- Of people who made or considered a formal complaint about a regulated service in the past year, 54 percent used an AI tool to help, according to Legal Ombudsman research published on 1 October 2026.
- A review of 100 complaint files found evidence of AI use in 51 percent of them, with unsupported legal references and no clear resolution request among the markers.
- AI is not only producing more complaints. Of those who used it to decide whether to complain at all, 47 percent decided not to and 32 percent decided to.
- Better written complaints do get better outcomes. In 1.1 million filings to a United States regulator, those appearing AI written won relief about 50 percent of the time against 40 percent for human written ones.
- The researchers traced that gap to presentation rather than substance, which tells you exactly where to aim your reply.
- The invented legal citation is the trap. Answering it as though it were real concedes a rule that does not exist.
The email is four pages long. It is laid out with headings and bold text, it cites two regulations by number, it uses the phrase "pursuant to" twice, and it closes by reserving rights. It is from a customer who bought a 40 pound lamp.
Your first instinct is that you are being threatened by a lawyer. You are not. You are reading a prompt result. An AI written complaint sounds like counsel and is usually a customer complaint with better typography, and the person on the other end is as uncertain as they were before they opened the tool. The difference is that they now sound certain, and you have to answer the substance without being pushed around by the packaging.
How common is this actually?
Common enough to be the normal case in regulated sectors, and we now have numbers rather than impressions. The Legal Ombudsman published research on 1 October 2026 based on an Ipsos survey of 2,196 adults alongside a review of 200 complaint files. Legal Futures reported the headline figures: 22 percent of people had made or considered making a formal complaint about a regulated service in the past twelve months, and of those, 54 percent used an AI tool to help them.
The file review is the part a business should care about. Evidence of AI use appeared in 51 percent of the cases examined. The age split is steep: 70 percent of under 35s used AI when pursuing redress, against 16 percent of over 55s. If your customers skew young, you are already past the halfway mark and have probably been assuming otherwise.
Financial services saw it earlier. Which? reported the Financial Ombudsman Service's own sampling, in which roughly a third of recent complaints looked generated or heavily assisted by AI. The service's chief operating officer described it as an emerging risk needing coordinated attention, and gave one detail that sticks: some responses ran past 200 pages, replying to provisional decisions that were six pages long.
Do AI written complaints actually win more often?
Yes, and that is the finding that should change how you reply. Researchers examined more than 1.1 million complaints filed with the United States Consumer Financial Protection Bureau between 2015 and 2024. Yale's writeup of the study reports that the share appearing AI generated went from essentially zero before November 2022 to nearly 10 percent by the end of the period, and that those complaints obtained relief, monetary or otherwise, about 50 percent of the time against 40 percent for human written ones, a relief rate gap of roughly ten points.
Then the authors did the thing that makes the result useful. Across three lab experiments they tested whether the advantage came from the content changing or the presentation improving, and found that editing by a language model improves overall presentation quality without altering the underlying facts. Jiwoong Shin, Minkyu Shin and Jin Kim conclude that clarity, not substance, drove the difference.
Sit with that for a second, because it cuts both ways. It means a complaint you would have dismissed when it arrived as three angry sentences is the same complaint when it arrives as four tidy pages. It also means the tidy version is not stronger on the merits than the angry one, and treating it as though it were is how a small business pays out on a claim it should have answered.
How do you tell whether a complaint was written with AI?
You mostly should not try, and the markers are worth knowing anyway because of what they predict. The Legal Ombudsman's file review named four: excessive formatting such as bold text, legal references that could not be substantiated, regulatory requirements that could not be substantiated either, and complaints that lacked a clear or practical resolution request.
That last one is the useful signal and the only one worth acting on. A complaint with no stated remedy is not a negotiation, it is a grievance looking for a shape. When you get one, the productive reply is a question rather than a defence: what outcome are you asking for. Half the time the answer is smaller than the document implied.
Resist the urge to accuse. Detection is unreliable, the accusation insults the honest majority, and the ombudsman services have gone out of their way to say AI use is legitimate. The Financial Ombudsman's own view, as reported, is that it can help vulnerable and neurodivergent complainants put an argument clearly. A business that replies "this was written by ChatGPT" has lost the moral ground and told the customer nothing about their lamp.
What do you do about a legal reference that does not exist?
Name it, neutrally, and do not answer it. This is the single highest value move in the whole exchange and almost nobody makes it.
The failure mode is quiet. A customer cites a regulation requiring a 60 day refund window. No such rule applies to you. You answer the complaint by explaining why your 30 day window was nonetheless reasonable, and in doing so you have accepted the premise that 60 days was the standard you fell short of. Every later message is now argued on their invented ground.
The alternative takes two sentences. State that the invented citation, or at least the provision cited, does not apply to this transaction, or that you cannot locate it. Then state the rule that does apply and answer under that. You are not being clever. You are refusing to let a hallucinated citation set the test.
Check before you assert. If you tell a customer their citation is invented and it turns out to be real, you have handed them the better letter. A statute number, a regulation number and a case name each take under a minute to verify against the official register, and that minute is the cheapest insurance in this process.
