Step-by-Step Teaching Guide
STEP 1: The Foundation — Start With Outcomes, Not Tools
The Core Mistake Most Organizations Make
Many organizations think:
"Give people AI tools → Productivity improves"
This is incomplete thinking.
The Correct Mental Model
AI Tools → Changed Workflows → New Bottlenecks → Adaptation → Desired Outcome
Real Example From Dropbox
| Action | Unexpected Consequence |
|---|---|
| Engineers code faster | Too many code reviews pile up |
| More code generated | Infrastructure and development tools get overloaded |
Key Takeaway
AI is a starting point, not a solution. You must examine the entire workflow end to end and keep adapting it.
"There's no silver bullet." — Ali Dasdan, CTO
STEP 2: Driving Adoption — The Two-Layer Approach
Simply telling employees "use AI tools" is insufficient. Effective adoption requires two simultaneous layers:
Layer 1: Top-Down (Leadership Direction)
- Leadership mandates and encourages AI usage
- Sets strategic direction and goals
Layer 2: Bottom-Up (Ground-Level Engagement)
- Boot camps — hands-on training
- Show-and-tells — peer learning and demonstration
- Aligned incentives — managers understand what ROI they are getting
The Product Mindset Principle
Treat AI deployment like a product launch:
- Understand your users (employees)
- Gather feedback continuously
- Iterate based on real usage data
"You have to approach it with a product mindset." — Uma Namasivayam
STEP 3: Measuring AI Productivity and ROI
This is one of the hardest challenges in AI deployment. Here is how to think about it clearly.
The Two-Tier Measurement Framework
Tier 1: Real Outcome Metrics (Ultimate Goals)
These are what actually matter to the business:
| Metric | What It Measures |
|---|---|
| Revenue | Did AI help generate more income? |
| Cost reduction | Did AI lower operational expenses? |
| Customer satisfaction | Are customers happier? |
| Retention | Are customers staying longer? |
Tier 2: Proxy Metrics (Leading Indicators)
These connect daily AI activity to business outcomes:
| Category | Example Metrics |
|---|---|
| Speed | How fast are features shipped? |
| Effectiveness | How many experiments are run? |
| Quality | Change failure rate (how often deployments break things) |
| Impact | Pull request throughput (rate of completed code changes) |
Critical Warning
Producing more does not automatically mean producing better.
Example: Writing one million lines of code per day means nothing if it is the wrong product that customers do not want.
Dropbox's Real Performance Numbers
- ~70% AI-generated code (comparable to Uber at the same level)
- Top 5% in pull request throughput among similarly complex companies
- 75th percentile in change failure rate (quality maintained while moving faster)
- Among the lowest token usage in peer group (indicating efficiency)
STEP 4: Making Investment Decisions — Where to Spend on AI
The AI Spending Portfolio
Organizations typically invest across multiple layers simultaneously:
Individual Function Tools (e.g., specific team software)
+
Platform Integrations (e.g., AI features in Zoom, Slack)
+
Company-Wide Deployments (e.g., ChatGPT Enterprise)
+
Automation Tools
+
Coding/Development Models
The Core Investment Challenge
Organizations must decide which ideas to pursue before they can prove ROI. This requires:
- Domain expertise — understanding where AI can genuinely help
- Right incentives — ensuring decision-makers are rewarded for good choices
- Judgment — accepting uncertainty and making informed bets
Smarter Spending: Value Over Volume
Old thinking: How many AI tokens are we consuming?
New thinking: What engineering value do those tokens actually create?
Dropbox's internal tool Nova connects agent usage directly to engineering workflows and outcomes — measuring value produced, not just resources consumed.
STEP 5: Human Skills That Become MORE Valuable With AI
This is counterintuitive but critical. As AI handles more execution, human judgment becomes the scarce resource.
The Three Most Valuable Human Skills
1. Problem Solving
AI can help solve problems faster and at greater scale, but humans must:
- Define the right problem in the first place
- Provide accurate context to AI agents
- Ask the right questions
- Know how to iterate on answers
- Understand how solutions fit into broader systems
2. Judgment
As AI lowers the cost of producing software toward near-zero:
- Choosing the right problems becomes the primary competitive advantage
- Evaluating AI output for validity, accuracy, and appropriateness becomes essential
- Poor judgment at low cost = massive amounts of wrong work produced quickly
3. Leadership
Human leaders must:
- Navigate ambiguity as work fundamentally changes
- Build and maintain team trust
- Communicate clearly across changing roles
- Align incentives across the organization
- Make consequential decisions that AI should not make alone
Summary Table
| Skill | Why AI Makes It More Important |
|---|---|
| Problem solving | AI executes — humans must define what to execute |
| Judgment | More output means more opportunities for costly mistakes |
| Leadership | People and culture challenges grow as workflows transform |
| Communication | Agents need clear context and specifications from humans |
| Systems thinking | AI solutions must fit into larger organizational systems |
STEP 6: What Distinguishes Organizations That Use AI Effectively
The Context Problem
Even the most capable AI models fail without:
- Knowledge of which information the organization trusts
- Awareness of decisions already made
- Understanding of who has access to what
- Clarity on what needs to happen next
Without context → Generic, incomplete, or wrong results
The Winning Formula
Capable AI Model
+
Organizational Context (files, history, decisions, permissions)
+
Human Judgment (choosing problems, evaluating output)
+
Durable Home for AI Output (saving, sharing, reviewing, approving)
=
Sustained Competitive Advantage
STEP 7: The Broader Vision — AI Across the Entire Organization
Current State (Engineering Focus)
Most organizations, including Dropbox, have concentrated AI productivity gains in engineering teams first because:
- Output is measurable (code, deployments)
- Workflows are well-defined
- Tools are more mature
Future State (Organization-Wide)
The goal is extending AI-enabled workflows to:
- Product teams
- Design teams
- Customer experience teams
- Any team that identifies problems and builds solutions
Concrete Vision Example
A customer experience team spots a recurring customer issue → Uses AI agents to develop a solution → Tests it with real customers → Ships a fix
This represents a fundamental shift: non-technical teams gaining the ability to build and ship solutions.
Consolidated Summary
| Concept | Core Lesson |
|---|---|
| AI Deployment | Start with desired outcomes, not tools |
| Adoption | Combine top-down mandates with bottom-up engagement |
| Measurement | Use both real outcome metrics and proxy metrics |
| Investment | Measure value created, not resources consumed |
| Human Skills | Problem solving, judgment, and leadership grow in importance |
| Competitive Edge | Context + human judgment + durable workflows |
| Future Direction | Expand AI productivity beyond engineering to all teams |
Self-Check Questions
- Why is giving employees AI tools alone insufficient for productivity gains?
- What is the difference between real outcome metrics and proxy metrics? Give two examples of each.
- Why does AI increase rather than decrease the importance of human judgment?
- What three human skills does Dropbox identify as most valuable in an AI-enabled workplace, and why?
- What does it mean to measure AI spend by value created rather than tokens consumed?
- What organizational conditions must exist for AI models to produce useful rather than generic results?