Learn How the Best Work with AI — and How to Train Others to Work Just Like Them (part 2)

In the last article, I introduced a concept for training and upskilling people to adopt AI effectively and supercharge their productivity. In this follow-on, I want to dig into the typical challenges with professional training and how AI can make a meaningful difference with the right approach.

In my experience observing AI adoption — across my peer group, in the media, at conferences — the story looks fantastic. Companies are buying licenses for AI models and features. Every vendor's new AI capability comes with compelling quotes about impact. But if you sit next to actual knowledge workers — Product Managers, Software Developers, Research Analysts, UX Designers, Cybersecurity Experts — you see something else.

Most people use AI regularly for search, rewriting emails, or drafting content. They get middling results from basic prompting — sometimes improved, sometimes slop. Many get stuck in old habits because "I don't have time to figure this out right now."

But a small group of people are killing it. They've quietly invested the time to learn, practiced with tools, and redesigned their workflows around AI. In my personal observations it feels like they're easily 2–5× more productive compared to the average person in the same role. But a recent report by OpenAI revealed that it's closer to a 6x productivity gap between AI power users and everyone else. And more importantly, they're poised to get even better as AI models and tooling improve.

Combined with the report where BCG recently noted that only ~6% of companies have started any meaningful AI upskilling program. Everyone else has essentially done: "Tools first, skills later (or never)."

The result is a growing divide inside organizations:

• A small group (maybe 2%) of AI-skilled "super workers" who ship more, faster, and often with better quality
• A much larger group of smart, capable people whose workflows haven't changed much at all

So while license adoption is high across organizations, actual workflow change is low — and your usage analytics will probably confirm this. If left alone, the gap between AI-skilled and AI-unskilled will continue to widen.

This is bad on multiple fronts:

• Career risk for individuals who never learn to work effectively with AI
• Productivity drag for teams when slower colleagues become bottlenecks or regularly deliver un-reviewed slop
• Retention risk if super-skilled people get frustrated carrying the load and leave
• Missed ROI for companies — and eventually for economies — in productivity and competitiveness

Not to mention the fact that a generation of young adults entering the workforce are going to be the worst hit as junior skills and entry level tasks get automated away by AI. Young people will have to spend more time training, shadowing and being mentored or apprenticed by experts while honing their AI competencies to make that jump from 1x to 6x more impactful as a hire. If left unaddressed, this gap in skills can lead to bankruptcies, job losses, and widening gaps between the haves and have-nots based on how effectively you can use technology to operate and grow a business.

The Core Problem: We're Training People for AI the Wrong Way

Most people aren't doing much, if any, formal training on how to maximize AI in their role. Partly because there's not a lot available beyond YouTube tutorials and blog posts. But best practices and patterns are emerging — there's a growing set of AI strategies that can be applied to specific skills.

Through educational research, we actually know a lot about how people build durable skills. You don't get good at something by watching a presentation. You get good by doing the real thing, with:

1. Clear goals — understand what "good" looks like for a given task. And expanding that understanding of 'good' by setting progressively challenging goals to master techniques and learn new skills.
2. Repeated practice — not once, but many times and embedded across real tools used in your role. You don't have to reach 10,000 hours, but investing a couple hours a week now for can save you dozens of hours a week forever.
3. Fast, specific feedback — not vague praise, but targeted correction from the persona of your boss, a customer, a board member on real work assignments and drafts, so that it's tied to real outcomes.

Most corporate AI training lacks most if not all of this. It tends to look more like one-off webinars or Lunch-and-Learns on "The Basics of ChatGPT", Slack channels full of link-sharing, emoji reactions, and AI user tips and whatever blogs, podcasts, tutorials and courses that keeners can find for free or in org approved Learning and Growth subscriptions.

The result is sub-optimal adoption, inconsistent quality, and no clear way to measure ROI on your AI investments. The missing piece is simple to describe and hard to execute:

People don't just need to know "how AI works," but "how you should work with AI, in your role and current work environment.".

If you zoom out, this is the macro thesis: The people who win in the next decade aren't "AI experts." They're domain experts who've rebuilt their craft around AI. It takes time and energy to get there. But it's a marathon and the race has just started.

I already see the prototypes of those people in Product Management, Software Development, CyberSecurity, Marketing and more. The problem is the playbook is trapped in their heads and their Slack threads and their notebooks. They're spending all their time trying to keep up and experiment — they don't have time to write things down or train others.

We need to capture those playbooks (by role), use them to 'train the trainer', or train our AI Coach in this case, and then use an army of these AI personal trainers to coach the next 10,000 people in each role up the curve.

