Video summary

The AI Workflow That Puts You in the Top 1% | Practical Steps to Level Up

Main summary

Key takeaways

Business

Executive takeaway

Effective AI use isn’t “having a tool”—it’s building an internal workflow system (agents + connected data + reusable “skills”/SOPs) that reduces weekly operational overhead while keeping humans responsible for strategy and approvals.

Adoption reality check (why most companies fail)

  • Only ~15% of CEOs (as referenced) are getting “real value” from AI.
  • Reported productivity gains depend on breadth:
    • 45% report real productivity gain when using AI for 1–2 tasks
    • 90% report real productivity gain when using AI for 7+ tasks
  • Training/enablement matters:
    • Successful users spend at least ~8 hours/week building AI skills
    • Some companies measure AI value in performance reviews, but without onboarding + clear use cases, it becomes a checkbox and overwhelms employees.
  • Operational constraint:
    • 58% of US companies require employees to use AI (per the referenced employee research).

Framework: “AI as an operating system” (connect context + automate process)

A recurring pattern across leaders:

  1. Connect AI to where work happens (emails/docs/Slack/spreadsheets/CRM-like records).
  2. Give it process context so it doesn’t require constant manual prompting.
  3. Package repeatable workflows into reusable units (“skills”).
  4. Use it for execution + analysis, but keep humans for strategy.

Playbook: Turn repeatable work into reusable “skills” (process library)

Concrete process/productization approach:

  • Identify a task you repeat (especially weekly).
  • Convert it into a skill:
    • Save the prompt/workflow once
    • Run it repeatedly with consistent inputs
    • Improve the skill via feedback loops (agent conversations + prompt updates)

Improvement loop (recursive learning)

  • After using a skill, hold a back-and-forth with the AI.
  • Ask AI to update the skill so it gets closer to the desired output “in one shot”.
  • Review updates (human-in-the-loop).

Case example #1 (Kalshi cofounder): Weekly company-state reporting + performance follow-through

Operational need: a fast-growing team needs weekly visibility:

  • what changed across the company
  • what’s falling behind
  • who promised what vs. who didn’t follow through

AI workflow components

  • Company-wide updates pulled from: emails, docs, Slack
  • AI produces:
    • summaries for weekly planning/state-of-things
    • tracking signals such as over-promising / under-delivering
  • Build/enablement:
    • an AI team is building enterprise agents
    • major operational hurdle: permissions/access control (including legal/surveillance sensitivity)

Concrete “agent” KPI/output idea

  • A bot in chat can summarize:
    • how many times a person caused decision rework (e.g., “pulled back on decisions”)
    • how many times re-explanations were needed (Used to improve decision throughput + accountability.)

Case example #2 (Founder / creator system): Spreadsheet-native analytics + reusable weekly prompts

Productivity system

  • Goal: replace manual re-typing with reusable workflows.
  • Example workflow:
    • use “AI sheets” to:
      • rank items (e.g., last 30 episodes) by views
      • tag by topic
      • compute top 3 outperforming topics
  • Mechanism:
    • convert the workflow prompt into a one-click skill used by the whole team

Operating principle

  • AI acts as analyst/co-worker; humans decide strategic action.
  • Example rule stated explicitly:
    • “Never ever let AI make your strategic decisions.”

Case example #3 (Skills to media production pipeline): Document SOPs → agent-run execution

Described process breakdown for content production:

  • steps such as guest scoring, packaging, dossier briefing, teaser, distribution, tracking
  • approach:
    • write down the repeating processes
    • create a skill library where each step becomes a skill
  • starting data sources:
    • existing SOPs
    • 40+ full episode transcripts
    • performance tracker to evaluate outputs

Key outcome target

  • Achieve “AI gets me to 90%; the last 10% is human judgment.”

Case example #4 (Box / Aaron Levie): Automate research + prototype faster

Recommendations

  • Start with mainstream tools (Codex, Claude, Perplexity are named).
  • “Automate a process” and run bigger research tasks as an agent:
    • e.g., analyze ~100 companies worth of trends without manual Google searching
  • Markets/product research:
    • AI replaces the traditional research workflow by “fan out” scanning + source verification by a human.

Cost/realism framing

  • Encourage safe usage:
    • understand how agents query other systems (email/data connectors)
    • verify outputs rather than fully trusting blindly

Case example #5 (Khan Academy / Sal Khan): Token cost management + rapid prototyping ROI

AI spend + engineering use

  • Khan Academy spends about $1.2M/year on AI.
  • Cost driver:
    • multiple agents used by engineers for coding + reviewing (described as 5–10 simultaneous agents).
  • Compute cost example:
    • one engineer spent about $3,000 in a day (compute costs), framed as acceptable because it reduced time from 3–4 months to a few hours.

Decision criteria (business execution)

  • Pay for experimentation when it produces:
    • rapid prototyping
    • faster cycle time to ship:
      • feature “fit” target: once “by next school year”
      • with AI/hackathon development: shipped in ~1 month

Governance guidance

  • connectors to Slack/Gmail/Docs are approved
  • employees are trained:
    • don’t let AI act autonomously without review
    • e.g., don’t auto-post/send emails; draft + human check
  • Chief-of-staff usage:
    • AI reads Slack/Gmail/Docs and flags what’s “falling through the cracks”
    • acts like an operational coordination layer

Risk control

“Be critical of me / push back” against seductive AI suggestions

  • Emphasis: token-maxing should be intentional because cost can be unnecessary for the scale of the org.

Metric & KPI mentions (explicitly stated or implied)

  • Productivity adoption / output:
    • 52% of employees use AI at work
    • 15% use AI daily
    • Productivity gain: 45% (1–2 tasks), 90% (7+ tasks)
  • AI adoption benchmark:
    • 15% of CEOs getting “real value” (as referenced)
  • Time investment:
    • successful users: ≥8 hours/week building AI skills
  • Cost:
    • Khan Academy AI run rate: ~$1.2M/year
    • Example compute spend: ~$3,000 in a day
  • Business growth/financial KPI:
    • GenSpark crossed $250M annual recurring revenue (ARR) in 12 months (framed as proof of “system” approach)
  • Engineering throughput/cycle time:
    • time reduction: 3–4 months → a few hours
    • shipping: ~1 month vs. later-cycle expectation
  • Operational performance:
    • “AI gets to 90%” output quality; human retains 10% judgment responsibility

Actionable recommendations (what to do next)

  • Pick one weekly repeatable task and turn it into a skill (reusable prompt/workflow).
  • Connect AI to your workflow sources (Slack, Gmail, Docs, spreadsheets) so it has context.
  • Build a permission/access model for agents (especially for legal/compliance/sensitive operations).
  • Measure value as outcomes, not tool usage checkboxes:
    • track how often AI-enabled work reduces rework, improves follow-through, accelerates shipping.
  • Use AI for execution + analysis; keep humans for strategic decisions (explicit rule repeated).
  • Avoid blind token-maxing:
    • create a rule to assess whether automation is actually worth the cost/time.

Presenters / sources mentioned

  • Lana Lopes — cofounder of Kalshi
  • Eric (unnamed in subtitles) — runs the Digital Economy Lab at Stanford
  • Eric (also the main host/creator in the video) — builds/tests AI workflow system; also references GenSpark
  • Peter Yang — former product lead at Meta, Reddit, Roblox; runs a media empire; creator of “skills”/workflow automation
  • Aaron Levie — CEO of Box
  • Sal Khan — founder of Khan Academy
  • Boris Cherny — referenced via a tweet (“doesn’t prompt his AI anymore; it prompts itself”)

Original video