AI Agents vs Workflows: What Actually Works Today (Zapier CEO Explains)
Most businesses struggle with knowing when to use deterministic workflows versus AI agents. Zapier CEO Wade Foster reveals the AI automation spectrum - where workflows excel, where agents shine, and the surprising "90% rule" that determines what you should automate first.
The AI Automation Spectrum
Business leaders often assume AI agents will completely replace traditional workflows, but Zapier CEO Wade Foster reveals a more nuanced reality. There exists an automation spectrum ranging from deterministic workflows to fully agentic systems, each with distinct strengths.
On the deterministic end (what we traditionally call workflows), systems follow predefined rules with perfect reliability. These excel at high-volume tasks like "when a new lead arrives, send a text message and add to CRM." The inputs, transformations, and outputs are completely predictable.
Key Insight: Most people who say they want an AI agent actually just need a well-designed workflow. The determinism of traditional automation often provides better reliability and lower costs for routine business processes.
At the opposite end sit pure AI agents that reason about unstructured inputs, make decisions, and take actions based on learned instructions. While powerful, these currently struggle with reliability for complex tasks - they might "keep spinning and burn a lot of tokens" without reaching useful outputs.
The sweet spot lies in the middle - AI-enhanced workflows that combine deterministic steps with targeted AI interventions. These "agentic workflows" maintain reliability while adding flexibility where needed most.
Real Email Triage Agent Case Study
Wade Foster shared a production example from his own workflow - an email categorization agent built with his executive assistant. The agent handles Wade's daily email volume (100+ messages) by:
- Categorizing emails into priority groups (requires Wade's attention, EA should handle, informational)
- Enriching customer emails with CRM data and web searches
- Applying appropriate labels and archiving non-essential messages
Result: Less than 10 emails per day actually require Wade's attention - a 90% reduction in cognitive load. The agent handles the triage work that would otherwise consume hours of human time.
This showcases a key principle: Effective agents start small and focus narrowly. The email agent began with simple categorization before gradually taking on more responsibilities as it proved reliable. Like training an employee, you incrementally increase an agent's scope based on demonstrated competence.
At , the most successful agents handle specific, well-defined tasks rather than attempting to be general assistants. As Wade noted: "The more you can constrain the task, the more likely it is you're going to get something you're pretty satisfied with."
The 90% Rule for AI Automation
A powerful framework emerged from Wade's experience: the 90% rule. AI can typically handle about 90% of a knowledge work task, with humans still needed for the final 10% of quality control and refinement.
This plays out in several ways:
- Email responses: AI drafts replies for 90% of messages, humans review sensitive communications
- Sales support: AI provides 90% of customer context, salespeople focus on relationship-building
- Content creation: AI generates 90% of draft content, humans polish the final version
The rule acknowledges AI's current limitations while maximizing its productivity benefits. As Wade explained: "You still need humans at the other side to review it and make sure that it's hitting your standards and your goals."
This approach also solves the "blank page problem" - AI handles the initial 90% of work that people often procrastinate on, while humans contribute their unique value in the final 10%.
How to Build Trust With Agents
Adopting AI agents requires a trust-building process similar to onboarding new employees. Wade recommends this progression:
- Observation: Start with the agent labeling/categorizing items (emails, tickets, etc.) without taking action
- Drafting: Have the agent create draft responses or content for human review
- Limited action: Allow autonomous actions only for well-understood scenarios
- Full delegation: Gradually expand scope as the agent demonstrates reliability
This "task-relevant maturity" approach (referencing Andy Grove's management philosophy) minimizes risk while building confidence. As Wade noted: "The worst thing you have is maybe an inbox that's a little cluttered with a bunch of drafts you don't like."
The key is starting with tasks where mistakes have minimal consequences, then expanding responsibility as the agent proves itself - exactly how you'd train a new team member.
APIs vs MCPs: Reliability Tradeoffs
The discussion revealed an important technical distinction between traditional APIs and Machine-Callable Procedures (MCPs):
| APIs | MCPs | |
|---|---|---|
| Structure | Deterministic inputs/outputs | Handles unstructured data |
| Reliability | High (predictable behavior) | Variable (depends on reasoning) |
| Cost | Low (efficient processing) | Higher (token-intensive) |
| Best For | High-volume routine tasks | Complex, variable scenarios |
Wade emphasized they're complementary technologies: "You're going to opt for APIs when you want reliability and cost advantages... MCPs let you tackle use cases you couldn't do with an API."
This aligns with the broader spectrum - APIs power the deterministic workflow side, while MCPs enable agentic capabilities where needed.
The Coming Orchestration Layer
Looking ahead, Wade sees a growing need for orchestration between specialized agents: "If you want to do more complex tasks, the best way is to build agents for each part of the task and orchestrate them."
This mirrors modern software architecture, where microservices each handle specific responsibilities while communicating through APIs. For AI, it means:
- Specialized agents focused on narrow tasks (email triage, research, drafting, etc.)
- Orchestration logic to sequence their work (either deterministic or agentic)
- Standardized interfaces for passing context between components
Interestingly, Zapier's existing workflow engine turns out to be well-suited for this orchestration role. As Wade noted: "This workflow engine we built over 14 years is really good at connecting agents together."
This suggests established automation platforms may have an advantage in the emerging agent ecosystem, contrary to predictions that AI would disrupt them.
Proven AI Adoption Strategies
With Zapier achieving nearly 100% internal AI adoption, Wade shared effective strategies for organizations:
1. Hackathons with Judging: Dedicated time for experimentation across all departments (not just engineering), with demos to share learnings.
"When everyone in your organization has dedicated time to put their hands on the keyboard, the fear fades to the background because they see what's possible."
