £10,000. Garfield has not published aggregate win/loss statistics or case volumes, so there is no way to assess whether this win is typical or anomalous. Cost comparisons use Clio's average US rates applied to a UK case for illustrative purposes, and actual UK solicitor rates for small claims work may differ substantially. Shah and Levy's filing volume data does not distinguish between substantive and frivolous filings, making the 158 percent increase difficult to evaluate on quality grounds. And critically, Garfield's trial was argued by a human barrister, not by AI; the system handled preparation, not advocacy.
Strongest Counterargument
Strongest counterargument: the Nippon Life lawsuit reveals a genuine harm where an AI system encourages litigation the claimant would not otherwise have brought, generating legal costs and organizational disruption even when the underlying claim is meritless. If AI consistently inflates people's confidence in weak cases, the net effect on the justice system could be negative even as individual access improves. Courtready's data showing 79 court rulings flagging non-existent cases in the first half of 2026 alone, against seven in all of 2024, suggests the quality problem is growing faster than solutions, and that acceleration deserves honest engagement: making it easier to file lawsuits is only good if the lawsuits have merit.
Measuring the Framing
The Economist's article is itself an example of a detectable editorial pattern: bury the strongest evidence for the thesis you want to undermine, lead with anecdotes that support your preferred frame, and let structure do the arguing your prose cannot. This is not speculation. MediaScope, an open-source toolkit for tracking editorial bias and coverage asymmetry across publications (MIT license, 960 tests, 41 framing device types, 253 detection patterns), can identify exactly these techniques programmatically: editorial deflation (building up a finding then puncturing it in the same paragraph), selective omission of competing data points, and structural burial where the strongest counterevidence to the article's thesis appears in the final paragraph. The toolkit uses a difference-in-differences approach adapted from Card and Krueger (1994) to decompose institutional bias from individual journalist bias by tracking how coverage changes when writers move between publications. If we are going to hold AI legal tools to standards of accuracy and disclosure, the same standards should apply to the publications evaluating them.
The Bottom Line
A regulated AI law firm just won a contested trial in open court for £400, while the average American pays $353 per hour for a civil litigator, assuming they can find one willing to take a small case at all. Five billion people worldwide have legal needs that go completely unmet. When the Economist spent 600 words on hallucinated citations and one paragraph on the first AI courtroom win in history, it told the wrong story. Hallucinations are a real engineering problem and they are being solved by architecture, not by slowing down. What cannot be solved by architecture is an $1,114-per-hour average partner rate and a profession that serves less than half the population that needs it. AI lawyering's results are not mixed; the results are early, and the trajectory is unmistakable.