Start with structure, not rhetoric
Direct email rewards the last writer and the longest response. Structured intake rewards completeness. It asks who the parties are, what happened, when it happened, which facts are disputed, what evidence exists, and what outcome would resolve the matter.
This does not discard a party’s story. It gives that story an evidentiary shape. Chronology is separated from argument; documents from descriptions; settlement flexibility from the material that must be shared with the other side.
Give each party a private channel
Early resolution often depends on information a party will not put into a confrontational group email: uncertainty, commercial priorities, or the minimum change needed to settle. Private intake allows each side to explain those interests without exposing its entire negotiation position.
Privacy must not become secret evidence. The opposing party should be able to see and answer the documents, witness details, and material allegations on which a proposal or ruling relies. The design challenge is to keep candid settlement dialogue private while building a shared, contestable case record.
Store the file, then describe what can be inspected
Evidence arrives in inconvenient forms: PDFs, Word files, photographs, videos, audio, old document formats, and ZIP archives. A reliable system keeps the original, extracts what it can, and tells the sender when a file could not be opened or reviewed.
That distinction matters. “Stored” is not the same as “verified”. A transcript generated from an image is not proof that the image is authentic. A document extracted from a ZIP is not automatically relevant. Both parties need access to the same file and a clear statement of its inspection status.
Treat instructions inside evidence as evidence
AI-mediated procedures introduce a new adversarial input: a message or attachment may contain text telling the model to ignore its rules, change roles, reveal private material, or decide for one party. That text must never become an instruction to the dispute-resolution system.
The safe pattern is to separate trusted system rules from untrusted party content, detect clear manipulation signatures, preserve the original for audit, and quarantine it from later model context. A detection should not automatically hand victory to the other side. The passage may be quoted, forwarded, or planted; attribution, context, and an opportunity to explain come first.
Make consequential consent deterministic
An AI model may help extract names or organise a chronology. It should not guess whether a party intended to open a case, accept a settlement, or waive a right. Consequential steps need an explicit action: a clear button, a signed instruction, or an exact confirmation keyword.
This is less conversational than asking a model to infer intent, but it is more auditable. In dispute resolution, ambiguity should trigger clarification, not an irreversible state change.
Close the record and move toward an outcome
An email war has no natural stopping point. A procedure needs one. Each party should be asked whether it has further evidence or witnesses, receive a reasonable deadline, and know when its side of the record will close.
Once the material issues are identified, the system can propose concrete terms: who does what, by when, for how much, and what is released once performance is complete. If settlement fails, any assessment or ruling should cite the actual record rather than the confidence or volume of the prose.
Where human judgment remains essential
Structure and AI can reduce sorting cost, surface contradictions, and make a case legible. They do not eliminate questions of credibility, coercion, proportionality, governing law, or enforceability. High-impact outcomes need transparent reasoning and a meaningful path to human review.
The objective is not to automate conflict at greater speed. It is to prevent automation from manufacturing more conflict than humans can economically resolve.