AI-assisted complaints: what the Ombudsman’s 54% finding means
New Legal Ombudsman research separates AI use from complaint quality. What the samples show, and how firms can keep the facts and requested remedy clear.
The brief
The Legal Ombudsman’s 1 October research announcement reports AI use by 54% of people who had made or considered a complaint about a regulated service in the previous year. That is not a finding that AI wrote 54% of all legal complaints. The study combines a public survey with separate reviews of complaint files.
Our take is to preserve the person’s account, the supporting evidence and the outcome they seek as distinct parts of the record. An AI-generated summary can help organise a complaint, but it should remain traceable to the original. A polished letter should neither receive extra credibility nor be dismissed simply because it appears machine-written.
The full record
A complaint may arrive as a fluent, formal letter and still leave the reader uncertain about what happened. The practical question is whether the firm can recover the underlying issue and respond to it. The Legal Ombudsman’s new research and guidance, announced on 1 October 2026, give firms reasons to focus on that task.
Two different kinds of evidence
The research report, pages 5–6, describes an Ipsos survey of 2,196 people aged 16–75 in England and Wales, conducted on 13–17 March 2026. Of these, 478 had made or considered a formal complaint about a regulated service during the previous year. The 54% figure applies to that subgroup. It covers help considering a complaint as well as making one, across regulated services rather than legal services alone.
Separately, LeO reviewed 100 randomly selected legal complaint files. It identified indicators of AI use in 51. Those indicators are an assessment method, not confirmed authorship. In that subset, reviewers judged the apparent impact positive in 17 cases, neutral in 16 and negative in 18. The report also contains a separate 100-file conveyancing review; the two exercises should not be collapsed into a population-wide rate. These figures describe the reviewed files, not a controlled test of whether AI causes better outcomes. Source: report, pages 7–9.
What the guidance asks firms to address
LeO’s service-provider guidance describes both the potential to clarify a complaint and the risk of exaggerated or inaccurate correspondence. It cautions against responding in a way that escalates the exchange. Its consumer guide asks people to check facts, protect personal information and focus on a practical resolution.
Our analysis: make the complaint recoverable
A useful complaint record should allow a colleague to answer four questions: what service was expected, what the client says went wrong, which evidence supports or disputes that account, and what outcome the client wants. Keep the original message even when a summary is prepared. Link each material statement in the working summary to a document, message or confirmed account.
If the requested remedy is unclear, ask a short clarifying question. Record the answer beside the original request. Treat disputed facts as disputed, and identify who must investigate them. A long chain of replies can otherwise make an assumption look like an agreed fact.
For an internal trial, use fictional complaint files with deliberately conflicting dates or an unsupported legal reference. Ask reviewers to find the service issue, distinguish evidence from assertion and identify the unresolved questions. Measure omissions, corrections and the time needed for a second person to understand the file. Do not measure success only by how quickly a reply was drafted.
This is a proposed test, not a benchmark we have performed. The value of formal communication here is a record that keeps the client’s words, the firm’s findings and the eventual resolution connected. Our conveyancing workflow analysis applies the same test to instructions and approvals.
We read the research announcement, report and both guidance pages. Survey results and case-file assessments are separate evidence; apparent AI use is not proven authorship. We have not replicated the research or measured a product’s performance. The proposed internal trial is our analysis.