Finding the complaint inside an AI-drafted submission

The submissions investigators now receive

You open a complaint about a boiler installation and find eleven pages under numbered headings. It cites the Consumer Rights Act 2015 and asks for punitive damages, a remedy the scheme cannot award. Halfway through, between a section on statutory interest and another headed “Data protection breaches”, the customer mentions that they were promised a callback three times and no one rang.

Several ombudsman schemes have said in public that a growing share of their complaints are drafted with AI. For a team carrying a backlog, the temptation is to treat the length as the problem and the tool as a sign of a weaker complaint.

I think that gets it the wrong way round. In my years running dispute resolution teams, the people who wrote at length were often the ones who felt least heard. Someone passed between departments for two months, or writing in a second language, has good reason to reach for a tool that makes them sound like a lawyer the firm will take seriously. The tool supplies that register and adds grounds of its own, such as a discrimination point, and people leave them in because they assume the tool knows better, whether or not anything of the kind happened.

Tribunals have less patience. This month the Employment Appeal Tribunal called a 300-page ChatGPT-generated skeleton argument “entirely unacceptable”. An ombudsman scheme exists so that people can complain without a lawyer, and criticising how someone asked for help would work against that purpose.

Going back to the person’s own words

Fairness requires an investigator to engage with what the person is complaining about, and length works against that. An investigator facing fourteen numbered grounds will be inclined to take each in turn. A decision written that way can answer every ground and still treat the unreturned calls as point nine of fourteen, and the customer reading it will conclude, with some justice, that no one listened.

The clearest account of the grievance is often what the person wrote before asking a tool for help. The respondent’s file holds the first contact, perhaps a webchat transcript or a call note an agent logged at the time, and the person may have filled in the free-text box on the scheme’s form without help. I would read those before the long submission and write down, in a sentence or two, what went wrong and what the person wants put right.

Then separate the issues raised from the issues the scheme can consider. Some grounds fall outside remit on their face, such as a claim under legislation the scheme does not apply. Others are new to the respondent. Most schemes consider a matter once the firm has had the chance to answer it, and points that first appear in an AI-drafted submission have often never been through the firm’s complaints process. Those go back to the firm, and the person should hear that at the start, in plain terms.

Agreeing the scope at intake

The most useful thing a scheme can do with a long AI-drafted complaint, in my view, is pick up the phone. A twenty-minute conversation about what happened and what the person wants will often bring fourteen grounds down to the two or three they care about. The investigator then writes to confirm the scope, listing the issues the scheme will look at and those outside its remit, with the reason for each. The person can correct the letter before the investigation starts, and from then on the decision answers the complaint it describes.

Firms can respond in kind, answering every paragraph, sometimes with their own drafting tools, and the investigator then reads two long documents to find one missed callback while the case ages against its timeliness target. Asking the firm to respond to the agreed scope is fairer to the firm and keeps the file to the size of the dispute.

Narrowing a complaint has its own risk. A scheme that trims too hard ends up deciding a different case from the one the person brought. The protection is the person’s agreement to the scope, recorded on the file, and an investigator willing to reopen it if the evidence later shows it was wrong.

Using AI to read AI-drafted complaints

Schemes will be tempted to answer AI with AI, and I understand why: a tool that reads a long submission and returns a list of issues is attractive when cases are waiting. The danger is that the person’s grievance passes through two machines and no human reads it in their words. An investigator who accepts a generated issue list without checking it has handed over the most important judgement in the case, deciding what the complaint is.

A generated issue list can be acceptable on two conditions. Each issue has to point to the page and paragraph where the person raised it, so the investigator can test the summary against the original. And an investigator, named on the file, has to confirm the list, amending it or striking points out, with a record of who checked what. Without both, the scheme cannot show either party how it decided what the complaint was about.

Those are the conditions we build to at Ctrl AI: each issue the platform lists against a scheme’s configured remit is cited to the place the person raised it, and nothing goes forward to assessment until an investigator has confirmed or amended the list.