Last updated: October 2026 · Reading time: about 10 minutes
What this guide covers
What "agentic" actually means, and what it doesn't
The distinction that matters is between a system that produces output and a system that takes actions.
A generative AI tool produces output. You give it a prompt — "summarise these medical records" — and it returns a summary. The summary sits there until a human does something with it. If the summary is wrong, the error stays in one place and someone notices.
An agentic system takes actions. You give it a goal — "process this intake form and open a matter" — and it decides the intermediate steps: read the form, extract the crash date, look up the statute of limitations for that jurisdiction, calculate the deadline, create the matter record, request records from the providers listed, assign the matter to a paralegal, and set a follow-up reminder. It does all of that without asking between steps.
That capability is real, and it is genuinely useful. It is also where the risk profile changes completely, which is the subject of the next section.
Where agentic AI is genuinely deployed in 2026
Setting aside vendor marketing, agentic workflows in US plaintiff firms are concentrated in four areas where the steps are well-defined and the success criteria are checkable.
| Workflow | What the agent does | Maturity in 2026 | Main risk |
|---|---|---|---|
| Intake triage | Reads inquiry, extracts crash details, checks against case-acceptance criteria, routes to human intake or declines | Widely deployed | Premature decline of a valid case |
| Records follow-up | Tracks outstanding record requests, sends reminders, logs responses, flags non-responders | Widely deployed | Low — easily reversible, low stakes per action |
| Deadline calculation | Calculates statutes of limitation and procedural deadlines from matter facts and jurisdiction rules | Deployed, needs verification | High — a wrong deadline can end a case |
| Document assembly | Assembles demand packages, discovery responses, and standard filings from matter data | Deployed, needs attorney review | Medium — errors caught at review if review happens |
| Full case handling | End-to-end from intake to demand | Marketing ahead of reality | Compounding errors across every stage |
The pattern is that agentic tools work well where each action is independently checkable and poorly where a single wrong fact propagates through the whole chain. Records follow-up is close to ideal — if the agent sends a reminder to the wrong provider, someone notices and fixes it. Deadline calculation is at the other end — if the agent misreads the jurisdiction or the accident date, nobody notices until the deadline passes.
The compounding error problem
This is the single most important concept for anyone deploying agentic AI in a law firm, and it does not appear in most vendor materials.
Consider a generative tool drafting a demand letter. If it misstates a treatment date, the attorney reads the letter, notices the date is wrong, and corrects it. One error, one catch.
Now consider an agentic workflow: the agent reads the intake form, extracts the crash date, calculates the statute of limitations, opens the matter with that deadline, assigns it to a paralegal, and sets an automated reminder. If the agent misread the crash date at step one, every downstream action is built on the wrong fact — and none of them looks obviously wrong. The deadline looks like a normal date. The reminder fires at a normal interval. The paralegal sees a normal case. The error only surfaces when someone independently checks the crash date against the intake form, which nobody does because the system already "handled" it.
This is not a hypothetical. It is the predictable behaviour of any system that chains automated steps on extracted facts. It is also the reason agentic tools need verification points, not just review at the end.
Where human checkpoints must sit
The practical answer is not "review everything" — that eliminates the time savings that justified the tool. The answer is to put human verification at the points where a wrong fact becomes hard to undo.
| Checkpoint | What a human must verify | Why here |
|---|---|---|
| Before matter creation | Crash date, jurisdiction, party names, injury description against the source intake document | Every downstream action depends on these facts |
| Before any deadline is set | Statute of limitations calculation against the actual jurisdiction rule | A wrong deadline is unrecoverable |
| Before any external communication | Recipient, content, and any factual claims about the case | Sent messages cannot be recalled |
| Before any filing | All factual assertions and every citation | Professional responsibility attaches to filings |
| Before case decline | The decline decision itself, by a human with authority | Declining a valid case is a lost client, not a corrected error |
In practice this means the agent can act freely on records follow-up and internal routing, but must stop for a human before setting a deadline, sending external correspondence, filing, or declining a case. Those are hard stops in the workflow configuration, not suggestions in a policy document.
ABA Formal Opinion 512 and supervision
ABA Formal Opinion 512 addresses lawyers' use of generative AI and sets out duties around competence, confidentiality, communication, and fees. It does not specifically address agentic AI — the opinion predates the current wave of agentic tools — but the underlying duties apply directly, and in some cases more sharply.
- Competence now includes understanding what your agentic tools actually do, including where they can act without confirmation. "The vendor set it up" is not a defence.
- Confidentiality requires knowing which systems receive client data and under what terms. Agentic tools often touch multiple systems, which multiplies the surface area.
