Built-In or Built for You? How Plaintiff Firms Should Use AI Agents
Every plaintiff firm has work that looks the same from one firm to the next, and work it does its own way. Choosing AI for plaintiff law firms comes down to how a platform handles both, and who checks the result before it leaves the office.
What should a plaintiff firm decide before choosing an AI agent?
A firm has three decisions to make: which work can run on agents as they ship, which needs an agent built around the firm's own method, and how much of either should run without attorney review. This guide takes each in turn, then covers what ABA Formal Opinion 512 says about supervising the result.
The Decision Hiding Inside the Demo
A platform demo is built to show output: a chronology assembled from a few thousand pages, a demand draft pulled from the record. Those outputs matter. They rarely answer the question firm leadership faces a few months after signing, though, which is whether the agents follow the firm’s method or ask the firm to follow theirs.
That question has two parts. The first is scope: which work an agent can take on as it ships, and which work depends on how one particular firm does things. The second is control: which tasks an agent should finish on its own, and which should stop and wait for an attorney.
Both parts depend on what an agent is. A traditional AI feature waits for a request and returns an answer. An agent carries out a defined sequence of work on a case, draws on what’s already in the file, and hands the result to an attorney or other designated reviewer at the point where judgment is required. How Anytime AI 3.0 Coordinates Agents Across a Case covers that difference in depth; this piece focuses on the buying decision.
What AI Case Preparation Software Handles on Day One
A large share of what plaintiff firms want to automate is common across firms. Turning medical records into a chronology, drafting a demand grounded in the case facts, reviewing a high-volume discovery production and summarizing depositions look broadly similar from one firm to the next. Built-in agents are designed for that shared work, and they’re usable without a build project.
The useful test for a built-in agent is how well its defaults fit the practice area. In a nursing home case, a chronology that lists dates and providers is less useful than one that lines up what the care plan called for against what the chart shows was done. That’s why it matters who shaped the defaults. Anytime AI’s nursing home litigation workflows were co-designed with trial lawyers in that practice, including Ernest Tosh, who spent more than 100 hours shaping the platform.
Built-in drafting agents can also work from a firm’s own templates, so a demand letter or discovery response comes back in the firm’s format rather than the vendor’s. For firms looking to automate plaintiff case preparation, this is often where the volume sits: records in, chronology out, then a first draft an attorney edits. The medical chronology walkthrough shows what that first step involves and where an attorney stays in the loop.
Built-in agents are more useful when they work from the case rather than from a blank prompt. Anytime AI’s agents keep case-level memory, so an agent drafting a demand can work from what’s already been established in the file, such as the treatment history and prior findings, instead of each task starting over. That matters because a case isn’t a series of unrelated prompts: the more an agent can work from what has already happened in the file, the less the team has to rebuild that context for every task.
When a Firm Needs Its Own Agent
Built-in agents know the practice area. They don’t know the house rules that make one firm’s work different from another’s: the institutional knowledge that usually lives in a senior attorney’s head. A custom agent turns that knowledge into a repeatable step anyone at the firm can run, and it becomes worthwhile when the step recurs on every case in a way no vendor default would guess.
Common signals include:
How a senior partner screens a new matter, applied the same way to every intake
What must be in the file before a demand goes out
Which discovery gaps the firm won’t accept, and what triggers a follow-up request
How the firm organizes medical evidence for its own case memos
What gets escalated to an attorney, and at what point
The common thread is repetition plus specificity. A one-off task isn’t worth an agent. A step the firm does identically across dozens of cases, in a way that reflects its own judgment, often is.
There’s a cost on the other side. Someone has to describe the method precisely enough for an agent to follow it, and someone has to check the first results against how the firm actually does the work. If three attorneys run the same checklist three different ways, the agent will encode one of them, so the firm has to settle which one first.
Built-In vs. Custom AI Agents at a Glance
Built-in agent | Custom agent | |
|---|---|---|
Handles | Work common across plaintiff firms | A step specific to one firm |
Knows | Practice-area patterns | Firm-specific rules, criteria and checklists |
Setup | Ready to use | Defined, tested and maintained |
Adapts to | Supported templates and workflows | The firm’s own process and decision points |
Best for | Chronologies, demand drafts, discovery review | Intake screens, pre-demand checks, escalation rules |
Who Builds Custom AI Agents for Law Firms?
There are two routes. A firm can build its own agents, which suits firms with someone on staff who can spell out a process step by step and test the result. Or the vendor can build the agent with the firm, which suits firms whose method lives in a senior attorney’s head and who would rather describe it than configure it. Anytime AI supports both approaches.
Whoever builds the agent, the question after launch is maintenance. Methods change when case law shifts, a carrier changes its review process, or the firm adds a practice area. A useful thing to settle before any build is who updates the agent when the firm’s method changes, and how quickly.
Who builds it doesn’t change who answers for it. That’s where the ethics guidance comes in.
Automatic or Supervised: Setting the Mode
Some agents can run end to end without a person in the loop. Others work under attorney supervision, pausing for review or taking direction at defined points. The choice is less about the agent than about the consequence of an error.
Work that stays inside the firm and is easy to check, such as organizing and tagging an incoming record set or producing a first-pass summary, is a reasonable candidate for automatic runs. Work that leaves the firm or rests on judgment, such as a demand sent to a carrier, a discovery response served on opposing counsel, or a valuation shared with a client, should stop for attorney review.
ABA Formal Opinion 512 offers a practical way to move a task from one column to the other. It describes a lawyer using an AI tool to summarize a large set of contracts: if the lawyer first tested the tool on a smaller subset, compared its summaries against a manual review and found them accurate, the lawyer would not necessarily need to review every document by hand afterward. Applied to agents, that suggests a sequence: run the agent supervised on a sample of real files, compare its output with the firm’s own work, and loosen review only where the results hold up.
