A Human-Accountable Arbitration Workflow: Synthetic Process Model

How can a hypothetical arbitration workflow keep human decisions accountable?

Give every consequential transition an owner, an input record and an observable result. Preserve the distinction between a party's assertion, a source document, a tool summary and a human finding. A workflow should make corrections and objections visible before the responsible decision maker relies on disputed material.

Use states that describe evidence, not confidence

The proposed model has six states: received, organized, summarized, challenged, reviewed and decided. Each state answers a different question. Received means the material arrived; reviewed means a specified person performed a stated check. Neither label alone means the underlying assertion is true.

• Received: preserve original document and receipt metadata.
• Organized: assign a stable exhibit identifier without changing the original.
• Summarized: attach source locations and unresolved ambiguities.
• Challenged: record a party's identified objection.
• Reviewed: identify the human response and any correction.
• Decided: distinguish the decision maker's findings from the tool's earlier draft.

Synthetic record: two invoices and a disputed delivery

The fictional record contains E1, an invoice dated April 4; E2, a delivery email dated April 6; and E3, a customer message saying the package was incomplete. The agent's first summary says delivery was completed on April 6. That statement overreads the email and ignores E3.

A corrected summary says the supplier reported a delivery on April 6 and the customer disputed completeness. It links both statements. The human decision maker then considers the relevant agreement and evidence. The correction is a better description of the record; it does not itself resolve the dispute.

• Original tool assertion: complete delivery on April 6.
• Challenge: E3 disputes completeness.
• Correction: preserve both attributed positions.
• Human issue: what was actually delivered and what the governing obligation required.

Assign a review task at each transition

A case administrator can check that a file is readable without deciding its admissibility or weight. A representative can verify a quotation without agreeing with the opposing party's interpretation. The process record should say which check occurred.

SVAMC Guideline 7 addresses tribunal reliance on AI-generated information outside the record and opportunities for party comment. This model uses an explicit external-information flag to make that issue visible. [S1]

• External-information flag: identify material not supplied in the case record.
• Objection route: who receives a challenge and when it is considered.
• Correction scope: identify summaries and drafts affected by an error.
• Decision record: the human's reasoning and disposition, subject to the applicable procedure.

Observe completion without overstating it

For the synthetic model, complete means the responsible human has addressed the identified issues and the approved result has been recorded through the defined process. A successful export proves only that a file was produced. A delivery receipt proves only the delivery event it records.

AAA's published process description retains human responsibility for deciding the outcome and issuing the award. [S2] The broader lesson for workflow design is to make that human step distinguishable from automated preparation.

• Check the final document against the approved version.
• Record the actual distribution event and recipients.
• Preserve unresolved procedural questions instead of marking them silently complete.
• Apply the governing retention and confidentiality requirements to the record.

Scope and limitations

An annotated fictional process model. It is not an adopted procedure, a real arbitration or a statement that this sequence satisfies any jurisdiction's law.

Every exhibit and event is synthetic. The model has not been validated as a compliant arbitration procedure or used to decide a real dispute.

Sources

1. SVAMC Guidelines on the Use of Artificial Intelligence in Arbitration, First Edition (2024) — Principle-based guidelines covering confidentiality, disclosure, diligence, non-delegation and due process; adoption is case-specific. Checked 2026-11-05
2. AAA AI Led Rules, Forms, and Fees — Provider describes AI assistance and states a human arbitrator decides the case and issues the final award. Checked 2026-11-05

Explore further

Explore the central Agentic Attorney project at https://agenticattorney.com/ for the broader legal-workflow context. This process model offers concrete questions about evidence and human review without implying arbitration administration by that project. Explore Agentic Attorney.