Agentic AI in dispute resolution: what it is and the rules that apply

The short version

• Agentic AI is an emerging label for AI that is given an objective and plans, takes several steps and acts through tools (running a search or sending a message), rather than only answering a single prompt. There is no official statutory definition.
• No rule we checked lets an autonomous tool carry a lawyer’s or a decision-maker’s responsibility. The State Bar of California’s guidance says an agentic system must not make substantive legal determinations, give legal advice or file without meaningful lawyer review.
• The failure that got lawyers sanctioned over invented citations, a model confidently producing false content, is more dangerous once a system acts on it before a person checks.
• Existing conduct duties already apply: competence, confidentiality, communication, candour to the tribunal and supervision, as ABA Formal Opinion 512 sets out.
• The NIST framework expects a person to be able to override or deactivate a system acting outside its intended use, and the EU AI Act requires high-risk systems to be overseen by a person who can stop them (Article 14).
• Courts and arbitral bodies want AI use open to checking: a US federal court requires generative-AI use in briefs to be disclosed, and the SVAMC and Ciarb guidelines keep responsibility and the duty to verify with the human.^{1,2,4,5,6,7,10,11,12}

What is agentic AI, and how is it different from generative AI?

Generative AI responds to a prompt with new text, images or code, and a chatbot does this one exchange at a time. Agentic AI is given an objective and then plans and carries out several steps, acting through tools, with less step-by-step prompting from a person. Treat the term as emerging: it has no official statutory definition, and we keep claims about how far it is deployed cautious.^1,2

The EU AI Act, the European Union’s law on artificial intelligence, defines an AI system as one “designed to operate with varying levels of autonomy” that “infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments” (Article 3(1)). That is a description of autonomy, not a definition of agentic AI.²

The closest official source is the State Bar of California’s 2026 practical guidance, which says agentic capabilities “can enable systems to autonomously perform tasks or workflows without human prompting,” giving “autonomously facilitating client intake” as one example. Agentic AI there is a characterisation, not a statutory category.¹

Keeping a person in control has its own vocabulary. Human in the loop means a person approves each step before it takes effect; human on the loop means a person watches the system and can step in; human out of the loop means no one is watching. Agentic systems push work toward the second and third, which is why the rules below turn on who stays answerable.

Where is agentic AI being pointed in legal and dispute work?

In legal and dispute work the tasks pointed at agentic AI are the ones that chain steps: client intake, discovery responses, document review and iterative drafting. Whether the setting is arbitration or mediation, the limit is the same: no official source we read confirms an autonomous system deciding a case. The State Bar of California’s guidance bars letting an agentic system make substantive legal determinations, give legal advice or file without lawyer review.¹

Can an AI agent act as arbitrator or decide a dispute? None of the rules we checked lets it. A neutral, the arbitrator, mediator or other person who decides or helps resolve a case, stays a person, and the duty to verify stays with that person. We leave adoption claims out: vendor accounts of autonomous negotiation rounds or named deployments are not citable against a primary source.¹

The broad picture of AI across dispute resolution is the overview in AI in dispute resolution. What the tools themselves do is set out in AI legal tech. Institutional AI deployments and guidance are in AI in ADR and the SVAMC AI guidelines.

What goes wrong when an agent acts without a human checking it?

The core failure is confabulation. NIST, the US National Institute of Standards and Technology, defines it as “the production of confidently stated but erroneous or false content,” what is colloquially called a hallucination: output that looks right but is invented. When a system only drafts, a person can catch it; when a system acts, the error reaches the record first.⁵

NIST says “legal confabulations have been shown to be pervasive” in current large language models, and that outputs can carry “confabulated logic or citations” that look like justification but are not.⁵

Two failures matter more once a system acts through tools. Prompt injection is when someone plants instructions in data the system will read, directly or at a distance, so it behaves in a way no one intended. Automation bias, NIST’s term for excessive deference to automated systems, is the habit of deferring to a machine’s output. NIST treats automation bias under Human-AI Configuration and prompt injection under Information Security.⁵

The concrete consequence is in the sanctions record. In Mata v. Avianca, Inc., a New York federal court found “nothing inherently improper about using a reliable artificial intelligence tool for assistance” but fined two lawyers and their firm US$5,000 under Rule 11, the US federal rule that lets a court penalise a baseless filing, for a brief built on non-existent opinions, after they stood by the cases when the court questioned them. The full, dated record of courts sanctioning AI-invented citations is in the AI hallucination sanctions record; the arbitration-interface cases are in AI arbitration case law.⁷

Who is responsible when an AI agent gets it wrong?

The human stays responsible. Every rule we checked keeps a named person answerable for what is filed or decided, and none lets “the AI did it” stand as an excuse.^1,7

Mata treated the invented citation as the lawyers’ and the firm’s failure under the existing duty of reasonable inquiry, not the tool’s. The State Bar of California’s limits bind the person who deploys the system: no substantive legal determinations, no legal advice and no filings by an agentic system without meaningful supervision and review. This is the rule of non-delegation, that a professional cannot hand the core of the job to a machine.^1,7

Where an AI service is sold straight to the public, consumer-protection law applies too. The US Federal Trade Commission’s order against the “robot lawyer” DoNotPay, Inc. prohibits it from representing, without competent and reliable evidence, that its service operates like a human lawyer, and required US$193,000 in monetary relief, for deceptive claims, not for the unauthorized practice of law. Whether a legal chatbot crosses into the unauthorized practice of law, doing a lawyer’s work without a licence, is a separate question we take up in AI legal advice chatbots and the unauthorized practice of law.¹³

What do the professional-conduct and institutional rules already require?