The same opinion draws a hard line on the other end. Lawyers may not leave it to AI tools alone to give clients legal advice or to negotiate their claims. For a plaintiff firm, that puts settlement negotiation firmly in the supervised column no matter how good the agent’s draft is.
Does ABA Opinion 512 Cover AI Agents?
The ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512 on July 29, 2024. The opinion addresses generative AI tools in general, meaning tools that produce new content such as text in response to a prompt, and it does not address AI agents specifically. Its guidance is still relevant to firms using agents that generate, summarize or analyze legal work, and three parts of it bear directly on how a firm deploys them.
Supervision. The opinion applies Model Rules 5.1 and 5.3, which govern managerial and supervisory lawyers. Managing lawyers must set clear policies on the firm’s permissible use of AI, and supervising lawyers must make reasonable efforts to ensure the firm’s lawyers and staff follow their professional obligations when using it, including training on the tools they use. When a firm relies on an outside provider, the opinion applies the vendor diligence expected under Rule 5.3(b): checking credentials and references, understanding the provider’s security policies, using confidentiality agreements, and finding out whether the tool keeps the firm’s information after the relationship ends. Those considerations apply whether the firm builds its own agents or works with an outside provider.
Confidentiality. Under Model Rule 1.6, the opinion warns that self-learning tools can surface one client’s information in response to prompts on another matter. Before a lawyer puts client information into such a tool, the client’s informed consent is required, and the opinion says boilerplate language in an engagement letter isn’t enough. This is why a vendor’s training policy is an ethics question, not only a security one. Anytime AI maintains zero data training, meaning client files are never used to train models; the security page sets out the controls.
Fees. Under Model Rule 1.5, the opinion separates AI costs that are ordinary overhead from those that may be passed to a client as an expense, and it bars billing clients for time spent learning a tool the lawyer will use regularly. It also says the same reasonableness factors apply to flat and contingent fees, which is worth raising with ethics counsel for firms whose economics depend on contingency work.
ABA formal opinions interpret the Model Rules; they don’t bind any state on their own. Several state bars have issued their own AI guidance, and a firm’s state rules and opinions control where they differ.
This article provides general information, not legal or ethics advice. Firms should consult their state’s rules and ethics counsel about their specific use of AI.
What AI Agents Still Don't Replace
An agent can assemble a chronology, draft a demand and flag inconsistencies across a production. It doesn’t decide what a case is worth, whether to accept an offer, or which theory to take to a jury. Those calls stay with the attorney, and Opinion 512 is explicit that the lawyer remains fully responsible for the work regardless of how much review the lawyer chooses.
An agent also doesn’t replace a retained expert. It can surface where the record suggests a standard of care problem, but it doesn’t supply the opinion testimony a case may need. And its output still needs verification. The practical safeguard is citations back to the source page, so a reviewer can check a finding against the record rather than taking it on trust.
Some capabilities also depend on the platform. For example, Anytime AI handles document compilation but doesn’t currently assemble image exhibits. A custom agent is also only as good as the method it’s given: if the firm’s process is inconsistent, the agent will reproduce the inconsistency faster.
10 Questions to Ask an AI Litigation Support Platform
These questions apply to any vendor, Anytime AI included. They’re written to separate legal AI for plaintiff firms that adapts to the firm’s method from tools that expect the firm to adapt.
Which agents are ready for our practice areas at launch, and what does their output look like on a closed case like ours?
Can your drafting agents work from our templates?
Can we customize agents without replacing the workflows we already use?
Who builds custom agents: our staff, your team, or both? Who updates them when our method changes?
Can multiple agents work from the same case context, or does each workflow run separately?
Which tasks can run automatically, and where can we require attorney review before anything leaves the firm?
What happens when an agent encounters something outside its instructions?
Is our data used to train models, and what happens to it if we end the contract?
Does “SOC 2” in your materials mean an audited report or alignment with the framework?
Does every finding cite the page it came from?
Final Thoughts
The build-or-buy question is narrower than it first looks. Most case preparation work can start on built-in agents, and a custom agent becomes worthwhile only where a firm repeats a step in a way no one else does. Whichever route a firm takes, Opinion 512 leaves responsibility where it already sat: with the supervising attorney who signs off on what the agent produced.
A reasonable first step is small. Pick one closed case, run it through the agents you’re evaluating, and compare the output with what your team actually filed.
FAQs
Can a plaintiff firm build its own AI agents?
Yes, on platforms that support custom agents. The firm needs someone who can describe the process step by step and check the results, or it can have the vendor build the agent; Anytime AI supports both.
Do AI agents need attorney supervision under ABA rules?
ABA Formal Opinion 512 applies Model Rules 5.1 and 5.3 to generative AI use, so managing lawyers set policy, supervising lawyers make reasonable efforts to ensure compliance, and the lawyer stays responsible for the work. State rules control where they differ.
Can a firm bill clients for AI agent work?
Opinion 512 says a lawyer can’t bill clients for time spent learning a tool used regularly, and a tool that works like office equipment is usually overhead. A per-use charge from a third-party provider can ordinarily be passed through as an expense with disclosure.
What is a custom AI agent for a law firm?
It’s an agent configured to follow one firm’s own steps, checklists or decision points, such as an intake screen or a pre-demand check. Built-in agents, by contrast, ship ready for work most plaintiff firms share, like chronologies and demand drafts.
Does Anytime AI train on firm data?
No. Anytime AI maintains zero data training, so client files are never used to train models, and data is encrypted with AES-256 at rest and TLS 1.2 or higher in transit.
Get Started
Ready to go deeper — and safer?
See how Anytime AI gives plaintiff firms the strategic edge
and the security their clients deserve.