Introduction
James Chou: We’re excited to talk about AI agents. It’s a hot topic in the legal tech industry, and hopefully from this webinar you’ll come away understanding a little bit more about AI agents, at least the way Pramata views them. We have a special guest, Alex Bubier with Smarsh, who will walk us through his journey of building out his own agents and seeing them come to life.
The GCs who are winning are actually starting smart. It can seem like a daunting task to build your first agent, but by partnering with Pramata and going through trials of building different agents, you’ll get rapidly acclimated and build the confidence to really start scaling and benefiting the whole organization.
A little background on Pramata for those hearing about us for the first time: we’re an enterprise-grade contract AI that actually works. We provide full end-to-end CLM capabilities, we have over 20 years of experience in the contract AI space, and we offer flexible deployment models. That includes full end-to-end or a best-of-breed approach where we work with your existing vendors to get you the contract intelligence your organization needs.
Alex Bubier: Smarsh is a software and technology provider predominantly operating in the financial services industry.
What Is an AI Agent?
James Chou: The definition of an AI agent varies depending on who you ask. Google describes AI agents as software systems that use AI to pursue goals and complete tasks on behalf of users, a more autonomous framing. AWS defines them as software programs that can interact with the environment, collect data, and use that data to perform self-directed tasks that meet predetermined goals.
The definitions exist on a spectrum and there are ongoing debates about how autonomous an agent must be to qualify. But rather than arguing definitions, what matters is the common thread: agents need context to act. For legal teams specifically, that context lives within your contracts.
At Pramata, our AI agents are purpose-built for specific contract tasks. They’re grounded in your actual contract repository, not a generic agent using general knowledge. If you’re asking questions about your contracts, the agent had better be reading your actual contract data. Our agents are also user-initiated, giving users the flexibility and control to place agents where they need them and trigger them when needed. And the outputs are repeatable and trustworthy, consistent results every time, whether you’re surfacing contract information, generating dashboards, or doing analysis in Microsoft Word.
The Three AI Agent Archetypes
James Chou: We’ve defined three AI agent archetypes. These are not sequential. You can start with any one of them or a combination. The most important factor is identifying what business pain point you’re trying to solve. Some archetypes overlap in functionality, and in practice, solving a real business problem often involves elements of all three.
The three archetypes are the Automator, which frees your team from repetitive high-volume work; the Analyzer, which unlocks strategic insights from your contract portfolio; and the Accelerator, which streamlines and speeds up contract processes end to end.
Archetype 1: The Automator — Eliminating Repetitive Work
James Chou: Start with the Automator if your team is drowning in routine, high-volume contract requests like NDAs, amendments, and renewals. Legal teams often get bombarded with questions like: Do we have an NDA with this counterparty? What are the payment terms? When is this account up for auto-renewal?
The goal is to give requesters immediate, accurate answers without routing every question through legal. This frees your legal team to focus on complex deals and strategic initiatives rather than answering hundreds of routine emails per day. Ideally, your legal team is working on negotiating complex deals, working on strategic initiatives, and partnering cross-functionally, not answering common contract questions all day.
Some outcome-based examples: agents can instantly check whether an NDA or MSA is on file for a given account without contacting legal. They can triage contract requests and route them to the appropriate approver automatically. They can answer routine contract questions on demand.
James Chou: Alex, how did you get started with AI agents? What led you to where you are today, and what role did Pramata play?
Alex Bubier: I started slow, which is probably the best answer. Because of the company I work at, privacy and data protection are very important, so that led me to experiment with local options first. I had colleagues experimenting with Ollama, and I figured with a JD it couldn’t be that hard. The outcome was I absolutely failed. I couldn’t even get the program started. There were a lot of different dependencies and it was frustrating.
The next step was I really started to read. I got a Medium membership and started educating myself on different tools and approaches. As service providers enabled more functionality, the technical hurdle was removed. Partners like Pramata made it much simpler to start testing and using these tools.
With Pramata, I was able to build an agent very quickly and easily. The reason I was successful was I had the benefit of hindsight, and I really didn’t try to do too much. I started very small. I took a simple, repeatable problem the legal team was having. I think I started with: Does this contract have termination for convenience? I built an agent to extract that information, and it worked. Because I limited the context and limited the ask, I was able to build an agent that successfully pulled that information.
The outcome was I could extend that agent to my internal legal team, and they too were able to quickly and easily use it because it was a single click. They didn’t have to understand the background or the tech behind it. They just knew how to use that single-purpose agent, and they recognized that the information being returned was relatively valid.
James Chou: For the most part it is difficult to get up and running, whether it’s with Pramata or a different vendor. It seems like such a daunting task because you’re thinking of all these business use cases. A lot of folks run into paralysis by analysis. They’re thinking, if this doesn’t do everything my team wants, they get stuck and don’t know where to start.