The existing conduct rules already cover anyone deploying AI, agentic or not. ABA Formal Opinion 512 (29 July 2024), the American Bar Association’s first formal ethics opinion on generative AI, applies the duties of competence (Model Rule 1.1), confidentiality (1.6), communication (1.4), candour to the tribunal (3.3), meaning not misleading the court, supervision (5.1 and 5.3) and reasonable fees (1.5). It is persuasive, not binding.⁶

Under 512, a lawyer needs a reasonable understanding of a tool’s capabilities and limitations, not expertise; how much checking is enough depends on the tool and the task; output must be reviewed before it goes to a court; and informed client consent is needed before information about a representation goes into a self-learning tool, which general engagement-letter wording does not cover. Managers must set clear AI policies and train their staff.⁶

The NIST AI Risk Management Framework gives the risk vocabulary and the governance structure. It organises the work around four functions: govern, map, measure and manage. Its Generative AI Profile expects systems that can be deactivated, and controls that let a person supersede, disengage or deactivate a system whose outputs are “inconsistent with intended use,” and it tells users to review and verify sources and citations. These are voluntary and bind no one by themselves.^4,5

Dispute-resolution bodies’ own guidance keeps responsibility with the person. The SVAMC Guidelines (1st edition, 2024) from the Silicon Valley Arbitration & Mediation Center and the Ciarb Guideline (2025, updated September 2025) from the Chartered Institute of Arbitrators are soft law: they bind only where the parties agree or a tribunal adopts them. SVAMC’s Guideline 6 says an arbitrator “shall not delegate any part of their personal mandate to any AI tool.” The detail is in the SVAMC AI guidelines and AI in ADR.^11,12

For the US state statutes and the EU AI Act in full, we hand off to US state AI laws: Colorado and Texas and the EU AI Act and dispute resolution.

Must you disclose agentic AI, and keep a person able to stop it?

Increasingly, yes. A US federal court already requires generative-AI use in briefs to be disclosed, and the EU AI Act requires high-risk systems to be built so a person can oversee, override and halt them.^2,10

The US District Court for the Northern District of Texas, under Local Civil Rule 7.2(f) (effective 1 September 2026), requires a brief prepared using generative AI to say so on page one, under the heading “Use of Generative Artificial Intelligence”; filing without that statement certifies that no generative AI was used.¹⁰

Article 14 of the EU AI Act requires that high-risk AI be “designed and developed in such a way … that they can be effectively overseen by natural persons.” A high-risk AI system is one the Act singles out for the strictest duties because of what it affects. The person overseeing it must be able to stay aware of the tendency to over-rely on the output (automation bias), to “disregard, override or reverse” the output, and to stop the system through a “stop” button. The organisation that uses the system under its own authority is, in the Act’s words, the deployer. The high-risk duties for administration-of-justice and alternative dispute resolution (ADR) uses apply from 2 December 2027, and the classification is covered in the EU AI Act and dispute resolution.^2,3

Separately, US rule-makers are weighing how courts should treat machine-generated and deepfake evidence, a deepfake being a convincing fake image, audio or video made with AI. Proposed Federal Rule of Evidence 707 is a proposal, not in force as of 10 November 2026, and we cover it, and the draft deepfake rule, in AI-generated evidence and deepfakes.¹⁴

The State Bar of California’s guidance closes the loop on autonomous action: no autonomous filing or communication with a court, and no sending client information out without safeguards and human review.¹

What does this mean for anyone using AI in a dispute?

In our view, the striking thing about agentic AI in dispute resolution is how little new law it needs. Read across ABA Formal Opinion 512, the NIST AI Risk Management Framework, Article 14 of the EU AI Act, the Northern District of Texas disclosure rule and the SVAMC and Ciarb arbitration guidelines, and the same three requirements keep surfacing, whatever the jurisdiction. A named human stays answerable for what is filed or decided, and no instrument we read lets an autonomous tool carry that responsibility. AI use that touches the record or the outcome has to be open to checking, through disclosure, oversight or an audit trail. And a tool’s output has to be verified before anyone relies on it, because, as Mata v. Avianca showed, a confident citation can be wholly invented. The pace of AI is ahead of the rulebooks, so the safest reading is that the duties already on the books, competence, candour, supervision, human oversight, verification and consumer protection, already govern agentic AI, and the newer rules are making that explicit.^{2,4,5,6,7,10,11,12}

Related: AI in dispute resolution, AI in law, the AI hallucination sanctions record, AI legal advice chatbots and the unauthorized practice of law, US state AI laws: Colorado and Texas, the EU AI Act and dispute resolution and the SVAMC AI guidelines.