AI agents have been technically capable of useful legal work for a while. The problem was not only accuracy or governance. The operating cost often made the maths hard to defend.

A first-pass document review agent that had to read a large discovery batch, reason across dozens of documents, draft a structured memo, and send the result to a supervising lawyer could consume a meaningful volume of model tokens. For a larger matter, the monthly spend could feel uncomfortably close to the cost of assigning more junior associate hours.

That equation is changing.

Writer’s recent Palmyra X6 announcement claimed a 52% reduction in AI agent costs for its new model. One vendor’s claim does not set the price for every model or every agent workflow. It does point to a real market direction. Model costs are falling while the quality of structured agent work is improving.

For law firms doing $1 million to $25 million in annual revenue, this matters because the work that was marginal six months ago can now be economically sensible. That includes intake screening, first-pass contract review, discovery sorting, matter updates, and routine follow-up.

The firms that benefit won’t be the ones that throw a chatbot onto their website. They will be the ones that identify a defined workflow, set clear boundaries, connect the agent to the right systems, and measure the economics against the current cost of delay and attorney time.

Lower token costs change the automation threshold

Most partners do not care what a token costs in isolation. Nor should they. They care about the total cost to acquire a good client, progress a matter, deliver quality work, and keep the team focused on billable legal judgment.

AI agent economics are best understood as a comparison between four numbers:

1. The cost of the work today, including attorney and staff time.
2. The revenue lost when the work is delayed or missed.
3. The cost to run and supervise an agent workflow.
4. The cost of mistakes, including the controls needed to prevent them.

The fourth number is critical in law. An agent should not decide whether your firm takes a matter, give legal advice, clear a conflict without a defined process, or send substantive legal conclusions directly to a client. Those are lawyer-led decisions.

But plenty of work around those decisions is repetitive, time-sensitive, and structured enough for an agent to complete safely with review points.

Take an after-hours intake call. The firm does not need an AI system to advise the caller. It needs the call answered, the caller’s contact details captured, the matter type classified, the urgency identified, a basic conflict information set collected, and an appropriate consultation offered. If the matter involves a time-sensitive court date, an injury, a criminal charge, or another trigger your firm defines, it needs escalation.

That is not a vague automation project. It is an operating process.

Historically, firms could justify a simple call answering service but hesitate at a more capable voice agent that listens, asks follow-up questions, records structured notes, checks availability, and completes a booking. As model use becomes less expensive, the capability gap between basic routing and a well-managed agent becomes more affordable.

The same applies to documents. A junior associate billing $200 to $400 per hour may spend a day sorting a discovery production, identifying standard clauses in a contract set, or producing an initial chronology. Some of that work is essential legal training. Some of it is the kind of repeatable first pass that consumes expensive capacity without requiring senior legal judgment at every step.

The manual work that creates the biggest leaks

Law firms often look first at headline overhead. Rent, salaries, legal research platforms, and marketing are easy to see. The more damaging leaks are usually scattered across the week.

We usually see three areas create an annual leakage band of roughly $80,000 to $250,000 for firms in this revenue range. The exact figure depends on practice area, matter values, staffing structure, and how consistently the firm records non-billable work.

Intake that arrives when nobody is available

A prospective client calls at 6:40 pm. The lawyer is in a client meeting or at home. The receptionist has left. The voicemail asks the caller to leave a message.

They may do that. They may also call the next firm on the search results page.

For practices dependent on consumer or urgent commercial inquiries, 30% to 40% of after-hours intake can fail to convert when the response is slow. That is not always because the firm was a poor fit. Often, the next firm simply replied first.

The manual version of intake also creates a slow internal chain. A receptionist transcribes a message. An office manager forwards an email. A partner asks two follow-up questions. A consultation is offered the following day. Notes are incomplete, the referral source is missing, and nobody knows which leads were never contacted.

An Omni Voice deployment can support the Intake Voice Agent through this first interaction. The agent answers after-hours, during lunch, and on weekends. It identifies the caller, asks approved qualification questions, captures the relevant parties for a preliminary conflict workflow, records the practice area and urgency, and books a consultation into the right calendar.

It does not give legal advice. It makes sure a prospective client gets a prompt, professional response and that your team begins with useful information.

Attorney time that never reaches the invoice

Most firms have an uncomfortable amount of work that is real but never billed. We commonly see an average of four to six hours per attorney per week spent on document sorting, client status messages, intake follow-up, internal coordination, and matter administration that never makes it to an invoice.

Some non-billable time is necessary. You should not attempt to turn every client touchpoint into a timer running in six-minute units. The question is how much of the administrative load is being carried by people whose time should be reserved for legal work, client counsel, supervision, and business development.

An agent can draft a client status update from approved matter events, prepare a list of missing documents, create a structured handoff from an intake form, or turn a scanned inbox into tasks for the correct team. A human reviews exceptions and anything client-facing before it leaves the firm.

The aim is not to eliminate judgment. It is to remove the repetitive preparation around judgment.

First-pass review that ties up associates

Document review is often where partners see the strongest potential return and the highest perceived risk.

Both instincts are reasonable.

A Document Review Agent should not be treated as a lawyer. It is a controlled first-pass assistant. It can read a defined batch of documents, identify specified clauses or categories, produce a chronology, flag missing information, compare documents to an approved checklist, and draft an associate-grade memo with citations back to source material.

For contract work, that may mean flagging change-of-control clauses, indemnities, liability caps, notice requirements, assignment terms, or non-standard governing law provisions.

For discovery, it may mean grouping documents by custodian, identifying likely responsive communications, extracting dates and entities, surfacing repeated issues, and preparing a review queue. The supervising lawyer still makes privilege, relevance, legal interpretation, and production decisions.

The cost decline matters here because a well-designed agent may need to process substantial material. If model cost falls, the firm can run more useful first passes without making a small matter uneconomic.

What an agent workflow looks like in practice

The useful unit is not “an AI tool.” It is a workflow with a beginning, a set of rules, a handoff, and a measurable outcome.

Consider a personal injury or family law intake workflow.

A prospective client calls outside office hours. The Intake Voice Agent answers with an approved greeting and explains that it can collect preliminary details and arrange a consultation. It captures the caller’s name, contact details, adverse party or parties, case type, location, relevant dates, and stated urgency.

The agent checks the firm’s consultation availability based on rules you set. It can offer a booking only when the matter fits the practice area and the conflict information is sufficient for a preliminary review. If the caller reports an emergency or a deadline inside a defined window, the agent follows an escalation script.

After the call, the Matter Triage Agent receives the call record, web forms, and relevant emails. It classifies the matter, scores fit against the firm’s intake criteria, checks for missing information, and routes the matter to the right partner or intake coordinator. That person receives a one-paragraph brief rather than a raw transcript.

The brief might include:

• Matter category and apparent jurisdiction
• Key parties for the conflict process
• Trigger dates and stated urgency
• Referral source
• Fit score and reason for the score
• Recommended next action
• Questions that still need a lawyer or intake staff member to answer

This is where Omni Ops becomes practical. It connects the agent process to the tools your firm already uses, such as email, forms, calendars, case management platforms, CRM records, and task queues. The work is routed into the operating system of the firm rather than sitting in another AI dashboard.

For document review, the process is similar. A lawyer or paralegal selects a matter folder and an approved review template. The Document Review Agent processes the files, records its findings against the template, cites document locations, highlights confidence limits, and prepares a review memo. A lawyer reviews the output, corrects anything needed, and decides what becomes work product or client advice.

That last review stage is not a failure of automation. It is the control that makes legal AI usable.

The right way to assess the financial case

Do not start with a promise that an agent will replace a headcount. It is usually the wrong frame, particularly in a growing firm.

Start with capacity, conversion, and leakage.

If an Intake Voice Agent recovers even a small number of high-fit consultations that would otherwise go unanswered, the return can be clear. The precise value depends on your average fee, close rate, and case duration. A firm with a $5,000 average initial engagement needs a much different conversion result than a firm whose average matter produces $75,000 in fees.

For document work, calculate the current annual hours spent on first-pass tasks. Use actual time records where possible. If time is not recorded because the work is routinely written off, ask associates and paralegals to estimate it for two representative weeks. Then separate legal analysis from sorting, extraction, initial summarisation, status preparation, and routine follow-up.

That gives you a realistic baseline.

A good business case also includes agent supervision time, implementation effort, quality checks, and usage cost. Lower model prices reduce one line of that calculation. They do not remove the need for process design.

This is why we do not recommend buying an agent because a vendor says it is cheaper than the prior version. The model is one component. The actual return comes from using it on work that has enough volume, enough structure, and enough cost or revenue impact.

If you want a practical first step before changing any system, download the AI Client Intake Checklist for Law Firms. It gives your team a worksheet for defining intake questions, escalation triggers, conflict data, booking rules, and handoff responsibilities. You can also access the direct intake checklist download for use in your next partner or operations meeting.

Where law firms get agent deployment wrong

The most common mistake is trying to automate a messy process before the firm has agreed on how the process should work.

If three partners use different intake criteria, no agent can resolve that inconsistency. If nobody owns the response queue, a matter triage system will create cleaner notes that still go nowhere. If document folders are unmanaged and templates vary by lawyer, review outputs will be less reliable than they should be.

The second mistake is using an agent without clear guardrails. For legal work, those guardrails should cover data access, approved sources, escalation rules, human review, client communications, retention, permissions, and audit trails. The firm should decide exactly what the agent can do independently and where it must stop.

The third mistake is chasing the most complex use case first.

Start with the work that has a clear input, a repeatable decision path, and a measurable result. After-hours intake is often a strong candidate. So are incoming email triage, matter status preparation, and standard first-pass document tasks.

You can review the broader operating options on See Omni for law firms, including how voice and operations agents can be designed around your actual workflows. Our AI resources and guides can also help your team build a working vocabulary before you commit to a project.

A 60-minute audit beats a generic AI roadmap

You do not need a slide deck about the future of legal AI. You need to know where money and capacity are leaking in your firm, which workflow has the clearest return, and what it would take to deploy it safely.

That is the purpose of an Omni Audit.

In 60 minutes, we work through three outputs. First, a map of the high-friction workflow and where the current handoffs break. Second, a prioritised agent opportunity with expected operating requirements and commercial logic. Third, a practical next-step plan that identifies data, systems, owners, controls, and a pilot scope.

There is no generic transformation pitch. We look at the calls, forms, emails, documents, calendars, and people involved in your firm.

If lower agent costs have moved automation from “interesting” to “worth assessing,” Book a 60-min Omni Audit. Bring one workflow that frustrates your team. We will work through the financial and operational case together.

The opportunity is cheaper capacity, not cheaper legal judgment

The 52% cost claim behind a new model is not a reason to make a rushed decision. It is a signal that the cost barrier is moving.

For a law firm, that opens a practical window. Work that was too expensive to run through an agent at volume can now be reassessed. Intake screening can be active every hour your firm is not. Document review can begin with a structured first pass. Client updates can be prepared from matter activity instead of assembled manually by attorneys and staff.

The firms that act well will keep lawyers responsible for legal advice and high-stakes decisions. They will use agents to make the firm faster at responding, more consistent at routing work, and less dependent on expensive people doing repetitive preparation.

Start with the work that is already costing you money. You can see the AI audit for law firms to understand the process, or Book my Omni Audit when you are ready to put numbers against the opportunity.