How to Build an AI Digital Twin in Cursor: Zapier's Internal Playbook
Most HR leaders spend days creating change management plans and tracking team goals manually. Zapier's Head of HR reveals how their AI digital twin automates these workflows - cutting planning time from 1 day to 10 minutes while providing longitudinal coaching no human could maintain.
The 3-Layer Digital Twin Framework
HR and people operations teams face a impossible challenge - tracking employee goals, analyzing meeting patterns, and planning organizational changes while maintaining human connection. Traditional tools create silos between knowledge (Google Drive), intelligence (ChatGPT), and action (Zapier).
Zapier's solution combines these three capabilities into what they call a "digital twin" - not just an AI assistant, but a persistent companion that grows with your role. At 4:22 in the video, Brandon Sammut demonstrates how their framework works:
The digital twin combines: 1) A context layer (company knowledge and personal goals) 2) An intelligence layer (multiple LLMs matched to tasks) 3) A connection layer (MCP protocols to 8,000+ apps). This creates what Sammut calls "one pane of glass" for automated people operations.
Building the Context Layer
The most common failure in AI implementations is lack of relevant context. Digital twins solve this through a structured folder system that's always accessible:
- Company HQ: Shortcut to Google Drive with strategy docs and product info converted to markdown
- Team Knowledge: People team playbooks and AI transformation docs
- Personal Context: README file with values/leadership principles + 6-month goals
At 7:15, Sammut reveals how their knowledge manager Elise solved the markdown conversion problem - building a tool in 24 hours that automatically converts all company wiki content to AI-readable markdown. This unexpected skills transformation is key to digital twin success.
Mixing Multiple LLMs
Unlike ChatGPT's one-model-fits-all approach, digital twins intelligently route tasks:
Cost/performance throttling: Fast/cheap models for simple queries, expensive/precise models for high-stakes analysis. At 9:48, Sammut shows how the twin selects models based on task complexity - like using Claude 3 for longitudinal meeting pattern analysis.
This multi-model approach solves two problems: 1) Reducing API costs by 60-80% for simple tasks 2) Preventing "model drift" where one LLM develops blind spots. The digital twin becomes more reliable as it learns which models work best for your specific workflows.
The Power of MCP Connectors
Model Context Protocols (MCPs) turn knowledge into action. Zapier's digital twin uses four core connectors:
- Databricks: Queries product usage and financial data
- Granola: Pulls AI-generated meeting notes and feedback
- Ordinal: Manages social media interactions
- Zapier: Accesses all 8,000+ connected apps
At 12:30, Sammut demonstrates how MCPs create what he calls "ambient automation" - where the twin can check Slack for unanswered questions, verify calendar time allocation against goals, and prompt action without explicit commands. This is the difference between an assistant and a true digital twin.
Weekly Accountability Co-Pilot
The first digital twin Sammut built addresses a universal leadership challenge - aligning time investment with stated priorities. At 14:22, he shows how the weekly co-pilot works:
- Data Collection: Pulls from Slack, Google Calendar, Granola meeting notes
- Analysis: Maps activities to personal goals and leadership principles
- Outputs: Markdown report + Google Doc + Slack summary with "stoplight" priorities
What makes this powerful isn't the automation, but the longitudinal tracking. At 16:45, Sammut shares an example where the twin flagged a 5-week pattern of "obtuse answers" in meetings - something no human would consistently track across dozens of interactions.
10-Minute Change Management
The most dramatic demo comes at 22:10, where Sammut shows how the digital twin creates comprehensive change plans:
From 1 day to 10 minutes: By combining the "Leading Change" skill with company knowledge and MCP connectors, the twin generates SCARF framework analysis, comms waterfalls, manager briefings and "root FAQs" - work that previously took skilled practitioners a full day.
This works because the twin: 1) Queries markdown-format playbooks 2) Applies Zapier's change framework 3) Pulls relevant data from past initiatives 4) Structures outputs in consistent templates. The quality is high because it's not generating from scratch - it's remixing approved components.
Scaling With Skills Libraries
At 28:40, Sammut addresses the adoption challenge - how to move from power users to companywide implementation. Their solution centers on skills libraries:
- Internal Skills: Coaching frameworks, context setup guides, SEO content agents
- Curated External: Lenny's Podcast library of interview-derived skills
- Skill Creator: Digital twin that helps build new skills through natural language
The key insight? "Get folks their first wow moment quickly." At 32:15, Sammut shows their "Find Your Next Best Skill" skill that helps employees discover pre-built automations for their specific role - creating the intrinsic motivation to build their own digital twin.