Why This Is Finally Doable

I've been thinking about this problem since Google announced LearnLM and Khan Academy partnered with OpenAI on Khanmigo — and even more once my teenagers started using these tools to help with school learning. Despite all the talk about AI's potential to transform education, I'm surprised there hasn't been a more meaningful breakthrough for professional skill-building.

I haven't come across anything game-changing from vendors in the Online Learning space — LMS platforms, professional training providers, K-12 classroom tools — have done minimal innovation. Google's LearnLM, and learning modes in ChatGPT and Claude are great starts, but they're mostly one-shot interactions without consistency or focus on the learner's progress over time. Yet they have the potential to do so much more, and it's encouraging that the major foundation model providers

A couple of things have shifted in the last 18–24 months that make this solvable:

1. AI Can Now Act Like a Coach, Not Just an Answer Box

Newer modes like Claude's Learning Mode, ChatGPT's tutoring features, and Google's LearnLM-based models can ask you questions instead of just replying or provide hints and feedback without giving away the answer to assess understanding. It can scaffold your reasoning step by step and provide critical feedback on work files. Over time, it can assess your skills and adjust based on your level, with consistent guidance from beginner to master.

In education, the race to build the best "AI tutor" has already started, with startups focused on (ToDo: insert examples ) and incumbents like (ToDo: insert examples) quickly following suit. But none of them have really impressed me so far as a way to upskill effectively or demonstrated meaningful results. It's early, and it's getting alot of attention so I am enthusiastic about breakthroughs in the coming year. But, I'm also not going to wait around for somebody else to solve it. AI has inspired plenty of hypotheses and new approaches to throw at this age old problem and I am throwing my hat in the ring (or stick in the middle). I think we have enough of the parts to build different solutions today that can teach a Product Manager how to use AI to produce better specs, a developer how to safely use AI for refactoring, or a security analyst how to use AI for alert triaging.

2. Why Can't Foundation Models in Teaching Mode solve this?

I wish that chatGPT, Claude or Gemini could help me accelerate my day-to-day work with a toggle button. As a Product Manager, I wish I could train it on my products and use that to support our users, enable our sales teams and many other things. But I've tried everything that comes from Computer use and browser Operator models to the most expensive versions of chatGPT and none of them can effectively walk me teach new workflows in standard business applications, software and operational tools. The big models are generalists. And they don't know what version of software you're using, what permissions you have if it even knows about it. Out of the box, the frontier models don't know your tools, configs, or custom workflows. They don't know what your definition of "good" for a design document or sales email is. They don't have deep knowledge of your domain policies or company processes. This is why generic AI doing training often feels... lacking. It's capable, but not situated in your reality. I think there's an opportunity to build that missing layer:

First understand "Here's how work actually gets done in this org, by these roles, to this standard"
→ Now use that to power coaching for AI assisted and agentic workflows.

The big models provide the intelligence. We can do a better job of providing the context that makes it applicable to your daily work and optimizing it based on your best performers approach.

The Idea: A Skill Graph + Learning Engine in a Virtual Coach that Teaches New AI Skills

The concept has two main components that, over time, become a coaching pipeline organizations can use to scale training — and eventually translate into semi-automated agentic workflows.

Component 1: The Workflow Capture → Learning Graph → Coaching Engine

This is the generalized backbone:

• Capture how people actually work
• Structure that into a Learning Graph for each role
• Drive personalized coaching and practice from the graph

Component 2: An AI Coach That Lives Where Work Happens

Once the Learning Graph and context are in place, the coach:

• Guides people through using AI on their real tasks
• Provides feedback in real time
• Adapts to their progress

The Longer-Term Extension

As we observe how people and AI solve tasks together, we can translate some workflows into agentic automations — using or partnering with existing AI orchestration platforms — while still coaching humans on the higher-order parts that require judgment.

Let me unpack each piece.

1. Capturing Workflows (Without Boiling the Ocean)

This is the messiest part, and I don't have a magically perfect technical solution yet. But there's a workable starting point.

The goal isn't "document everything about how this company works." It's:

Identify 3–5 high-leverage roles and workflows, then capture just enough of how the best people do them to build useful coaching.

For a first phase, I'd focus on archetypal, high-impact knowledge-work roles where I have access to strong domain experience:

• Product Management
• Software Development
• Cybersecurity / SecOps

Then expand into Digital Sales & Marketing, and potentially Finance or Legal.
Within each role, we define a small set of archetypal workflows:

How to Capture Them (v1)

Structured interviews and guided walkthroughs:

• Sit with top performers and ask them to walk through real past work
• Ask "why" at each decision point
• Capture their mental checklists, "red flags," and shortcuts

Artifact analysis:

• Ingest good examples of past work (docs, tickets, reports, campaigns)
• Review existing templates and playbooks
• Understand tool configs and constraints (CI/CD rules, detection rules, CRM setups)

Lightweight observer mode (optional, later):

• With explicit consent and privacy safeguards, let the system "watch" real workflows in tools
• Spot patterns in how tasks are completed
• Refine the model of "typical" vs "expert" behavior

At this stage, this is mostly consultative: real humans doing structured discovery, with AI helping summarize and extract patterns.