- Communication requires clients to understand material aspects of their representation, including the use of AI where it affects their case.
- Supervision is the sharpest issue. A firm that deploys an agent to calculate deadlines and does not verify the output has not supervised it. The duty to supervise non-lawyer assistants applies to software agents acting on the firm's behalf.
What's real vs what's marketing
Because "agentic" is the most over-used term in legal tech in 2026, it is worth being blunt about the gap between marketing and deployment.
| Claim | Reality in 2026 |
|---|---|
| "Handles an entire case end to end" | No. Individual workflows are automated; the case is not. |
| "Replaces paralegal work" | It replaces specific repetitive tasks. It does not replace the judgement those tasks required. |
| "Learns your firm's style" | Some tools can be configured with firm templates. That is configuration, not learning in any meaningful sense. |
| "Attorney oversight built in" | Usually means a review screen exists. Whether your firm actually uses it is a workflow decision, not a product feature. |
| "No hallucination risk" | No current system eliminates it. Retrieval systems reduce it; nothing eliminates it. |
How to evaluate an agentic tool
Ask these questions before signing, and ask for written answers where the stakes are high.
- What can the agent do without human confirmation? Get a specific list of actions. "It depends on configuration" is the correct answer — but you need to know the full set of possibilities.
- Where are the hard stops? Can your firm configure mandatory human approval before external communication, filing, deadline-setting, or case decline? If not, that is a serious limitation.
- What is the audit trail? For every action the agent takes, can you see what it did, when, based on what input, and which human approved it? An agentic tool without an audit trail is ungovernable.
- How does it handle conflicting facts? If the intake form says one crash date and the police report says another, what does the agent do? It should stop and flag, not pick one silently.
- What happens on error? If the agent takes a wrong action, how is it reversed, and who is notified? Test this with a real example before signing.
- Who are the subprocessors, and are they covered by the BAA? Agentic tools often chain multiple model providers. Every provider that touches PHI must be covered.
- Does it train on your data? Get "no" in writing.
- What is the off-switch? If something goes wrong, can your firm pause the agent immediately without losing the matter data? A tool you cannot turn off quickly is a tool you do not control.
For the compliance side of these questions — BAA terms, SOC 2 reports, subprocessor lists — see our guide to HIPAA and SOC 2 compliance for legal AI tools.
Frequently asked questions
What is agentic AI in a legal context?
It is an AI system that takes multi-step actions toward a goal rather than just producing text — for example, reading an intake form, calculating a deadline, creating a matter, and requesting records without a human clicking between each step. The distinguishing feature is that it can change records and trigger downstream actions.
Is agentic AI actually being used in US plaintiff firms in 2026?
Yes, but narrowly. Intake triage and records follow-up are widely deployed. Deadline calculation and document assembly are deployed with human verification. Genuine end-to-end case handling is largely marketing rather than reality.
What is the compounding error problem?
An agentic error is silent and self-consistent. A wrong fact extracted at step one propagates through every downstream action, producing a record that looks internally coherent. The error is only visible if you compare the final state against the original source document — which is why verification points must sit at the start of the chain, not the end.
Does ABA Formal Opinion 512 cover agentic AI?
The opinion was written before the current wave of agentic tools, so it does not address them specifically. Its underlying duties — competence, confidentiality, communication, and supervision — apply directly, and the supervision duty is arguably sharper for tools that can act without confirmation.
Can an AI agent calculate a statute of limitations deadline?
It can attempt to, and some tools do. A human must verify the calculation against the actual jurisdiction rule before the deadline is set in the matter record. A wrong deadline is unrecoverable, which makes this one of the points where a hard human checkpoint is essential.
How do I know if a tool is genuinely agentic or just a chatbot?
Ask whether it can change a record, send a message, open a matter, or set a deadline in a system of record without a human clicking between steps. If it can only produce text for a human to act on, it is generative rather than agentic, regardless of how it is marketed.
What happens if an AI agent makes a mistake in a law firm?
The firm is responsible. Software agents acting on the firm's behalf fall under the duty to supervise non-lawyer assistants. This is why audit trails, hard stops, and a fast off-switch are governance requirements rather than nice-to-have features.
Will agentic AI replace paralegals?
Not in the foreseeable future. It automates specific repetitive tasks — records follow-up, standard document assembly, intake routing — but the judgement those tasks required still has to sit somewhere. Firms deploying agentic tools are shifting paralegal capacity toward client communication, case strategy support, and quality review of agent output.
Related reading on Dedon Tech: HIPAA and SOC 2 compliance for legal AI tools · best AI legal software for US personal injury firms · full site map
0 Comments: