Mark Alcazar opened our recent Law Firm Profitability group session, “AI Agents for Law Firms,” with this question (See session recording here.) The evidence shows that AI is able to handle large parts of what law firms do. A significant portion of the cognitive scaffolding for legal work, including routine drafting and document formatting, is already handled at near-zero cost. Your judgment and your client relationships are not going anywhere, but the production layer around them is another matter.
Most of the law firms I speak with are using AI for narrow, low-stakes tasks, if at all. The session I hosted with Mark Alcazar and John Fitzpatrick from Apex Velocity Catalysts was designed to change that. Everything they showed was live. Tools they are building and using in their own work today. What those demonstrations showed is worth your attention.
Building Without a Developer
The first thing they demonstrated was how fast something useful can be built without a software developer. Mark built a working new-matter intake form in roughly five minutes using Claude Cowork. He wrote a detailed prompt describing what the form should do, and Claude returned working HTML: a form that populates a table and exports to a spreadsheet. Not production-ready, but a great starting point. The distance between “we need a tool for this” and “here is a working prototype” is now measured in minutes, not months. Cost and technical complexity used to keep custom tools out of reach, but both have dropped considerably.
Improving the Output
The second thing they showed was how to improve the AI’s output. Most of you are already using Claude or ChatGPT for some part of your work, and the output quality varies based almost entirely on how the prompt is written. Two techniques made an immediate difference. The first is context: a bare prompt like “draft an engagement letter” returns something generic. Adding jurisdiction, matter type, client background, and the purpose of the letter produces something that resembles actual work your firm would do. The second is more interesting. Instruct the model to create a scoring rubric, score its own output against it, and iterate until it meets a defined threshold. In the demonstration, Claude scored its first draft at 86 out of 100 and kept improving. The mechanism works, and the results are noticeably better.
Agents vs. General AI
The third area was the distinction between general AI and purpose-built agents. General AI is versatile but unfocused. An agent is built for a single defined workflow, such as your NDA process or conflict check. Because the scope is narrower and the instructions are specific, agents are measurably more reliable for the workflows they support. John demonstrated a full NDA workflow: intake, drafting, multi-agent review, negotiation with version tracking, client approval, and e-signature integration. Early indicators from firms adopting this kind of focused approach suggest the gains come precisely from that focus. One workflow done reliably outperforms a broad tool used inconsistently.
Before You Deploy
Before any of this goes into practice, certain things need to be settled deliberately.
What can the AI do without human approval? That is a decision your firm has to make deliberately, not by default. Mark’s own agent is instructed never to send an email on his behalf, even though it is technically capable of doing so. He set that limit, the system did not.
How does AI-generated work get verified before it reaches a client? Whether that is a second agent reviewing against a rubric, or a structured checklist process, the principle is the same: one step produces, another step evaluates.
Where is your client data going? Consumer versions of Claude and ChatGPT carry different protections than enterprise accounts. If your firm is using consumer tools for client-related work, that needs to be resolved before any other adoption decision.
Where to Begin
On the question of where to begin: every firm has what Mark called “drag.” It is the repetitive, error-prone work that consumes time without generating strategic value. It is spread across every role and usually invisible because everyone is too busy doing it to examine it. The exercise is simple. Identify the drag and put a number on it: hourly rate multiplied by time spent each week. Prioritize based on what automation would recover.
Then pick one workflow and make it work before touching anything else. Forward-thinking firms are not trying to transform everything at once. They are proving value in one place and expanding from there.
Colin Cameron is President of Profits for Partners and founder of the Law Firm Profitability group on LinkedIn. The session was presented by Mark Alcazar and John Fitzpatrick of Apex Velocity Catalysts.