The key is that all of this feeds into a structured representation…

2. The Learning Graph: "How People Get Good at This Role, With AI"

The output of workflow capture is not a PDF. It's a Learning Graph:

A structured representation of tasks, sub-skills, decision points, and quality criteria for a role — including where and how AI should be used.

For each workflow, the graph encodes:

Tasks & Sub-tasks

e.g., "Write PRD" → clarify problem → define success metrics → map use cases → identify risks → align stakeholders

AI Interaction Patterns

For each step, we tag:

• AI-delegable: Let the model generate, human reviews
• AI-assisted: Brainstorm together, explore options
• Human-only: High-risk judgment, policy-sensitive calls

Quality Rubrics ("What Good Looks Like Here")

Examples:

• A strong incident report at this company: clear root cause, explicit customer impact, crisp timeline, actioned follow-ups
• A strong outbound sequence for this ICP: tailored, insight-driven, compliant with brand and legal constraints

Progression Levels

• Level 1: Can follow a coached flow
• Level 2: Can independently choose when/how to involve AI
• Level 3: Can design new AI-assisted workflows for others

This Learning Graph is what makes the coaching contextual rather than generic. It drives:

• What scenarios the coach generates
• What feedback it gives
• How it adapts as the learner improves

Crucially, this engine is what you'd focus on building: a generalizable, configurable layer that can represent new roles and workflows over time — not a one-off consulting deliverable.

3. The AI Coach: In the Flow of Work, Not in a Separate Training App

With a Learning Graph in place, you put an AI coach on top that does three things:

1. Guides you through real work
2. Critiques what you produce
3. Teaches you how to use AI intelligently, not blindly

Crucially, it lives where people already are: in the doc editor, issue tracker, IDE, CRM, or chat (Slack/Teams).

Its job is not to be the AI that does the work for you. Its job is to:

• Help you use AI assistants more effectively on your real tasks
• Provide deliberate practice and feedback as you go

Guided Mode — "Walk Me Through This"

A PM says: "I need a PRD for X."

The coach:

• Pulls context from tickets and notes
• Walks them through the sections their org expects
• Shows exactly how to use an AI assistant at each step
• Enforces the internal rubric for "good enough to circulate"

The coaching style matters here. The coach doesn't just give answers — it asks questions that build judgment: "What would make engineering push back on this scope? Have you addressed that?"

Feedback Mode — "Critique What I Did"

A security analyst pastes an incident report.

The coach:

• Scores it against the internal standard
• Highlights missing or weak sections
• Suggests edits and explains the why

Simulation Mode — "Let Me Practice Without Consequences"

This is where the coach creates immersive practice scenarios. Imagine:

You're dropped into a simulation: The CEO needs a pitch deck for Thursday's board meeting. An AI persona plays the CEO — direct, time-pressed, asking tough questions. You clarify the task, do the work, and submit.

Then the CEO persona debriefs you: walking through your deck slide by slide, pushing on weak assumptions, showing you how a stronger version might look.

For different roles:

• SecOps: Simulated alerts with realistic stakeholder pressure
• Sales: Simulated prospects with real objections
• PM: Simulated tradeoff decisions or launch go/no-go calls

The coach then debriefs: here's where you followed the expert pattern; here's where you deviated; here's targeted practice to close the gap.

Meta-Coaching — "Teach Me How to Use AI in This Job"

Within clear privacy bounds, the coach can see patterns across your work:

• You never invoke AI on tasks where top performers always do
• You routinely over-delegate judgments that should be human-owned

It can then push back:

• "People in your role often use AI here to cut the time in half — want to try that flow?"
• "This is a high-risk decision. Here's why your policy says you should not rely solely on AI."

The Difference: Generic Assistant vs. Context-Aware Coach

This is the difference between a capable tool and a coach that's explicitly trying to upgrade you.

4. Measuring Impact (Without Pretending It's Trivial)

You won't get McKinsey-grade causal inference on day one. But you need more than vibes. The realistic approach is staged:

Phase 1: Perceived Impact + Engagement

• In-flow micro-surveys:
  • "Roughly how much faster can you do [workflow] now?" (bucketed)
  • "Confidence using AI on [task]?" (1–5)
• Engagement metrics:
  • Which workflows see repeat use?
  • Do people graduate from guided → feedback → independent use?

This is crude, but it tells you whether you're moving the needle.

