Superintelligent Agents for Law Firms: What to Build Now

Superintelligent agents for law firms are, for now, a direction rather than a product: agents far more capable than today's at reading, reasoning and acting across the full record of a matter. No one can say when such agents will be reliable enough to trust, but a firm can prepare today. The preparation is a set of foundations: scoped memory, confidentiality boundaries and approval rules that let a stronger agent take on more work without taking on more risk.

Key takeaways

• Treat "superintelligent" as a planning assumption, not a purchase. Build for agents that are more capable than today's, and make sure each new capability arrives inside rules you already control.
• The biggest gains would come in matter intake, research support and deadline tracking, where work is repetitive, document-heavy and easy to check afterward.
• Memory should be scoped to the matter, with sources attached to every fact, so a capable agent never blends one client's information with another's.
• Confidentiality and approvals are the real limit on agent autonomy. Define what an agent may read, draft, send and change before you increase what it can do.
• Start with one narrow role and an observable success criterion. Widen it only when the evidence supports it.

What does "superintelligent" actually mean for a law firm?

The term is used loosely. For a managing partner or operations lead, the useful meaning is practical: an agent that can hold an entire matter in view, notice what a tired human would miss, reason through several steps without losing the thread, and use the firm's tools correctly. It does not mean an agent that replaces legal judgment, signs filings or owns the client relationship.

That distinction matters because it tells you where to invest. If capability keeps improving, the scarce resource will not be raw intelligence. It will be trust: knowing what the agent knows, what it is allowed to touch, and who answered for each decision. Firms that have solved those three questions for a modest agent will be able to extend them to a stronger one. Firms that have not will find that a smarter agent is simply a faster way to make an unreviewed mistake.

It is also worth being honest about uncertainty. Today's agents already make errors that sound confident, can be misled by hostile text inside documents, and need supervision. Plan for the capability you can verify now, and design the structure so that capability can grow without the structure being rebuilt.

How would matter intake change?

Intake is where a firm first meets a prospective client, and it is full of small judgment calls: Is this within our practice areas? Is there a possible conflict? Which attorney should see it? What is time-sensitive? Which facts are missing?

A far more capable agent could do the following well before a lawyer opens the file:

• Read the inquiry and any attached documents, and separate what the person stated from what the documents show.
• Check the parties named against the firm's conflict records and flag possible overlaps for a human to resolve.
• Identify dates that could create urgency, such as a notice period or a hearing mentioned in passing.
• Draft a short intake summary with open questions, each one tied to the sentence or page that raised it.
• Propose a routing decision and the reason for it, without sending anything to the prospective client on its own.

The important part is the last line. The agent prepares and proposes. A person decides whether to accept the matter, because accepting a matter creates duties. Intake also carries a subtle risk: a prospective client may disclose sensitive information before any engagement exists. The agent should treat that information as confidential from the first message, and the firm should decide in advance where it is stored and who can see it.

For a practical starting point on this role, the setup guide for law firm intake and deadlines covers the basic configuration. This article looks at what comes after.

What would research support look like with a much stronger agent?

Research support is the area where people most expect dramatic change, and also where the risk of false confidence is highest. A stronger agent could read a large body of authority, compare how different sources treat the same question, trace how a position developed, and draft a research memo in the firm's house style.

The working model that holds up is that the agent assembles and the lawyer concludes. Every proposition in a research draft should carry a pointer to a source the reviewing lawyer can open. A statement without a source is treated as a lead to check, not as a finding. The agent should also say what it did not find, and which questions it could not resolve, because silent gaps are what cause trouble later.

Two design choices make this safer today and scale better tomorrow:

• Retrieval from approved sources. Rather than relying on what a model remembers, the agent retrieves from the firm's own precedent bank and from research databases the firm licenses, then cites what it retrieved. This is the idea behind [retrieval-augmented generation](https://aws.amazon.com/what-is/retrieval-augmented-generation/), and it keeps answers tied to documents you can inspect.
• Separation of firm knowledge from matter facts. Your standard clauses, templates and prior work product are firm-wide. The facts of a client's dispute are not. The agent should know which is which every time it retrieves.

The second point connects directly to confidentiality, which is the next layer to get right.

How would deadline tracking change?

Deadlines are the clearest case where an agent can add value and the clearest case where it must never act alone. A missed date is among the most serious operational failures a firm can have, so the standard for this role is not cleverness. It is reliability and traceability.

A more capable agent could read a new court order, a contract or a notice, extract the dates and triggers it contains, apply the relevant rule to propose a computed deadline, and place it on the calendar as a pending item. It could watch for changes in a matter that move an existing date, remind the responsible attorney in the right channel, and prepare an end-of-week summary of what is coming and what has no owner.

The controls matter more than the cleverness:

• Every proposed date shows the document, the rule and the calculation that produced it.
• A named person confirms each date before it becomes official. The agent cannot remove or move a confirmed deadline without a human approval.
• Dates that are close, or that rest on an ambiguous trigger, are escalated rather than quietly computed.
• The agent never treats its own calculation as the final word. Rules differ by court and jurisdiction, and the calendar owner is accountable for them.

