AI in GTM School · Q2 2026
Class 1 of 8 · Recap & Takeaways

The State of AI in 2026

A working artifact for the cohort — what was taught, what was asked, what to do next. Built from the full audio and chat transcripts of the April 29 session.

Instructors: Andy Jolls & Jonathan Moss
Date: April 29, 2026
Cohort size: ~348 attendees
Length: 91 minutes

TL;DR

AI won't fix what's broken — it scales it. Start at data, not at the application. Move from tasks to systems. Most of us are still at the chatbot stage of the maturity ladder, and that's fine. The unlock for this cohort is to stop typing prompts and start talking to Cowork like an intern — describe the outcome, ask it to build the skills, schedule the workflow. Then iterate the skill, not the prompt.

Foundations

The 5 Big Ideas

Five principles Andy and Jonathan kept returning to. If you internalize only these, you'll outperform the cohort.

1

AI doesn't fix broken — it scales it

Bolting AI onto messy data, undefined process, or disconnected systems just makes the mess faster and louder. Sequence matters: data → process → systems → application.

"People start at the application layer. Start at the data layer and work your way up." — Jonathan Moss
2

Think in systems, not tasks

"Write me an email" is a task. "Watch my calendar, prep me for next week's calls, and surface signal from CRM activity" is a system. Tasks are great — but the system is what compounds.

3

The 30% coordination tax

Roughly 30¢ of every dollar gets eaten by searching, decision delay, missed signals, and threading people together. The biggest AI wins target coordination — not just individual productivity.

4

Most of us are at chatbot — that's the starting line

The 5-question maturity assessment showed the cohort clustered at Levels 1–2 (chatbots and simple automation). The class isn't behind. The point of these 8 weeks is to climb.

"Be authentic with where you are. Resist the urge to run if you're really crawling." — Andy Jolls
5

Talk, don't type

Both demos started with a voice-style description of the desired outcome — not a hand-crafted prompt. Cowork's skill creator turns that conversation into reusable skills, scheduled tasks, and connectors automatically.

The Tool Landscape

Who's winning where, as of April 2026

The cohort's tool mix has shifted dramatically in one quarter. In January the class skewed ChatGPT; today it skews Claude. Use the right tool for the job — most attendees run two or more.

Tool Where it wins Where it struggles Cohort signal
Claude (Chat / Cowork / Code) Enterprise, deep work, agentic workflows, building skills Capacity limits hit by noon for heavy users; ARM laptop gaps in Cowork Dominant — ~80% of class
ChatGPT Consumer reach, breadth of integrations, polished output Less flexible for skill-based agentic patterns vs Claude Still strong globally
Microsoft Copilot "Approved at work" — already wired into M365 stack Less expressive; many here feel "forced" to use it The default many didn't choose
Gemini Visuals, Google Workspace deep integration, "Gems" Lagging on agentic UX vs Claude/ChatGPT Up and coming
Perplexity Quick research, multi-model access in one UI Less suited for building durable workflows Personal/research go-to
Specialty (Lovable, Replit, Wispr Flow, NotebookLM, Manus, Clay) Purpose-built lifts: portfolio sites, voice capture, doc synthesis, vibe-coding, prospecting Stack sprawl & cost adds up Power users layer them in
It's a one-tool world only if you let it be. Most instructors run multiple. Pick the right tool for the complexity of the job. — Synthesized from the tool debate thread, ~27:00 mark

Hands-on Recap

Two demos, one pattern

Both demos showed the same loop: describe the outcome conversationally → let Cowork create the skills → schedule the workflow → iterate the skill, not the prompt.

Andy Jolls

Monthly Google Trends scheduled task

Built so the April 1, May 1, June 1 reports run themselves — instead of Andy doing it manually each quarter.

  • Started in Cowork, not the schedule UI — the UI hid the "monthly" cadence option.
  • Plain-English prompt: "I want to check Google Trends every month on how ChatGPT, Gemini, Claude and Perplexity are doing. How can you help with this?"
  • Pointed it at his existing slides as a style example.
  • Layered in news-monitoring schedules for Anthropic, OpenAI, Gemini changelogs — silent on no-news days.
Lesson: If the UI fights you, talk to the agent instead. It writes the schedule for you.
Jonathan Moss

Monthly Competitive Comparison workflow

Built end-to-end in one prompt: research → PowerPoint deck → interactive HTML dashboard, on the 1st of every month.

  • Folder of context: brand guidelines, pricing, value props, .pptx template.
  • One prompt asked Cowork to create three skills: competitive research, PowerPoint design, UX/UI dashboard.
  • Cowork asked clarifying questions back: audience? format? baseline data?
  • Same prompt produced a working version in ChatGPT — skills are portable markdown.
Lesson: Don't pre-build skills in isolation. Build them in the flow of work, the way you'd brief a teammate with multiple capabilities.

Skills 101

Global · Project · Skill · Prompt

A common point of confusion. Here's the mental model the instructors landed on.

Global instructions

Always-on rules. Tone, voice, output formats you want everywhere. Set once.

Projects / folders

Scoped context for one body of work. Files, references, reusable assets. Folders = local + flexible. Projects = cloud + tighter scope.

Skills

Specialized "interns." Domain expertise + an output spec. Invoked when the situation calls for them. Iterate the skill, not the prompt.

Prompts

The actual ask in the moment. Should mostly say what and why — the skill handles the how.

