Welcome to our exclusive AI Summit series! Over the next three months, we'll be bringing you weekly video sessions from HubSpot for Startups' annual AI Summit in San Francisco. Each week, we'll feature powerful insights from speakers representing tech giants like Anthropic, Atlassian, Clay, Hightouch, Replit, and HubSpot, alongside wisdom from the Valley's leading tech investors. This transformative content will help you understand and implement the AI technologies reshaping business today.
We're kicking off with two power-packed presentations that cut through the hype around AI agents to reveal how companies are actually implementing agentic AI to scale their operations and transform their workflows.
Breaking Down AI Agents: HubSpot's Framework for Success
In his keynote address, Nicholas Holland, SVP of Product at HubSpot, shared how the company is approaching AI agents with a mission to help small and medium businesses scale efficiently. With 50% of businesses failing in their first five years, Holland positioned agents as a potential "cheat code for scale" that could fundamentally change this trajectory.
Key Takeaways from Nicholas Holland's Keynote:
The Four-Layer Agent Architecture
Holland presented HubSpot's internal framework for understanding agents through four distinct layers:
- Data and Context Layer: The foundation combining structured data (CRMs, databases), unstructured data (80% of customer-created data), and external/world data (intent signals, web data)
- Intelligence Layer: Including reasoning capabilities (how agents make plans), memory systems (crucial for the next 12-24 months), prediction models, and skills (complex prompts that aren't full agents)
- Integration/Orchestration Layer: How agents work with existing systems and each other, including emerging protocols like MCP and A2A
- Agent-to-Human Experience: The critical layer focusing on adoption, training, and management
Real-World Impact at HubSpot
- Customer agents resolving issues 39% faster
- Clients creating 80% of their content through AI agents
- Prospecting agents booking 11,000 meetings in Q1 alone
The Human Element
Holland emphasized that technology might not be the limitation anymore—it's organizational adoption. He introduced concepts like "agent trainers" and "agent managers" as emerging roles, with some customers already requesting to see their AI agents listed alongside human employees in performance dashboards.
From Theory to Practice: Industry Leaders on Implementing Agentic AI
The panel discussion, moderated by Matt Thompson from AWS, brought together leaders from HubSpot, Turing, AWS, and Glean to discuss the practical realities of deploying AI agents today.
Key Insights from the Panel
Defining Agents vs. Agentic AI
The panelists offered complementary perspectives:
- Nicholas Holland (HubSpot): Agents move from completing tasks to doing entire jobs, eventually combining jobs into larger bodies of work
- Emrecan Dogan (Glean): AI agents follow instructions to complete tasks, while agentic AI sets its own course
- James Raybould (Turing): Likened the evolution from agents (musicians) to agentic AI (conductors orchestrating multiple elements)
- Randy DeFaux (AWS): Emphasized agents as workers that figure out how to accomplish tasks using available tools
The Augmentation Reality of 2025
- James Raybould stressed that 2025 is about augmentation, not replacement—what takes 100 people today might take 80-90 people
- Repetitive, low-risk tasks will see the most immediate impact
- High-strategy, high-autonomy, high-risk roles will take much longer to automate
Technical Challenges and Solutions
Randy from AWS outlined three critical layers for agentic platforms:
- Building blocks: Memory, observability, security guardrails, tool connections
- Orchestration frameworks: LangGraph, CrewAI, AWS Strands
- Scale infrastructure: The biggest gap—moving from a handful of agents to hundreds requires catalogs, registries, marketplaces, and security scaffolding
Adoption Over Technology
Multiple panelists agreed that organizational change management, not technology, is becoming the primary challenge:
- Nicholas Holland's recommended playbook: Every employee needs an assistant, every team needs to hire their first agent, then customize and build from there
- Emrecan highlighted the importance of "AI champions" within organizations who create visibility for how to use these tools effectively
- James Raybould introduced the "verify everything → verify some → verify important" framework for gradual AI adoption
Embracing Imperfection for Progress
A surprising consensus emerged around accepting higher error rates to accelerate learning:
- Nicholas shared how customers preferred a helpful agent who occasionally hallucinated over one that constantly refused requests
- James noted that if no hires have failed, the bar is too high—the same principle applies to AI agents
- The panel agreed that 90% accuracy is often good enough, especially when humans typically operate at 70-75% accuracy
Transformative Use Cases
- Personal productivity: Managers using agents to continuously track team performance rather than relying on biannual reviews
- Healthcare innovation: Research agents discovering new treatments from genes previously only used for diagnosis
- Organizational insights: Google using AI to identify high-performing but quiet engineers by analyzing codebases
The Path Forward
The panelists' closing thoughts painted an optimistic but realistic picture:
- Randy (AWS): We're not thinking big enough—with 5x productivity gains, what new markets could companies enter?
- James (Turing): Like Bill Gates' famous quote, we'll overestimate one-year progress but underestimate ten-year transformation
- Emrecan (Glean): The solution to AI problems is more AI—agent-to-agent communication creating infinite feedback loops for improvement
- Nicholas (HubSpot): The best software is the software you use—focus on moving from "novel" to "necessary" for customers
The session made clear that while the technology for AI agents is rapidly maturing, the real challenge—and opportunity—lies in thoughtfully integrating these tools into human workflows, accepting imperfection as part of the learning process, and thinking bigger about what becomes possible when every knowledge worker has AI-powered assistance.
~~~~~
AI Disclaimer: AI helped us summarize the key points from these videos, and our editorial team reviewed everything to make sure it's clear and correct.