Where we’ve seen a lot of successful implementations is that customers start small. It might not tackle every single use case they’re going after, but successfully building that first agent and showing the benefits to team members, where it’s easy enough for them to use, is where you really start snowballing in terms of gaining buy-in from cross-functional partners and scaling.
Demo — Automator in Pramata: Within Pramata at the account level, you can chat directly with your contracts. For a given account, you can ask questions across the entire family of contracts: the master agreement, order forms, and amendments. In seconds, without manually reading through documents, you can surface a high-level summary of payment terms across all related contracts.
Agents like this can be built for any information your team frequently needs. The key benefit: your cross-functional teams in sales, finance, and deal desk can get accurate answers to contract questions without going through legal, with the AI grounded in your actual contract data.
Another example: within a contract request workflow, a sales rep requesting a new NDA can click an agent that searches the repository to check whether an NDA with that counterparty already exists. If one is already in place, the user knows immediately. One less Slack message or email to legal. Some customers have made the entire NDA process fully self-serviceable this way.
Alex Bubier: The biggest paradigm shift is moving from a chat box to an agent. For buy-in, from my team and from others, the outcome had to be relatively consistent. With a plain chat box, different individuals use different terminology or different prompting approaches to get to the same answer. By using an agent, the team crafts a unified prompt each time. This increases ease of use and, more importantly for me, increases the accuracy and consistency of the output, especially when extending it to other legal team members or business stakeholders.
James Chou: You can fully customize agents in Pramata so that a required output format is always produced when someone presses that button, whether that’s specific charts, specific sections, or a specific structure. No matter who interacts with the agent, they get the same output format.
Archetype 2: The Analyzer — Unlocking Strategic Insights
James Chou: Start with the Analyzer if you need to demonstrate legal’s strategic value to your executives. Legal should not just be the group that reviews contracts. You want to be a strategic advisor providing informed inputs based on your contract data, and a true business partner rather than a cost center.
The Analyzer builds on the Automator. While the Automator surfaces contract information consistently and simply, the Analyzer performs a level of analysis you can present to your C-suite.
Outcome-based examples: these agents can extract critical terms like price escalators, liability caps, and specific clauses across your entire contract portfolio. They can accelerate M&A due diligence by surfacing risks such as change of control, assignment rights, and non-compete provisions and categorizing them by risk level. They can generate customizable dashboards exportable to PowerPoint for executive presentations. And they can identify auto-renewal deadlines across your portfolio for proactive planning.
James Chou: Alex, how did you scale from one agent to multiple agents? You started with a couple of simple agents but now you’re managing multiple agents that vary in complexity. What did that process look like?
Alex Bubier: Initially we were addressing internal legal group problems, questions coming to us that we were trying to solve. Once some of our business teams heard about what our legal group was doing with agents, they asked if their problems could also be addressed in this manner.
Legal teams tend to play strategically between a lot of different groups, so we had insights into problems other business teams were facing. We started scaling when we began partnering with other business teams to address their problems. For example, our marketing team wanted to understand whether contracts enabled publicity and use of marks. Instead of answering a single question, we were now looking at a group of questions with some level of analysis on top.
We also started getting requests from executives for a risk dashboard across all contracts. Once we took those problems and cut them into smaller component pieces, it allowed us to create agents to extract information, analyze it, and produce content we could review and refine.
James Chou: That’s an important point. When you’re trying to solve a business problem, you want to break it into smaller subcomponents. If you’re trying to make the NDA process self-serviceable, you might start with one automated agent that checks whether an NDA is in the repository, then add an analyzer-type agent that does a first-pass review of a third-party NDA. Through a combination of agents, you can really start building these workflows out.
One of the biggest things is showing even simple agents to your counterparts, letting them see how easy it is to use and letting them trust the results. That gives the confidence to build a hub-and-spoke model, one agent built on top of another. At the end of the day, when you’re trying to solve a business problem, it’s usually going to be a combination of all three archetypes.
Demo — Analyzer in Pramata: Going back to the M&A use case: you’re doing an analysis across all contracts within a given subset, extracting risks like change of control, assignment rights, and non-compete, and categorizing them into risk levels. The M&A dashboard agent generates specific charts and graphs every time, with the same structure but content that changes based on what’s in the report. Executives can see all risk data in a visual format and you can export it directly into PowerPoint for deal team presentations.
A similar pattern works for vendor product comparisons: across all your vendor contracts, you extract the terms you want to analyze and get a visual comparison, customized to what your organization cares about. Once you click that agent, it always produces the same output structure with updated content.
Alex Bubier: For those who answered in the poll that they’re still learning what an agent is: you can start small and smart, solve one problem, and then build on it. What you start to realize is that these agents and these situations all build on each other. I also had all the information I needed to build these agents because a lot of teams have playbooks, or attorneys have preferred approaches to how they address certain issues. You can effectively feed those into what we call an agent. It’s essentially a playbook or guideline with a reasoning component on it.
