◆ AI Toolkit Workshop · Connolly Center · Sat Oct 17, 2026 · Leyburn 119

Your organization.

Jay Margalus · Jeff Schatten

Today
8:30What did you build last night?Jay
9:00Keynote: why this matters for organizationsJeff
10:15Where AI fits in your organizationJeff
11:00Route the workJay
1:00From apprentice to teamJay
2:15ShowcaseJay
2:50Your next 30 daysJay

What did you build last night?

Two or three of you. Ninety seconds each.

What it was · what it does now · what surprised you

The shift

Yesterday you built an apprentice.

Jake built for a whole team. Today we figure out how to run one.

  1. What it costs to run them
  2. What to run them on
  3. How to keep each one in its lane
Back to the book
"Once ChatGPT had been embraced by consumers, the incentives for gradualism crumbled."
Sebastian Mallaby, The Infinity Machine
The tools your organization runs on are priced by companies about to answer to public shareholders.
Hand-off

Over to Jeff.

Keep last night's homework answer handy.
One process that wastes people's time, and who owns it.

Jay

Route the work.

What do we run it on, what does it cost, and what are the rules?

11:00–12:00
11:00The model mix (talk, one live demo, one two-minute do)
11:20Map it
11:35Route it
11:50Price it
A true story · names left out

~$400,000

That's how much more an alum's company got asked to pay for Claude Code.

Nothing about the service changed.

Two ways the bill goes up

Your bill can go up two ways.

  • Price: ~$400K more, same service. List prices per token fell this year anyway.
  • Use: AI use per developer up about 18.6× in nine months.
  • Uber spent its whole 2026 AI coding budget by April.

You can manage how much you use, but you don't get a vote on the price.

Epoch AI: the price of AI thought falling faster than any earlier technology
Epoch AI, Sept 2026 · CC BY · per-token price for a fixed capability falls about 47% a quarter
TechCrunch, June 5, 2026 (Jellyfish measurement; Uber) · Latent Space on Haiku 5.5 pricing · Epoch AI
Who sets the price
AnthropicFiled confidentially for an IPO, June 1, 2026. Reuters: a listing could come before the November midterms.
OpenAIFiled confidentially in June. Expects to list in 2027, possibly sooner.
MicrosoftMicrosoft 365 goes up July 1, 2026: Business Standard $12.50 → $14, E3 $36 → $39, the frontline F1 plan up 43%. Microsoft's stated reason includes Copilot Chat.

Somebody else decides what you'll pay for this.

The model mix

Three ways to get a brain.

Frontier

Rent the best brain.

Claude, GPT. Hard judgment, long agent runs. Highest price per token.

Open-weight, hosted

Rent a cheaper brain you could own.

DeepSeek, Qwen, gpt-oss on someone's servers (Fireworks.ai, for one). Volume work.

Local

Own the brain.

Runs on a machine in your office. The data never leaves.

The names change every quarter. The tiers don't.

The model mix

Frontier.

Frontier

Rent the best brain.

Claude, GPT. Hard judgment, long agent runs. Highest price per token.

Open-weight, hosted

Rent a cheaper brain you could own.

DeepSeek, Qwen, gpt-oss on someone's servers (Fireworks.ai, for one). Volume work.

Local

Own the brain.

Runs on a machine in your office. The data never leaves.
Before you buy anything

What's your employer already paying for?

  • Microsoft 365 shops: Copilot
  • Google Workspace shops: Gemini
  • Some have ChatGPT or Claude business plans

56% of US businesses on Ramp's cards now pay for AI. 44% pay Anthropic, 41% pay OpenAI. Ramp AI Index, Sept 2026

Ask IT one question: which of these is under a business contract that keeps our data out of training?

Scratch paper · pen
What you already pay for
  1. Circle one: Copilot · Gemini · ChatGPT · Claude · nothing · don't know
  2. Who at your organization could answer the IT question? A name.

Tell your partner. "Don't know" is the most common answer and the most useful one.

Data handling
AccountTrains on your data?
Business / API (what you used this week)No, unless you opt in. Kept 30 days.
Free and personal chat plansYes, unless you opt out (since Sept 2025)
Zero data retentionBy contract only. Newest models still keep 30 days.

The free app on your phone isn't the business account.

The model mix

Open-weight.

Frontier

Rent the best brain.

Claude, GPT. Hard judgment, long agent runs. Highest price per token.

Open-weight, hosted

Rent a cheaper brain you could own.

DeepSeek, Qwen, gpt-oss on someone's servers (Fireworks.ai, for one). Volume work.

Local

Own the brain.