Replying to a long complaint without writing a long reply
Length is contagious and expensive. Four pages in does not require four pages out, and matching the volume is how a 40 pound dispute consumes an afternoon.
| What arrived | The tempting reply | The reply that works | Why |
|---|---|---|---|
| Four pages, headings, bold | Four pages back | One page, numbered to their points | Length signals nothing. Structure does, and numbering proves you read it |
| A regulation that does not apply | Explain why you complied anyway | Say it does not apply, then answer under the rule that does | Arguing inside a false premise concedes it |
| No remedy requested | Guess and offer something | Ask what outcome they want | Your guess is usually more generous than their ask |
| A deadline they invented | Rush to meet it | State your actual response time and keep it | A deadline you accept becomes the one you are judged against |
| Threat to escalate to a regulator | Treat it as a bluff or panic | Answer as though the file will be read, because it may be | Your reply is the evidence either way, and it costs nothing to write it well |
The pattern across those rows is one idea. Answer the facts of your own case in your own words, at the length your own case deserves, and let the asymmetry in page count sit there unremarked. A reader deciding this later, whether a regulator or a card issuer, is looking for whether you addressed the points, not whether you matched the volume.
Should you use AI to write the reply?
For the draft, yes, with a condition that is not negotiable. The same mechanism that makes their complaint clearer makes your answer clearer, and the asymmetry only exists while one side uses the tool and the other does not.
The condition is that every fact in your reply comes from your records rather than from the model. Order date, delivery date, what the courier scanned, what your returns page said on the day of purchase, what your messages actually promised. A model will fill any of those in plausibly if you let it, and a confident wrong date in a business reply is worse than no reply, because it is now in writing from the party with the records.
Draft with the model, fill the facts from your system, read it once as the customer and once as a regulator. If you want that loop to run on a schedule rather than in your head, that is the kind of job to hand to an agent with a verifiable output you check, with the facts supplied from your own data rather than invented.
Does any of this change your actual obligations?
None of it. Your refund window, your warranty, your delivery promises and the consumer law of wherever you sell are exactly what they were before the customer opened a chat window. That is worth saying because the tone of these documents is engineered to suggest otherwise.
What does change is the evidentiary environment. More complaints arrive in a form somebody else can read and assess, which means more of them travel further. A grievance that would have died in your inbox now has structure, a numbered list and a stated escalation path. Your written answers therefore matter more than they used to, and the sloppy one line reply you could get away with for years is now the weakest document in a file somebody else will read.
The ombudsman services are explicit that the burden of accuracy stays with the complainant. The Ombudsman Western Australia's guidance says it plainly: it prefers people to write complaints in their own words, warns that AI can lead to inaccuracies and make a complaint unclear, and tells complainants they remain responsible for the accuracy of what they submit. It also notes that decisions are always made by a person. If a complaint reaches a third party full of unverifiable assertions, that is the complainant's problem, not yours, provided your own reply is accurate.
The volume question nobody expected
The assumption was a flood. The research says something stranger. Of the people who used AI to help them decide whether to complain, 80 percent said it influenced the decision, and the split was 47 percent deciding not to complain against 32 percent deciding to.
So the tool talks more people out of complaining than into it. That makes sense once you picture the interaction: somebody describes their situation, the model explains what the rules actually say, and a fair number discover they have no case. The complaints that survive that filter are, on average, better founded than the ones that would have arrived without it.
Which reframes the whole thing. If the complaints reaching you are both better presented and better filtered, the right posture is not defensive. It is to treat a well structured complaint as more likely to be legitimate than the equivalent arriving in anger, and to resolve the valid ones faster than you used to. The same filter is at work in reviews, and reading them for the broken expectation rather than the sentiment is covered in how to answer a negative review without making it worse.
What should you change in how you log these?
One field, and it will pay for itself inside a quarter. Record, for every complaint, the remedy that was asked for and the remedy that was given. Nothing else, no sentiment score, no category taxonomy you will abandon.
The reason is that an AI written complaint arrives with its ask buried in prose, and the only way to see whether your complaint handling is drifting is to compare what people request against what you concede. Small businesses tend to discover, when they first look, that they have been settling above the ask on the long documents and below it on the short ones. That is a presentation effect, not a policy, and it is invisible until you write both numbers down.
The second thing worth logging is whether any cited rule turned out to be real. Three months of that column tells you whether invented citations are a genuine pattern in your market or a one off, and it stops you treating every future reference as suspect. Scepticism applied evenly is useful; scepticism applied to a customer who was right is expensive.
None of this needs software. A spreadsheet with four columns does it, and the support queue you already have is the source. What matters is that the record exists before the awkward case rather than after, because the pattern across twenty consumer complaints is what tells you whether your refund window is actually the problem.
Where this shows up next
Card disputes. A chargeback narrative is a complaint with money attached and a deadline you do not control, and the same drafting help is available to the cardholder. The evidence you submit is read by somebody with minutes to spend, which puts a premium on structure that we went through in assembling representment evidence that a reviewer can actually follow.
Support queues too. If a third of your inbound text is now longer and more formal than it was two years ago, the triage question changes from which message is angriest to which message contains a specific, answerable ask. That reordering is the practical work described in deciding which tickets to answer first.
What to do with the next one
Read it once for the ask and ignore the formatting. Write down the remedy being requested, or ask for it if there is none. Verify any rule, regulation or case it names, which takes a minute per item. Answer under the rules that genuinely apply, at the length your own facts need, with dates and references pulled from your own records.
Then file the exchange. The complaints that end up somewhere else are the ones where the customer felt unheard, and a clean numbered reply is the cheapest proof that they were heard. The lamp is still 40 pounds. Nothing about four tidy pages changes what you owe.