These agents will carry out tasks in a very similar manner to how you would if you provide them with enough information. While 50% of the people here might not have started yet, you might already have the tools available to get started. It’s really important to just get started, see how these agents work, and then refine.
James Chou: Something important: you need clean, curated contract data in your repository. Before you even think about agents, make sure your repository has the right foundational data, your actual executed contracts, without partially signed agreements or duplicates. If bad data feeds into your AI agents, the analysis will reflect that. Start with clean data, then start small with agents, and build on your wins.
Archetype 3: The Accelerator — Streamlining Contract Processes
James Chou: The Accelerator is where you really start to see agents working together. This archetype is for streamlining contract processes from end to end, including redlining, negotiation pattern surfacing, contract request workflows, and more.
Think about getting redlines from a counterparty on your MSA template. You might have the redlined version on the left and the template on the right, going line by line. That approach introduces the possibility of errors, is very manual, and doesn’t scale for complex agreements. With an Accelerator agent, you can have it do the redline comparison against your template, surfacing changes systematically with suggestions based on your playbook.
Outcome-based examples: agent-driven redline comparison against your own contract template; first-pass review of a third-party NDA against a legal team checklist before the document reaches legal; playbook-driven contract analysis in Microsoft Word with suggested redlines and comment language for the other party; self-serviceable NDA workflows where the requester checks for existing agreements, runs a compliance check, and submits to legal with context already attached.
James Chou: Alex, at what point were you ready to build something more complex? How did you go from introductory agent building to taking the training wheels off?
Alex Bubier: I feel like to some degree I always have the training wheels on. But we started with data extraction agents, and one of the things we really spent time doing was validating whether the accuracy was there. Once we validated that the accuracy of the prompts was within our tolerance level, we moved on to working with our business to see if that type of agent was going to meet their needs. We quickly surpassed just data extraction in order to address our business stakeholder needs. Their request was for more comprehensive analysis.
Once we got buy-in from the data extraction work and realized their problems needed more comprehensive components, we started experimenting with what I’d call the analysis component agent, where you’re not only extracting the data but also performing an initial analysis. The other piece was recognizing that we had all the pieces: internal playbooks, an Excel sheet with different clause banks, and critically, the rationales behind each clause. Once you start feeding that into agents, they can start really adopting and using that contextual layer to provide analysis.
Once that got really fine-tuned, those complex agents provided serious value. I could run a complex agent, have it output issues and sections I needed to look at with citations, and I felt relatively certain the information was strong. It takes me from a place where I might have to read a whole contract to a place where I only need to review specific sections. It takes me from a two-hour task to maybe a 45-minute task. When you can start having the conversation about next steps 45 minutes earlier, whether that’s understanding risk or how to proceed with a large renewal or strategic engagement, it provides immense business value.
James Chou: To the playbook point: even if you don’t have a playbook, Pramata can help you create one. We can take a template or representative contract and build out the playbook, then in Microsoft Word the AI takes context from the draft contract, uses the playbook to identify risky provisions, suggests redlines, and even drafts the comment language you would give back to the other party. This also helps ensure you don’t miss anything. If someone made changes with track changes turned off, the AI will still pick it up when doing playbook analysis or a diff compare.
Alex Bubier: Taking that one step further: legal teams have playbooks, position statements, and an understanding of what clauses are problematic from a business perspective. You can build those into checklists and agents, and allow your team members to run these checks upfront before the documents even get to legal. The individual is now triaging for you. Your review cycle might be one, two, or three weeks, and you’re now moving to the next phase merely by adding that checklist upfront through the agent. It’s a powerful tool, with the caveat that you have to build it correctly. But you can get started, start small, build, and find new and interesting ways of using the information sets you’ve already built to enable and scale your business teams.
Choosing Your Starting Point: A Decision Framework
James Chou: Here is a framework for deciding where to start with your first agent. The bottom line is to start based on your business needs and your biggest pain points right now.
To Alex’s point and his journey: he started small, but very quickly he was able to build on the small wins, show his business partners, get buy-in, and now he’s at the point where he can build more complex agents that provide even more ROI, more business insights, and more strategic-level analysis that benefit his organization even more.
Throughout this whole journey, our motto is: start smart and scale fast. And you don’t have to do it alone. This isn’t something where we expect our customers to do it by themselves. We understand this is a learning journey. At Pramata, we partner with all of our clients throughout their entire journey, whether it’s their first simple agent or the more complex ones. We’re here to help along the whole way.
Q&A: Access Controls and Permissions
Alex Bubier: We have access controls for who can see what documents. Privacy and data protection are very important, contracts need to be locked down, and confidential agreements need specific viewing permissions. That’s why Pramata has specific user roles and profiles, with a level of permissioning in place so that only the people who need to see certain agreements can view them.
James Chou: And there are different modules, so sales individuals could be submitting requests from different places where they don’t actually have access to the underlying documents.