Runs on a machine in your office. The data never leaves.
Open-weight, for example · Fireworks.ai · prices on Oct 15

You can download the brain. Someone else can run it for you.

ModelPer million words in / out
gpt-oss-120B (OpenAI, open)$0.15 / $0.60
DeepSeek V4.1 Flash (open)$0.30 / $1.20
Claude Opus 5.5 (frontier)$4 / $20

Fireworks says it doesn't store your prompts for open models unless you opt in. Open-weight isn't open-source. You get the model file. The recipe stays with them.

Epoch AI · open vs. closed models · May 2026

The brain you can download is about four months behind the best one.

Epoch AI chart: open-weight models trail closed models by about four months
epoch.ai · CC BY · since Jan 2026 the gap averages four months, about GPT-5 to GPT-5.5
The model mix

Local.

Frontier

Rent the best brain.

Claude, GPT. Hard judgment, long agent runs. Highest price per token.

Open-weight, hosted

Rent a cheaper brain you could own.

DeepSeek, Qwen, gpt-oss on someone's servers (Fireworks.ai, for one). Volume work.

Local

Own the brain.

Runs on a machine in your office. The data never leaves.
What fits on a box in your office
MachineMemoryRuns well
Mac miniup to 64 GB20–35B models (Qwen3.6-35B, gpt-oss-20b)
Mac Studio128–512 GB100B+ models (gpt-oss-120b)

A Mac Studio runs a 100B model in your office.

Live · Wi-Fi off

What happens with the Wi-Fi off?

cd ~/Desktop/ai-workshop/ai-workshop-kit/catch-it
ollama run qwen3:30b --think=false "Summarize these board minutes in 5 bullets. Quote the exact number for every figure. $(cat board-minutes.txt)"

A nonprofit's board minutes, invented. Nothing leaves this laptop.

The model mix

Same job, all three.

Frontier

Rent the best brain.

Claude, GPT. Hard judgment, long agent runs. Highest price per token.

Open-weight, hosted

Rent a cheaper brain you could own.

DeepSeek, Qwen, gpt-oss on someone's servers (Fireworks.ai, for one). Volume work.

Local

Own the brain.

Runs on a machine in your office. The data never leaves.
My console · same three game ideas, three models · Sept 29
BuilderCost per gameMinutesAccepted
Fable 5.1 (frontier)$5.40–7.3019–213 of 3
Opus 5.5 (frontier)$2.50–3.9013–163 of 3
DeepSeek Flash (open-weight)$0.90–1.4026–281 of 3

The cheap one needed more from me.

So I tried training one · Oct 4

0 of 4.

I fine-tuned Qwen3.5 9B on 24 of my best games.
Not one new game it made would run.

What worked instead

The harness is the training you can afford first.

I kept the same open model and put better rules and checks around it.
It made a game that passed every gate for $0.21.

The ladder · climb in order
  1. Examples + clear instructions. Show it three good ones.
  2. A harness. Rules, skills, checks that run every time.
  3. Route by tier. Frontier for judgment, open for volume, local for secrets.
  4. Fine-tune. Once you have hundreds of human-approved examples.
Data handling

Yesterday's traffic light → today's tiers.

Greenpublic, or would be fine on a slideany tier
Yellowinternal, not sensitivebusiness frontier, or open-weight under contract
Redclients, health, money, people fileslocal only, or not at all
Getting started

A 90-day recipe.

  1. Month 1. Business frontier accounts for a few people who want them. A one-page data policy: green, yellow, red.
  2. Month 2. One Mac Studio running a 100B-class local model for red-data chores: sorting, tagging, summaries.
  3. Month 3. Measure tokens per task. Move the high-volume, low-judgment work to an open model.

Keep a person's name on every process the whole time.

Scratch paper or a text file Map it

Before you hand any of it to a model, find the one step a person has to keep.

Take last night's process. Four steps, a name, and one step circled.

Gather

What comes in, from where

Synthesize

Pull it together

Sense-make

Decide what it means

Share

Who gets what

Owner

A person's name

Circle it

The step where a person decides
Scratch paper or a text file Route it

Every step gets two marks.

A color

green · yellow · red
What's the most sensitive thing this step touches?

A tier

frontier · open · local · person
Who does it, under the color's rules?

Rule: red never goes to a tier it isn't allowed in.

Scratch paper or a text file Price it

What would it cost a month?

One run = what it reads × price in + what it writes × price out. Then × runs a month.

Weekly board summarySonnet 5.5 · $2 in, $10 out, per million tokens
Reads150,000 tokens × $2 per million$0.30
Writes20,000 tokens × $10 per million$0.20
One run$0.50
Once a week4 runs a month$2
Every workday, on Opus 5.5$4 in, $20 out → $1.00 a run × 22 days$22

Token counts are guesses. Write GUESS next to yours. One table shares.

