Dentons’ AI correspondence research: preserve the claim behind the prose

The brief

Dentons’ 26 October research announcement describes legal teams receiving AI-generated complaints, grievances and data-protection requests. Interviewees reported difficulty identifying substantive issues within lengthy submissions. The evidence comes from interviews involving Irish and international businesses, not a representative measure of all legal departments. Read Dentons’ announcement.

Our analysis focuses on intake. Preserve the original submission, identify each requested action and connect every extracted claim to its source passage. Assign a person to assess the applicable deadline and approve the response. A polished document can contain a valid complaint, an unsupported assertion or both. The task is to examine those elements fairly; suspected AI use should not become a substitute for considering what the sender is asking.

The full record

Dentons published the findings on 26 October. Its account describes teams using AI for bounded tasks while retaining human legal judgement. The release reports interview themes; it does not establish the prevalence of AI-generated submissions across the UK or measure the accuracy of an automated intake system. Dentons’ research announcement.

Our analysis: separate receipt from interpretation

An intake summary is an interpretation of a submission. Preserve the original message and attachments before creating that summary. Keep the date received and the version used. If the sender later supplies a correction, retain the relationship between the versions instead of overwriting the first account.

A useful summary should distinguish what happened, what the sender alleges, what evidence they supplied and what outcome they request. Those are different categories. A statement that an organisation admitted an error should remain a reported allegation until the cited material has been checked.

Make each extracted claim traceable

For every material claim, include a short reference to the passage or attachment supporting the extraction. If the summary adds an inference, label it. If an attachment is missing or a cited authority cannot be found, record that unresolved point for the reviewer. Do not turn a missing source into an assumption that the underlying grievance is false.

Assign deadline assessment to a named person who can determine the relevant process and law. This article offers no universal response period. Receipt, acknowledgement, investigation and final response should remain distinct stages, with an owner for the next action.

Test the intake before trusting the summary

Use a fictional long submission containing one important request, a repeated allegation and a reference to an attachment that was never supplied. Ask a reviewer to compare the summary with the original. Can they locate the request, recognise the repetition and see the missing evidence? Record omissions and unsupported additions separately.

The review should also check whether the proposed response answers the sender’s actual request. A shorter answer can still fail if it responds to a tidy summary rather than the substance. This is a suggested evaluation method, not a completed benchmark or a claim that any particular product passes it.

The practical conclusion

Measure whether a colleague can reconstruct the intake decision from the original submission, the extracted issues and the approved response. That gives a legal team something concrete to inspect. It also makes it easier to correct a mistaken interpretation without losing the record of what was received.

Read Dentons’ original announcement dated 26 October 2026, included as a catch-up item from the 2 November briefing. This article is based on the announcement, not an independently inspected complete interview dataset. No prevalence, causal productivity or detection-accuracy estimate is inferred. UK-specific representation is not asserted. The intake design and test are original analysis.

1. Dentons: research on AI-generated submissions, 26 October 2026