Watch the Full Tutorial
See the digital twin in action from 14:22-18:30 where Brandon Sammut demonstrates the weekly accountability co-pilot analyzing his meeting patterns and goal alignment over time.
Key Takeaways
Zapier's digital twin approach transforms three HR pain points: inconsistent coaching, slow change management, and disconnected people data. By combining structured knowledge, multi-model intelligence and ambient automation, they've created what Sammut calls "AI that grows with you."
In summary: Effective digital twins need 1) Markdown-format knowledge 2) Model throttling 3) MCP connectors 4) Skills libraries. The result isn't just automation - it's persistent, personalized augmentation that improves over time.
Frequently Asked Questions
Common questions about this topic
An AI digital twin is a holistic AI companion that combines three capabilities: a context layer (company knowledge and personal goals), intelligence layer (multiple LLMs), and connection layer (integrations with tools like Slack, Google Docs and Zapier).
Unlike chatbots that handle individual queries, digital twins maintain persistent access to your work context and can execute multi-step workflows autonomously. Zapier's implementation in Cursor serves as a "single pane of glass" for daily operations.
- Context layer: Structured knowledge in markdown format
- Intelligence layer: Right LLM for each task
- Connection layer: MCP protocols to take action
Cursor provides an integrated development environment specifically designed for AI-augmented engineering work. Its three key advantages for digital twins are:
1) Native markdown support for knowledge files 2) Easy switching between LLM providers 3) Built-in tools for creating Model Context Protocols (MCPs). At 6:30 in the video, Brandon Sammut demonstrates how Cursor's folder system creates the persistent context layer.
- For technical users: Combines IDE features with AI
- For HR/ops teams: Pre-built skills reduce coding needs
- Enterprise-ready: Supports permissioned access
Zapier uses a standardized process to build the knowledge foundation:
1) A "Create Team Brain" skill walks users through setup 2) Company knowledge is converted to markdown automatically 3) Personal README files capture working styles 4) Goal documents provide current priorities. At 8:45, Sammut shows how their knowledge manager built tools to convert wiki content at scale.
- Critical step: Markdown conversion for AI readability
- Ongoing maintenance: Skills auto-update changed files
- Security: Permissioned access to sensitive data
The weekly accountability co-pilot demonstrates the digital twin's power:
1) Pulls calendar/Slack data via Zapier MCP 2) Analyzes meeting notes through Granola MCP 3) Compares activities to goals documents 4) Generates three formatted reports. At 16:20, Sammut shares how it identified a 5-week pattern of unclear meeting responses no human would notice.
- Inputs: Goals + calendar + meeting notes
- Analysis: Multiple LLMs for different aspects
- Outputs: Markdown + Google Doc + Slack alert
By creating a reusable "Leading Change" skill that:
1) Queries markdown-format playbooks 2) Applies the SCARF framework 3) Auto-generates comms plans and FAQs. At 22:10, Sammut shows how what took specialists 1 day now takes 10 minutes. The twin isn't creating from scratch - it's remixing approved components with current data.
- Quality control: "Kevin" skill adds rigor
- Consistency: Uses company templates
- Speed: Parallel processing of components
Yes, with important caveats:
Zapier imports high-quality skills from sources like Lenny's Podcast library, but stresses careful review since skills execute with your permissions. At 29:50, Sammut demonstrates their "Find Your Next Best Skill" skill that helps identify and customize pre-built options for specific roles.
- Best sources: Lenny's, Anthropic's skill library
- Safety: Skills run with your access levels
- Customization: Most need role-specific tweaks
Three fundamental differences:
1) Persistent access to live company knowledge 2) Ability to mix LLMs based on task needs 3) Direct action-taking through app connectors. At 12:30, Sammut shows how MCPs enable "ambient automation" ChatGPT can't match - like automatically checking for unanswered Slack questions.
- Knowledge: Always-current vs snapshot
- Action: Can execute workflows
- Cost: Optimizes model selection
GrowwStacks specializes in building custom AI digital twins for HR, operations and executive teams. Our implementation package includes:
1) Context layer setup with markdown conversion 2) MCP connector configuration for your key apps 3) Development of your first 3 skills 4) Team training and adoption playbooks. We can have your digital twin operational in 90 days.
- Phase 1: Knowledge audit and conversion
- Phase 2: Core workflows automation
- Phase 3: Skills library development
Automate Your HR Workflows With an AI Digital Twin
Manual people operations create inconsistent coaching and slow response times. Our digital twin implementation delivers Zapier-level automation in 90 days - with your knowledge, your tools and your workflows.