In this demo Christopher Parsons walks through building a Synthesis Knowledge Agent that reviews nondisclosure agreements (NDAs) for an AEC firm. He explains the hierarchy of agent creation—any employee can create a personal or shared agent, while only global and agent administrators can create global agents that anyone can use. He then creates a personal agent called the "Contract & NDA Review Assistant," assigning it a stock avatar and a concise purpose that informs both users and the agent about its role. Parsons shows that grounding knowledge is optional for Level 3 agents, so he skips uploading a knowledge base and instead defines a single action: "review NDA and flag risks." He pastes step‑by‑step instructions that tell the agent to (1) read the uploaded NDA, (2) compare it against a team‑generated NDA Best Practices document (derived from interviews with a retiring CFO and codified as preferred, negotiable, and deal‑killer clauses), and (3) output a severity‑rated table with explanations, suggested next steps, and escalation guidance. He highlights configurable response formatting options (verbosity, audience, structure) and notes how these can be tuned for other use‑cases like HR. After saving the agent as personal, he uploads a draft NDA, runs the process, and the agent returns medium‑ and high‑risk findings (e.g., a 10‑year confidentiality term flagged as high risk) along with negotiation recommendations and a final recommendation to involve legal counsel for super‑high‑risk items. The demo illustrates how tacit legal expertise can be democratized via an AI‑assisted workflow.
Chapters
So, um, who can create agents? So anybody in the firm can create an agent. Um, but, uh, only global administrators and agent administrators can create what we call global agents. And that's an agent anybody can use. So any employee can create a personal agent and a shared agent. So a personal agent is just for yourself. A shared agent is shared with your team or your department. So I'm going to do a little bit of copying and pasting here to make this easier for everybody. So you don't have to watch. Me type. So I'm going to call this agent that I'm making, the contract, an NDA review assistant. I'm going to give it an avatar so I can upload images, but in this example, I want to give it a stock icon. I happen to have one saved. I'm going to use that. And then I'm going to give the agent a purpose. And so this purpose is used both for people who are looking at the agent, to understand how it can help them and also helps the agent know at a high level what it should be doing. This is the agent identity at a high level. I can then configure the grounding knowledge, and we've been looking at grounding knowledge for the last 20 minutes or so. In this example, this agent actually doesn't need any grounding knowledge, and this is oftentimes true with these Level 3 agents, is it doesn't need anything to search to answer the question. You're going to upload a resource in this example, like a non-disclosure agreement, And it's going to compare it against best practices. So I can just go ahead and I can skip this step. But what I do want to do is define an agent action. And you can see when I click on create action, my choices are to create something. So it's a creation action, to run a process or ask a question. I'm going to pick run a process. And so I'm going to call this action review NDA and flag risks. And then I'm just going to paste in the instructions, but then we'll talk about them. Okay. So step 1, you're going to review the uploaded NDA. Fair enough. Step 2, you're going to compare it to NDA best practices. I happen to have a page that our team put together called NDA Best Practices. And this actually example comes directly from a client example. And the way that they created the NDA best practices was to interview their retiring CFO to look at previous recent examples of NDAs that they had marked up, and then just to talk through best practices of what are our preferred terms, what are things we always try and negotiate, and what are just unacceptable provisions that are always going to be a no every time, deal killers. And then it kind of— out of that, they were able— this isn't the document they put together, but it's something like this. They were able to create this best practices that helped understand what the preferred terms are and then how to negotiate them or what, what we would, we would accept, you know, as a response and really give that kind of like deep tacit knowledge that this one person had who was reviewing these NDAs over and over and over again and democratize that knowledge to put it on the intranet. Yes. But then also give it to the agent so it can do that work. So the agent will review the uploaded NDA, compare it to the best practices. We're back to that obvious disclaimer at, at the top of the page saying it's AI generated. And then, you know, I'm asking it to create a table based on severity, explain each rating, and then a suggested next step. Great. Now I could go ahead and add some additional actions, but I'm not going to. That's fine for now. There's a lot of detail here in terms of configuring the response formatting, like how it's structured, who it's written for. How long it is or verbose it is. I'm going to accept the defaults for now, but that's some good tuning you can do on these agents. This comes in really handy when you've got something like an HR agent and you want the fidelity to be like verbatim. Like you always want to make sure we're referring to content verbatim. That's fine for my agent. I'm going to start this as a personal agent so I can do some testing, but later you can see that I can make it a shared agent and invite other people into it. Or if I have the permissions, I could make this a global agent so that everybody in the firm could see it. But personal agent's fine for now. Cool. So in a couple minutes, not even that, maybe I built an agent. And now what I can do is I can upload a nondisclosure agreement, a draft from the client. Well, we're considering it a draft. They may not consider it a draft. And once that's uploaded, I can review it and have the agent flag the risks. So you can see it starts by reading its instructions. It's going to get that best practices document loaded into context, and now it's going to get to work. It's going to tell me it's AI-assisted, and now it's going to start working step by step. It's found, you know, a medium risk in the NDA, told me what the concern is and the suggested action. Here's a high risk, 10+ years confidentiality term. That's pretty long. And so we prefer 2 to 5 year range. And so it comes back and says, we recommend negotiating this down to 3 to 5 years. And you can see as I'm scrolling, the agent's going and creating this, some medium risk, additional high risk pieces, and so on and so forth. And then at the end, it summarizes the super high risk ones. And this agreement is warranting escalating to legal counsel. It's so off the mark that this is probably a deal killer and you're probably gonna want a lawyer involved.