Phase 2: Workflow-Level Metrics Where They Exist

For specific workflows, you can track:

• Time to complete (before/after)
• Rework / defect / incident rates
• Cycle time for standardized artifacts (PRDs, incident reports, campaigns)

Pick a few anchor metrics per pilot; don't try to instrument everything.

Phase 3: Sampling & Observer Mode (For Mature Customers)

With explicit consent:

• Sample real artifacts over time (PRs, reports, outreach sequences)
• Score them against your rubrics (AI + human spot checks)
• Correlate quality trends with coach usage

The goal isn't individual surveillance; it's giving leaders and teams a real sense of where practice is improving and where it's stuck.

5. Business Model and Buyer: Where Does This Actually Live?

This almost certainly starts as consulting + product:

Phase 1: Consultative Onboarding

• Pick 1–2 roles and 3–5 workflows each
• Do workflow capture with true experts
• Build the initial Learning Graph
• Configure the coach into the customer's tools

Phase 2: Subscription

• Access to the Learning Graph engine
• The AI coaching layer
• Ongoing refinement of workflows and rubrics
• Expansion to new roles and tools

Who Buys This?

[MODIFIED — tightened]
My hunch is the primary buyer is not central IT or a generic "AI office," but department leaders who care about productivity and quality:

• Head of Product / VP Engineering
• CISO / Head of Security Ops
• VP Sales / Marketing
• CFO / GC / functional leaders in Finance, Legal

These folks don't usually have an "AI productivity software" budget line item (yet), but they do have:

• Budget for tools and enablement
• Mandates to "do more with less"
• Frustration that their teams aren't getting full value from the AI they've already bought

[NEW — category framing]
Over time, I can imagine this sitting at the intersection of:

• AI enablement — how we train people to use AI
• Productivity tooling — how we help people work faster
• L&D / capability building — how we develop talent

There's probably a category name that will emerge. For now, "AI-powered coaching for knowledge worker productivity" is descriptive enough, even if it's not yet on a Gartner quadrant.

6. Longer-Term: From Coaching Workflows to Orchestrating Them

[MODIFIED — tightened, less redundant]
An interesting upside of this approach is that workflow understanding accumulates.

As we:

• Capture how people work
• See where they consistently partner with AI
• Understand patterns in tasks and handoffs

We're in a position to:

• Suggest semi-automated agentic workflows for repetitive, well-understood tasks
• Use the same Learning Graph as the spec for agents
• Still coach humans on oversight and the non-automatable decisions

[NEW — cleaner articulation]
In other words: the same engine that teaches humans can power safer automation. Once you've seen 1,000+ instances of a workflow and know where humans actually add value, you can automate the rest with confidence.

The Learning Graph becomes both the training curriculum and the automation blueprint.

To Anyone Still Reading This...

[MODIFIED — slightly tightened, stronger close]
If you zoom out, this is a simple thesis:

The people who win in the next decade aren't "AI people." They're domain experts who've rebuilt their craft around AI.

We already see the prototypes of those people in PM, engineering, security, sales, marketing. The problem is the playbook is trapped in their heads and Slack threads.

I want to:

• Capture those playbooks
• Turn them into Learning Graphs
• Use AI to coach the next 10,000 people in each role up the curve

This is obviously bigger than a weekend project. There are open questions I'd like to test early. I'd love to chat if any of this resonates.

[KEEP — these questions are good, work well as-is]

Specifically:

Problem reality:

• Do you see the "AI-skilled vs AI-unskilled" gap in your world?
• Is AI mostly licenses and talk, or is it changing real workflows?

Value / fit:

• If you lead or work in Product, Engineering, Security, Sales, Marketing:
  • Would a role-specific AI coach like this be compelling?
  • Where would you pilot it first — what workflow, what team?

Practical barriers:

• What would stop you from trying this?
  • Security and data privacy?
  • Change-management fatigue ("not another tool")?
  • Skepticism that AI can really "coach" at this level?

Buyer / budget reality:

• If you picture your own org:
  • Who would own this?
  • Which budget(s) would you tap?

Blind spots:

• Are there obvious pitfalls in:
  • Capturing workflows this way?
  • Relying on a Learning Graph abstraction?
  • Measuring impact through user-reported and sampled measures?

[NEW — stronger closing call to action]
If you've read this far, thank you. If anything here makes you think "this could be big if X" or "this will fail because Y," I'd genuinely love to hear it.

Reply with brutal honesty, quiet interest, or "you should talk to so-and-so." Even if this never becomes a company, the underlying question — how we help people actually work differently with AI, at scale — is one we're all going to have to answer.

[NEW — explicit CTAs]
→ Subscribe for updates as I build this
→ Reply if you want to be an early tester or contributor
→ Share if you know someone wrestling with this problem