2025, The State of AI Agents, and Engagement as the Ultimate Business Metric
Engagement > Growth
Good morning
Welcome back after the holiday break, and let’s start the first 2025 newsletter edition with a video:
In today's edition, among other things:
Books That Shaped My 2024
AI Agents: What's Actually Working
Engagement - The Key To Building An Enduring, Billion-Dollar Business
Startup Valuation Ranges
The Founder GTM Handbook
Understanding Startup Board Control
How AI Will Change Software Engineering
Onwards!
Books That Shaped My 2024
I didn’t have much time to read books in 2024, but I managed to read some over weekends. Here’s the list:
Artificial Intelligence: A Guide for Thinking Humans by Melanie Mitchell
The same AI systems that can beat grandmasters at chess still struggle with tasks any 5-year-old can do effortlessly, like understanding why a person can't use an umbrella that's been carried away by the wind.
A Brief History of Intelligence by Max Bennett
A fascinating dive into how octopuses evolved a different form of intelligence than vertebrates, suggesting there might be multiple paths to developing consciousness.
The Weirdest People in the World by Joseph Henrich
The Catholic Church's ban on cousin marriage in medieval Europe accidentally triggered a cascade of psychological changes that helped create modern individualistic societies.
The Science of Can and Can't by Chiara Marletto
Shows how focusing on what's impossible (like perpetual motion machines) has led to some of physics' greatest breakthroughs.
Reentry: SpaceX, Elon Musk, and the Reusable Rockets
The first successful landing of a rocket booster was compared to "trying to balance a rubber broomstick on your hand in the middle of a wind storm".
The Science of Rapid Skill Acquisition
The "85% rule" - practicing at a level where you're successful about 85% of the time leads to optimal learning.
The Human Use of Human Beings by Norbert Wiener
Written in 1950, predicted with eerie accuracy our modern challenges with automation and information overload.
Never Play It Safe by Chase Jarvis
The most successful creatives often have unconventional backgrounds - one of the world's top food photographers started by photographing skateboarders.
The Vital Question by Nick Lane
Makes a compelling case that life began in deep-sea hydrothermal vents, where conditions today are remarkably similar to those billions of years ago.
The Structure of Scientific Revolutions by Thomas S. Kuhn
Scientists initially rejected the idea of continental drift not because of poor evidence, but because they couldn't imagine how continents could possibly move.
Consolations and Consolations II by David Whyte
Explores how the word "alone" comes from "all-one," suggesting a completeness rather than loneliness.
Don't Believe Everything You Think
Most people spend 46.9% of their waking hours thinking about something other than what they're actually doing.
If you want to pick one from the list, pick Consolations or Consolations II.
AI Agents: What's Actually Working
The term "AI Agent" has become increasingly both popular (new startups) and murky (as vendors rush to rebrand existing tools - please see the mem at the end). But, as I wrote multiple times in the past two years - it’s one of the most important shifts in tech for years to come. Here’s Insights Partners:
Often a fun conversation starter, the academic definition — software that can reason on a task and take action independently — captures the high-level aspiration of AI Agents. We think of Agents as a new architecture combining core application logic and associated workflow automation in a unified flow, embedding LLMs to interweave planning and execution of complex tasks.
Agents could be as simple as a single-task Agent, which combines an LLM with a specific tool or function. Multi-agent platforms address complex workflow by breaking down the task into distinct Agents and modules, which are orchestrated to deliver the required output.
At its core, an Agent is software that can:
Take in a task description
Break it down into steps
Execute those steps using available tools
Adapt its approach based on feedback
The key distinction from traditional automation: Agents handle variance and uncertainty by replanning rather than failing when their happy path breaks.
Four Main Categories:
1. Task-Specific Agents
These focus on one well-defined job and do it reliably:
Rasa's Customer Service Agent
Handles support tickets end-to-end
Integrated with knowledge bases and ticketing systems
Clear ROI: 40-60% reduction in human agent time
Harvey (Legal)
Reviews contracts and flags issues
Cites relevant case law and regulations
Used by Allen & Overy for due diligence
The pattern: Take a repetitive knowledge-worker task, embed domain expertise, achieve reliability through constraint.
2. Workflow Agents
These coordinate multiple steps across tools:
Bardeen
Automates multi-step processes across SaaS tools
Example: Lead enrichment → CRM update → Email sequence
Key insight: Most knowledge work is chains of small tasks
GitHub Copilot
Suggests code completions
Generates tests
Explains code changes
The pattern: Replace manual context-switching and tool juggling with orchestrated sequences.
3. Platform Agents
These extend existing software platforms:
Salesforce Einstein GPT
Summarizes customer interactions
Drafts responses
Updates records automatically
Microsoft Copilot
Embedded in Office apps
Handles document creation/editing
Meeting summarization
The pattern: Add AI capabilities to where users already work rather than creating new destinations.
4. Specialized Interface Agents
These focus on specific interaction modes:
ElevenLabs Voice Agents
Natural voice interaction
Multiple languages
Used in customer service and education
Midjourney
Image generation and editing
Visual design assistance
Growing use in creative workflows
The pattern: Make complex technical capabilities accessible through natural interfaces.
AI Agents market map:
What's Working and What Isn't
Working:
Bounded tasks with clear success metrics
Integration with existing workflows
Clear handoffs between human and Agent
Strong guardrails and safety checks
Not Working:
General-purpose "AI assistants"
Agents that require perfect data
Complex multi-step tasks without human oversight
Replacing entire job functions
Also this, via CNBC:
Salesforce will hire 2,000 people to sell artificial intelligence software to clients, CEO Marc Benioff said on Tuesday, double the number the company indicated it was planning to add a month ago.
The cloud software company, which targets sales reps, marketers and customer service agents, is among the many technology companies hoping to boost revenue with generative AI features.
“We’re adding another couple of thousand salespeople to help sell these products,” Benioff said at a company event in San Francisco. “We already had 9,000 referrals for the 2,000 positions that we’ve opened up. It’s amazing.”
Three emerging pricing models:
Per-Task Pricing
Example: $X per contract reviewed
Works well for clear, measurable outputs
Harvey and other legal Agents use this
Capacity-Based
Example: $Y per Agent instance per month
Similar to hiring an employee
Common in customer service applications
Outcome-Based
Example: % of cost savings or revenue generated
Hardest to implement but most aligned
Emerging in sales and procurement Agents
Some considerations:
Enterprises
Start with bounded, measurable use cases
Invest in data quality and API infrastructure
Build expertise in Agent oversight and management
Builders
Focus on specific workflows vs. general intelligence
Build strong monitoring and safety features
Develop clear ROI measurement tools
Investors
Look for clear usage metrics and ROI stories
Prefer focused solutions over platforms
Watch for emerging middleware and infrastructure plays
And for everyone:
Engagement - The Key To Building An Enduring, Billion-Dollar Business
Keep reading with a 7-day free trial
Subscribe to Bartek Pucek to keep reading this post and get 7 days of free access to the full post archives.