2. Regular Show-and-Tell: Feature different teams' AI applications in all-hands meetings to inspire broader adoption.
3. Leadership Commitment: While "we must use AI" memos seem cringe, Wade believes clear executive signaling is essential - provided it's followed by concrete support.
The key insight? Hands-on experience transforms fear into opportunity. Employees who experiment with AI tools quickly move past existential concerns to practical applications that make their jobs easier.
AI Fluency in Hiring
As AI becomes integral to work, assessing candidates' AI fluency grows important. Zapier has evolved their approach:
- Self-reported experience: Early on, simply asking "What are you doing with AI?" revealed meaningful differences.
- Practical tests: For PM roles, giving candidates an hour with AI tools to solve a sample problem.
- Creative application: Evaluating how candidates leverage AI creatively within their domain.
Wade cautions against "outsourcing thinking to AI" - the goal isn't using AI for its own sake, but enhancing human capabilities. Effective candidates demonstrate:
- Understanding of AI's strengths/limitations
- Ability to frame problems for AI assistance
- Skill in evaluating and refining AI outputs
This aligns with the 90% rule - the best professionals use AI for the "middle work" while applying human judgment to inputs and final outputs.
Watch the Full Tutorial
For a deeper dive into Wade Foster's email triage agent (shown at 12:45 in the video) and additional examples of AI workflows in production, watch the full 38-minute discussion:
Key Takeaways
The AI automation landscape requires pragmatic choices, not hype-driven assumptions. Wade Foster's insights reveal a maturation beyond "agents vs workflows" to thoughtful integration:
In summary:
- Most "agent" needs are better solved with enhanced workflows
- Effective agents focus narrowly and prove reliability gradually
- The 90% rule maximizes AI's productivity boost while preserving human oversight
- Internal adoption requires hands-on experimentation, not just mandates
- Orchestration between specialized agents represents the next frontier
For businesses, this means starting with concrete pain points rather than chasing buzzwords. As Wade demonstrated with his email workflow, targeted AI applications can deliver transformative efficiency gains today.
Frequently Asked Questions
Common questions about this topic
Workflows are deterministic automations that follow predefined rules with perfect reliability. They excel at high-volume, routine tasks where consistency matters most.
AI agents can reason about inputs, make decisions, and take actions based on learned instructions. They handle more complex, variable situations that require interpretation, but with less predictable reliability.
- Use workflows when you need guaranteed outcomes (data processing, notifications)
- Use agents when you need adaptability (customer inquiries, content generation)
- Hybrid approaches often work best - deterministic workflows with targeted AI steps
The 90% rule suggests AI can handle about 90% of many knowledge work tasks, with humans still needed for the final 10% of quality control and refinement.
This plays out differently across functions:
- Email management: AI can categorize and draft responses for 90% of messages
- Content creation: AI generates 90% of draft content, humans polish the final version
- Data analysis: AI processes 90% of raw data, humans interpret key insights
The most practical AI agents currently in production focus on specific, narrow tasks rather than general assistance.
Email triage agents like Wade Foster's example automatically categorize incoming messages, enrich them with CRM data, and prioritize based on learned rules. These can reduce inbox volume by 90%.
- Coding assistants that handle routine boilerplate or suggest improvements
- Customer support agents for common, well-documented inquiries
- Research assistants that compile information from known sources
Building trust with AI agents follows a similar progression to training human employees:
Start with low-risk observational tasks where mistakes have minimal consequences (like labeling emails). Gradually increase responsibility to drafting content, then limited autonomous actions, before finally handling complete tasks independently.
- Phase 1: Observation only (categorization, labeling)
- Phase 2: Drafting outputs for human review
- Phase 3: Autonomous actions in constrained scenarios
- Phase 4: Full delegation as reliability is proven
APIs provide more reliable, cost-effective solutions for deterministic tasks where inputs and outputs are well-defined. They offer predictable behavior and efficient processing.
MCPs (Machine-Callable Procedures) allow more flexibility for agents to reason about unstructured data and determine appropriate actions, but with higher token costs and potential reliability trade-offs.
- Choose APIs for high-volume, routine data processing
- Choose MCPs when you need to handle variable, unstructured inputs
- Hybrid approaches often work best - APIs for core processing with MCPs for edge cases
Effective AI adoption requires more than executive mandates - it needs hands-on experience across the organization.
Zapier's successful strategies include hackathons with dedicated experimentation time, demo sessions to share learnings between teams, and regular show-and-tell of practical applications during all-hands meetings.
- Hackathons: Provide protected time for exploration
- Demos: Spread knowledge of what works
- Showcases: Highlight applications from different departments
The best AI agents are narrowly focused on specific tasks with clear boundaries. They perform better when given minimal necessary context and tools rather than broad access.
Like specialized employees, focused agents with clear responsibilities outperform generalists trying to handle too many diverse tasks. They're also easier to monitor, improve, and integrate with other systems.
- Narrow focus: One well-defined responsibility
- Clear boundaries: Limited tools/context to prevent confusion
- Measurable outcomes: Specific success criteria
GrowwStacks helps businesses implement the right mix of AI agents and deterministic workflows tailored to their operations. We design systems that combine reliability with adaptability.
Our team can assess your processes, recommend where AI adds the most value, and build customized automation solutions. Whether you need email triage agents, AI-enhanced workflows, or full automation systems, we handle the technical implementation so you can focus on your business.
- Process assessment: Identify automation opportunities
- Custom development: Build tailored AI agents and workflows
- Integration: Connect with your existing tools and platforms
- Free consultation: Discuss your automation goals with our experts
Ready to Implement AI Agents in Your Business?
Every day without automation costs your team hours of repetitive work. GrowwStacks can design and deploy customized AI agents and workflows tailored to your specific needs - often in just 2-3 weeks.