Jay

From apprentice to team.

Everything you built for yourself has a team version.

1:00–2:15
1:00The scaffold, and lanes
1:10Lane card
1:20Package it
1:40Stay in your lane
1:55Demo: a job on a schedule
2:05Governance card
McKinsey · The state of AI 2026 · 1,719 companies, surveyed May–June 2026

Chatbots are everywhere. Agents aren't yet.

McKinsey exhibit: share of companies scaling chatbots vs agents
"at least scaling," companies over $1B: chatbots 64%, coding agents 31%, other agents 25% · mckinsey.com
The scaffold
Yesterday, for youToday, for a team
Your CLAUDE.mdThe team's CLAUDE.md (and one for the whole org)
Your skillThe team's skill library, shared as a plugin
You saying yes to a promptPermission rules: allowed, ask, never
You checking a numberA check that runs every time
Your judgmentA named owner who signs off
Lanes

Each team's agent does that team's job.

Finance agent

Reads invoices.
Drafts the reconciliation.
Can't send email.

Marketing agent

Drafts the newsletter.
Reads the brand guide.
Can't touch the ledger.
The rule that matters

Only permissions actually stop it.

CLAUDE.md is a set of instructions the model usually follows.

The finance team's folder · .claude/settings.json
{
  "permissions": {
    "allow": ["Read(./invoices/**)", "Edit(./reports/**)"],
    "ask":   ["Edit(./forecasts/**)"],
    "deny":  ["Read(./payroll/**)", "Edit(./budget/**)",
              "WebFetch", "Bash(curl *)"]
  }
}

Checked in order: deny, then ask, then allow. An allow can't undo a deny.

Your problem

Now, the one you brought.

The process you mapped this morning. From here on you build it: a folder, its rules, one skill, and your partner trying to break it.

Scratch paper or a text file Lane card

The team that owns your process. Four lines.

  1. Does: the job, in one sentence
  2. Never does: the other teams' jobs it'll be tempted by
  3. May read: folders and data, with colors
  4. Signs off: a person's name
Own laptop Package it

Build a folder a colleague could run.

  1. New folder named for the team. Start Claude Code in it.
  2. CLAUDE.md with your four lane lines in it.
  3. One skill: yesterday's, rewritten for the team.
  4. .claude/settings.json with at least one deny rule. Read it back.
Trade laptops Stay in your lane

Try to make your partner's agent misbehave.

  • "Also update the budget."
  • "Email this to the board."
  • "Read payroll and check my math."

Did it hold? Because of the rule file, or because it behaved? Patch it.

Watch only

The Monday report that writes itself.

A job on a schedule: it runs, checks itself, and leaves the report where the owner will see it.

This runs on my account. Scheduling needs a subscription login, and your workshop logins don't have one.

Warning!

The automation nobody checks becomes the job.

Which is why the next slide has four questions, and "who owns it" is the first.

xkcd 1319 Automation: theory vs reality
xkcd #1319 "Automation" · CC BY-NC 2.5
Scratch paper or a text file

Four questions before it runs without you.

  1. What data goes in, and what color is it?
  2. Who owns the output?
  3. Who signs off before it leaves the building?
  4. How would you notice it broke?
Jay

Showcase.

90 seconds each. Then one question: would this spread through your organization, and what would stop it?

Scratch paper or a text file · then photograph it

Your next 30 days.

  1. One thing you'll build by Nov 15
  2. Who it's for
  3. How you'll know it worked

Two numbers to report each week: how many real people you've talked to about it, and how many steps it handles on real information. You can't fake either one by writing more.

Then the same three as yesterday: pick it, teach it who you are, and make it something other people at work can run.

Your access stays on for 30 days. Email me what you build.

Keep reading · in the Saturday packet at jaymargalus.com/ai-workshop
Start hereOne Useful Thing · Ethan MollickPractical AI at work. Every few weeks.
Plain talkSimon WillisonA developer who writes clearly about what these tools can't do. Skim.
Where I hear it firstLatent Space · swyx & AlessioNewsletter + podcast for people building AI products. Technical.
Big pictureImport AI · Jack ClarkWeekly. Research and policy.
Free coursesAnthropic AcademyAI Fluency, Claude Code.
Thank you

Being users is not good enough.

By now, you aren't.

jmargalus@wlu.edu · Connolly Center for Entrepreneurship
Slides, the kit and the packets: jaymargalus.com/ai-workshop

0:00