"A skill is like a cheat sheet you tape to my desk. I check it automatically whenever the situation comes up, so you never have to repeat yourself again." — Karen C., asking Claude to explain skills like she's 5
"If you do the same thing more than twice, opportunity to create a skill." — Scott E.

Where you are on the curve

The AI Maturity Ladder

From the live assessment, ~106 respondents clustered at Level 1. The 30-question version goes deeper — take it during the week.

Level 1
Chatbot
Ask, copy, paste. One-off conversations, no memory.
Level 2
Power user
Projects, files, custom instructions. Reusing context.
Level 3
Skills + scheduling
Codified expertise. Tasks run on cadence. This week's target.
Level 4
Multi-agent
Skills coordinate. Workflows span tools and connectors.
Level 5
Agentic ops
Agents own outcomes, plug into core systems, accountable to a team.

This week

Quick Wins to Try Before Class 2

Pick three. Resist the urge to do all ten.

Take the 30-question maturity assessment.Get a real benchmark. Save your score for class 8 to measure your move.
Schedule one recurring task in Cowork.Use Jonathan's competitive-comparison prompt as the template — swap in your competitors.
Stand up a clean folder structure.Steal Scott S.'s pattern: CLAUDE.md, _reference/, _templates/, then per-project folders with their own CLAUDE.md + HANDOVER.md.
Schedule heavy tasks off-peak.Avoid 8 AM – 2 PM ET on Cowork to save tokens. (Tip via Victoria A..)
Install Tom C.'s "reduce hallucinations" skill.Posted in the class Slack — tiered by task type. Less paranoia about made-up answers.
Ask Claude to roast your output."Now find the holes in this." Tiffany H.'s tip — surfaces gaps you'd miss on your own.
Build a passion-project workflow.Especially if you're between roles or blocked at work — synthetic data is a feature, not a workaround.
Write your top 5 jobs-to-be-done in a 2x2.Highest impact × lowest lift = your first 3 skills. If you can't pick, ask the LLM.
Refresh the Pavilion AI Pulse Report.15 minutes. Useful context for every other class.
Send your Class 5 skill request to Scott.Scott Wueschinski is curating a production-ready GitHub repo for Class 5. Submit the skill you want in the package via the skill-request form.

Things we didn't fully solve

Open questions surfaced by the cohort

These came up in chat or on mic and didn't get a clean answer. Future classes — and the Slack — will keep working them.

  • Who owns the agent? Built on personal time, on personal hardware, with company data — IP, liability, transferability. Kelly B. framed it as the missing infrastructure layer (no Active Directory for agents, no DNS, no SWIFT). 2–4 year horizon before this is settled.
  • Copilot-only orgs. Real tension between Pavilion's Claude-leaning curriculum and many members' enterprise stack. Workaround for now: synthetic data + personal Claude account; bring outputs back into Copilot for company use.
  • Cross-machine sync. Skills are tied to the local machine. Cowork doesn't yet ARM-natively support all laptops. Workaround: GitHub as the source of truth + sync scripts (Scott Wueschinski's pattern).
  • Privacy / PII / PHI. Sandesh raised the cybersecurity-AI angle. Walled-garden infrastructure (AWS, Azure, GCP) + a cross-functional Center of Excellence is the durable answer. Shadow AI is the bigger risk than sanctioned AI.
  • Cost discipline. Heavy users hit limits before noon. Tactics shared: caveman-speak prompting, Haiku for cheap routes, off-peak scheduling, semantic routers. Norm-setting: $20–$200/mo of personal investment is reasonable for a GTM professional in 2026.
  • External-facing output. Most of the cohort uses AI for inbound prep / internal research. Building agents that produce outbound, customer-facing output is where Class 5 (Scott's session) will go deeper.

Resources Dropped In Class

One place for every link

Pavilion
Pavilion AI Pulse Report — April 2026

Cohort survey: tool mix, deployment patterns, blockers.

Andy Jolls
AI Maturity Assessment

5-question and 30-question versions. Take the long one.

Andy Jolls
Class 1 Slides

The deck used in class.

Andy Jolls
Synthetic CRM Dataset

10K rows, 30+ columns. Use it if you're between roles or blocked from real data.

Jonathan Moss
Chat vs Cowork vs Code

Reference site explaining when to use which Claude surface.

Scott Wueschinski
LLMRouter (GitHub)

Open-source router for cost-aware multi-model orchestration.

Scott Wueschinski
vLLM Semantic Router

Routes prompts to the right model by semantic class.

Pavilion
School syllabus

All 8 weeks at a glance. Recordings live in the Pavilion Member Hub.

Coming in Class 5: Scott's production-ready repo

Scott Wueschinski (GTMify) is curating a deployable GitHub repo of Cowork + Claude Code skills built for GTM teams. He's taking skill requests from the cohort right now — send yours in before Class 4 to get it in the package.

Drop your skill request

From the chat — quotes worth keeping

Voices of the cohort

"I thought I was pretty far along on the journey until I got to this class." — Ryan P. (a feeling shared widely in chat)
"Lesson learned: put your 'product manager' hat on." — Scott M.
"We have agents that can act. What's missing is the infrastructure layer that makes them accountable. No Active Directory for agents inside organizations, no DNS or BGP across the internet, and no SWIFT between organizations. So we cannot answer a simple question: who owns the agent and who is liable when it acts?" — Kelly B.
"Move from AI whack-a-mole to a much more structured design approach to implementing — to get outcomes." — Andy Jolls, closing