Illustrative scenario: a firm's agent receives a scheduling order by email. It extracts three dates, proposes them as pending entries with the paragraph that supports each one, and notices that one date depends on a service event that is not yet recorded. Instead of guessing, it asks the paralegal to confirm the service date. The paralegal confirms, the supervising attorney approves the entries, and only then do they appear on the shared calendar. Nothing here required the agent to be right alone. It required the agent to make the human check fast and well informed.

Which foundations can a firm build today?

Three foundations decide how much a firm can safely delegate, now and later: memory, confidentiality and approvals. None of them depends on a model being more capable. All of them become more valuable as it does.

Memory. An agent that serves a firm needs durable context, but durable does not mean unlimited. Keep matter memory scoped to the matter, with each fact linked to its source document and the date it was recorded. Keep firm memory (templates, preferences, practice notes) separate. Decide what is retained when a matter closes and what is removed. Good memory design is mostly about context discipline: giving the agent the right information at the right time, not everything at once. The ideas in [effective context engineering for agents](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents) are a useful reference for this.

Confidentiality. Permissions should follow the matter, not the agent. If an attorney is walled off from a client, the agent acting for that attorney is walled off too. Retrieval should be filtered by matter before the agent sees anything, not after. Inbound documents also deserve suspicion: a PDF or an email can contain text written to steer an agent into disclosing or doing something it should not. The [OWASP prompt injection guidance](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) describes the defenses, and the [OWASP authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) covers least privilege and deny-by-default access, which fit legal work well. Professional conduct rules on client information vary by jurisdiction, so your own counsel or ethics resource should confirm what applies to you.

Approvals. Sort every action the agent can take into a few tiers, and write the tiers down:

• Read: the agent may look at specific sources for a specific matter.
• Draft: the agent may prepare a document or message that stays internal until a person releases it.
• Send: the agent may communicate outward only with named approval, or from a short list of pre-approved message types.
• Change: the agent may alter a record, calendar entry or file only after a human confirms, and every change is logged with who approved it.

A firm that has these tiers in place can raise an agent from drafting to sending in one area, with evidence, without redesigning anything. A firm without them has to guess.

What should stay with lawyers, however capable the agent becomes?

Some responsibilities are not tasks to be optimized. They are duties. Giving legal advice, deciding to accept or decline a matter, certifying a filing, communicating a significant decision to a client and judging whether a risk is acceptable all belong to people who answer for them professionally. A more capable agent can prepare the ground for each of these. It should not occupy them.

Framing this early helps the whole firm. Lawyers are less likely to resist an agent that has clear limits, and clients are more likely to accept one when the firm can explain who is responsible. The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) offers a vocabulary for mapping, measuring and governing these risks that translates well into a firm's internal policy. Firms that already run compliance programs can reuse much of their existing structure, as described in the piece on agents for compliance teams.

It also helps to look at neighboring professions that face the same pressure. A firm that must keep one client's work apart from another's will recognize the problem described in AI for engineering firms and client walls, and the discipline of record keeping is familiar from personalized AI infrastructure for accounting firms.

A step-by-step plan for the first role

A practical way to prepare is to build one narrow role well. The following sequence works for most small and mid-sized firms.

• Choose one role with a clear boundary, such as intake triage or deadline extraction for a single practice area.
• Write the success criterion in observable terms: for example, every proposed deadline has a cited source and a named confirmer, and no unreviewed entry reaches the calendar.
• List the memory the role needs, and for each item decide whether it is firm-wide or matter-specific.
• Define the read, draft, send and change tiers for the role, and the person who approves each one.
• Connect only the tools the role needs, with the narrowest access that works, and keep a log of what the agent saw and did.
• Run the agent in a shadow mode first: it prepares outputs, people do the work as usual, and you compare the two.
• Review errors and near misses on a regular schedule, then decide whether to widen, narrow or pause the role.
• Only after the review, extend the role to adjacent work, bringing the same tiers and logging with it.

How Be Human approaches this: Be Human designs, implements and operates personalized AI agent infrastructure with the firm's own people in control. A first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. Managed operations can then include monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements. The service is remote and available in English and Portuguese.

Frequently asked questions

Are superintelligent agents available to law firms today? No one can reliably label any current system that way, and a firm should not buy on that promise. What is available is capable assistance for drafting, extraction and organization, which works best inside clear limits. Build the foundations now so that stronger capability can be adopted deliberately.

Should a firm wait for better agents before starting? Waiting leaves the slowest work undone: deciding what the agent may know, who can see what and who approves which actions. Those decisions take time and involve partners, operations and your ethics resources. Starting with a small role also teaches you where your records and processes are weak, which is useful whatever the technology does next.

How do we keep client information separate between matters? Scope memory and retrieval to the matter, and apply the same access rules to the agent that apply to the people it works for. Filter what the agent can retrieve before it sees it, and log access so you can review it. Your own counsel should confirm how professional conduct rules apply in your jurisdiction.

What is the first thing to measure? Measure the quality of reviewed output, not the volume produced. For a deadline role, that could be the share of proposed entries a reviewer accepts without correction and the number of items escalated for good reason. A role that is accurate, traceable and boring is ready to be extended.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.