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

Jack introduces Lightfield, an AI-native CRM built using AI SDK V5, and identifies the fundamental problem with traditional CRMs: they were built decades ago assuming humans would manually enter data, resulting in rigid schemas that fail to capture conversational context scattered across email, Slack, and meeting notes. Traditional CRMs become mere reporting tools rather than active sales enablers, forcing users to search multiple systems to find customer context. The insight is that customer information lives across many communication platforms but CRMs only store structured fields, creating a gap between actual business interactions and the system of record. This disconnect is why most CRM adoption fails—the system doesn't reflect how business actually happens, making it easy to forget updating it.

Lightfield reimagines CRM as a system of memory and action featuring automatic capture of conversations, meetings, and emails without manual entry, lossless memory with customizable schemas, and action-oriented workflows like drafting personalized follow-ups and surfacing insights. The platform captures both structured and conversational data, making it valuable beyond traditional sales teams—customer success teams can identify patterns across support conversations, and engineering teams can ask natural language questions about customer issues. A live demo shows the agent searching across customer information to find five stalled opportunities and draft personalized emails to each, demonstrating autonomous action capabilities. The key innovation is that one unified capture and memory layer powers multiple use cases through natural language interfaces rather than rigid predefined reports.

Lightfield's architecture features three interfaces—UI for humans, natural language agents, and workflow jobs for automation—all interacting through a unified domain objects layer that ensures identical permissions, business logic, and data access patterns. The critical principle is 'agent UI parity': if a user can access data, an agent can access it with full read, create, and update capabilities, eliminating separate agent APIs with different rules or limited access. This unified approach means the agent is simply another interface to the same data layer, not a separate system with different constraints. This architectural choice was deliberate from day one because building AI-native systems from the ground up beats bolting agents onto legacy systems where different interfaces inevitably end up with divergent logic, permissions, and capabilities that become impossible to reconcile.

The team chose AI SDK because nobody has a complete playbook for AI products in 2025, so they optimized for learning speed over perfection, deliberately choosing frameworks that provide flexibility to iterate rather than enforce rigid abstractions. They adopted AI SDK V4 in January 2024 for model switching and stream text primitives, then seamlessly adopted V5 alpha in June 2025 when custom transport options were released, shipping features in weeks without major rewrites. Jack emphasizes the principle that 'duplication is far cheaper than the wrong abstraction,' noting that building software with AI is faster than ever, making wrong architectural decisions even more costly than code duplication. The team noted an internal joke that they'd identify a feature they needed from AI SDK and the next day see a tweet from the team announcing it, exemplifying how a good framework grows with your product needs instead of forcing rewrites or technical debt.

Lightfield implements adaptive context building where users don't need to explicitly specify account names or meeting details; instead, they ask natural questions like 'what's next for this account' and the system intelligently retrieves relevant context. This feature leverages AI SDK's Data Parts API, which enables the client to send signals about entities and context to the server, where the system fully hydrates the information with intelligent retrieval. The Data Parts abstraction allows sending different types of entity signals from the client while the server-side retrieval logic determines what actually matters for the user's query. This approach maintains clean separation between client-side context signals and server-side retrieval logic, enabling flexible context construction without forcing the client to know what data might be relevant, creating a more natural and intuitive interface.

Nikita explains that while AI SDK makes file injection into chat threads straightforward through the use chat hook, this creates practical security concerns around avoiding direct database persistence and preventing accidental exposure of private user data when using S3 URLs. Their solution has the client send an internal ID referencing files in their own data store, then the backend replaces these IDs with signed S3 URLs before sending to external LLM providers, and the expiration time on signed URLs prevents unauthorized access. Additionally, Lightfield uses contextual tool collections where user IDs are directly injected into tools at runtime, ensuring the LLM never directly queries the database but always goes through the unified data layer that enforces user permissions. This architecture maintains the principle that agents operate with identical permissions and through identical data access patterns as the UI, while also protecting against information leakage through external API calls.

Lightfield maintains UI parity by allowing agents to create CRM entities like accounts, opportunities, and contacts through the same modal interface available in the UI, leveraging AI SDK's human-in-the-loop abstractions to require user confirmation for tool calls. When the LLM issues a tool call requiring confirmation, it forwards to the front-end client which renders an interface, and then the backend executes the function based on user input, appending the result back to the tool output so the LLM sees the complete interaction history. For example, if the LLM suggests creating multiple accounts with specific names and domains, the user can edit the values and approve individual items, and the LLM receives feedback about which creations succeeded, failed, or why (e.g., account already exists). This design ensures users stay in control of high-stakes actions while the LLM has full visibility into the complete flow, enabling it to appropriately suggest next steps and maintain user trust through transparency.

Lightfield enables customers to construct custom data models for each CRM entity, allowing a B2B productivity tool company to track their specific needs like customer tech stack, engineering team size, and mutual investors as typed custom fields. The agent receives instructions about how to use these fields in background workflows, such as backfilling fields through web research or searching through meeting transcripts and emails to enrich account information. The system creates tools at runtime with schemas derived from the company's configuration, validates that suggested values match the schema, and builds flexible, highly reliable tools that adapt to each customer's unique business structure. This customization enables a product that molds to how customers structure their data rather than forcing customers to adapt to predefined schemas, transforming Lightfield from a tolerated tool into one customers can't live without.

Lightfield includes a knowledge section where companies provide business context and instructions for background workflows like automated meeting prep, where the system generates preparation documents listing key attendees, account information, and discussion points before each meeting. After meetings, the system suggests follow-up action items and field updates based on discussion content, and can access Google Calendar to draft follow-up emails with appropriate meeting times based on previous discussions. All suggested actions—draft emails, action items, and field updates—remain gated behind user approval to ensure agents never take action without explicit authorization, maintaining user agency over business-critical decisions. These automation capabilities combine the full context of sales interactions with customized knowledge about the company and its customers to dramatically accelerate the sales workflow while respecting the requirement that users maintain control over customer communications and commitments.

Jack synthesizes four core principles discovered while building Lightfield: (1) Agent-UI secure parity designed from day one with agents having full read-write access through the same data layer, avoiding separate agent APIs; (2) Fast iteration over perfect abstractions, accepting duplication as genuinely cheaper than wrong abstractions when conventions are still forming; (3) Human-in-the-loop workflows with full transparency and history so users see original suggestions, edits, and execution results to build trust; and (4) Programmable systems by users and agents, recognizing that real customers need custom data models and the product must adapt to how customers structure their data rather than forcing standardization. Jack concludes by inviting the audience to share what they're building and what patterns they're discovering, emphasizing that these principles represent the difference between products people tolerate and products they can't live without. The talk demonstrates how thoughtful architectural decisions, framework selection, and commitment to user control enable powerful AI-native products.