From Alpha to Action: How AISDK5 Helped Us Build an AI-Native CRM

[Music] Hi, thanks so much for having us. I'm Jack and along with my colleague Nikita who will be out here uh soon. We built Lightfield, an AI native CRM. We started using AI SDK V4 back in January and adopted V5 as soon as it hit alpha in June. Today we want to share with you how we built a production system where AI agents have secure full readr access

to customer data, how we handle human in the loop workflows, and the architectural decisions that made it all work. We'll walk through the patterns we discovered, the trade-offs we made, and how the AI SDK enabled us to move fast without painting ourselves into a corner. But first, let's talk about why CRM were broken and why this matters. So, who's familiar with CRM? Maybe some engineers. Yeah. So, so

here's what's supposed to happen, right? You start talking to customers. Maybe you're a founder doing sales. Maybe you're on a sales team. At first, it seems manageable. You remember everyone. Every conversation is fresh in your mind. Then you hit 10 customers, 20, 50, and someone on your sales team asks, "Hey, what did Sarah from Acme say about our pricing? Did she have concerns about the enterprise tier?" So now you're searching Slack, you're

searching your email, you're searching Google Docs, maybe that Zoom recording that hasn't been transcribed yet. You eventually find it buried in a thread from 2 weeks ago, but you realize you never updated your spreadsheet again. So, you buy a CRM, it promises to be your single source of truth, but it just becomes another place you forget to update. Here's the problem. Traditional CRM were built decades ago with the fundamental assumption that humans will do manual data entry. They gave you these rigid

fields and predefined schemas, but the actual context nuance of your conversation, it lives in your email, Slack, meeting notes, different places. And it just becomes CRM just becomes a reporting tool for your VP of sales, not something that helps you sell. So, we thought there had to be a better way. What if the system could just remember? What if it captured everything intelligently and could actually take action on your behalf? That's

Lightfield. So, Lightfield reimagines what a CRM should be. It's a system of memory and action for startups. So, it has automatic capture. Conversations, uh, meetings, emails, they all get captured and structured without manual entry. It has lossless memory. We support schemalist and customizable schemas. You don't need to know what to track up front or pay a consultant to set it up for you. and it turns memory into action.

Lightfield uses all that captured context, both the structured and the conversational data to draft follow-ups, surface insights, and automate workflows for you. Now, traditionally, CRM are built for sales team tracking sales deals. But because Lightfield captures and structures all this conversational data, it becomes really powerful for anyone who needs to remember and act on customer context. What were the most requested features from last week's onboardings? Customer

success teams understanding patterns across support conversations. Same system, different questions, but all powered by that same memory layer. That's the product. Let me show you what it actually looks like. So, here's an example of a asking lightfield agent. Um, I think we're asking find five stalled ops and draft a personalized email to each. So it can search across all of your customer information using an agent built on a AI