Who is the client in this report?

For confidentiality we use a pseudonymous company with an unspecified industry. It reconstructs, as a single day, the events that recur across multiple real Day 1 sessions. Numbers and time blocks reflect actual training.

Can non-engineers really finish?

The five attendees that day were mostly finance, PR, sales and admin with no programming background, yet all reached the afternoon demos. Only setup needs individual help, so one instructor can handle about 10 people.

What can attendees do after one day?

They leave with a working template that delegates one routine task to AI (e.g., triaging unread email into a spreadsheet). Day 2 onward puts it into real workflows.

Can you run this for our company?

Yes. We adapt the demos and curriculum to your industry and roles. For subsidies and class size, reach out via the white-paper form at the end of this article.

Field Report: A Day Inside a Corporate AI Agent Training

Published: 2026-05-27 · Author: AI Agent Camp (instructor: Kohei Nakamura) ・ Reading time: about 10 minutes

We shadowed Day 1 as five mostly non-engineering employees took a full day of AI agent training — check-in, the setup struggle, lecture, demos and wrap-up, in order.

One morning, five employees walked into a Tokyo meeting room with their laptops — mostly non-engineering roles: finance, PR, sales, admin. Everyone had used ChatGPT, but none had touched Claude Code or Cursor. Today is Day 1 of training to see whether these five can turn an "AI agent" into a tool for their own work. We shadowed the day from check-in to wrap-up.

Attendees with laptops in a meeting room while an instructor explains at the front

09:30Check-in — starting from skepticism

The pre-class chatter was honest: "I'm not sure what AI can actually do" and "no time to try it." The instructor immediately lowered the bar: "Today's goal is to leave having handed exactly one of your tasks to AI."

10:00The setup struggle — plan to lose the whole morning

This is the hardest part of Day 1. Confirming VS Code / Cursor, Google login for Claude Code, GitHub auth, cloning the repo, gog CLI auth — many steps. Within 20 minutes several people were stuck. The order is always the same: Claude Code UI version drift, corporate policy blocking extensions, then the "unverified app" OAuth warning. The instructor walks the room, checking each screen.

A code editor and terminal running a setup script (illustrative mockup)

During gog CLI auth, a terminal command opens a browser asking for approval. The most common question — "is it safe to continue here?" — is the perfect cue to explain least privilege, which later connects to the security segment. The first "wow" came when an inbox message was fetched successfully.

CLI authentication in a terminal handing off to a browser consent screen (illustrative)

12:30Lunch — really a catch-up session

The full lunch hour doubles as catch-up time. Fixing anyone who got stuck in the morning, over a bento, visibly raises the afternoon completion rate.

13:30Lecture — just four ideas, in plain words

Sixty minutes of lecture, held to four ideas: (1) an LLM is a text-to-text probability model, (2) context is how you supply surrounding information, (3) one chat, one task, (4) an AI agent = LLM + tools + loop. Framing hallucination as a training bias ("guessing beats leaving blank on a multiple-choice test") landed with the finance lead.

14:30Hands-on — demos mapped to their own work

The middle 90 minutes are demos and imitation. The first: unreplied Gmail → priority scoring → spreadsheet. The same prompt yields different results per person — the probabilistic world, felt firsthand. The sales rep stopped to take notes: "I could use this for the morning outbound summary."

Triaging unread inbox into a spreadsheet automatically (illustrative)

Then calendar free-slot search → event creation, and browser automation via a Chrome extension. Deliberately showing a "forgot to stop it" loop motivates human-in-the-loop. Finally, running over a dozen parallel Claude Code sessions made the room buzz — the clearest way to convey "zero wait time."

Many sessions tiled on a desktop running different tasks in parallel (illustrative)

16:30Security — don't get used to loosening guardrails

The late afternoon is defense: don't paste API keys into Slack or Notion; email automation stops at draft; never auto-accept meetings; don't let the model alone decide on legal, medical or sensitive data. Showing a prompt-injection example changed the PR lead's face from morning skepticism to "that's scary." Delivering risk with the same energy as the upside is what works in corporate training.

17:30Wrap-up — everyone leaves with three "homework" items

Each attendee writes three repetitive tasks they want to delegate. Finance: turning meeting notes into purchase tasks. PR: three-axis press-release review. Sales: morning outbound summary. Day 2 starts by putting these into real workflows. Morning skepticism had turned into "I want to try this on Monday."

An attendee writing AI homework items on a whiteboard (illustrative)

Takeaways

You can't make engineers in a day. But "hand one task to AI" is reachable in a day, even for mostly non-engineers. The keys: plan the morning setup to overrun, hold the lecture to four ideas, and deliver risk with the same energy as the benefits.

For the detailed agenda and demo steps, see "Corporate AI Agent Training in Practice — What Actually Happened on Day 1".

Talk to us about corporate training

We design corporate AI agent training with Claude Code, Cursor and gog CLI for your industry and roles. Ask us about class size, subsidies and more — the short form below takes about 30 seconds.

Want to run this for your company?

We adapt the demos and curriculum to your industry and roles. Single-company programs, subsidy-eligible options, class sizes 4–20 — reach out anytime.

Learn more about AI Agent Camp

FAQs

Who is the client in this report?: For confidentiality we use a pseudonymous company with an unspecified industry. It reconstructs, as a single day, the events that recur across multiple real Day 1 sessions. Numbers and time blocks reflect actual training.
Can non-engineers really finish?: The five attendees that day were mostly finance, PR, sales and admin with no programming background, yet all reached the afternoon demos. Only setup needs individual help, so one instructor can handle about 10 people.
What can attendees do after one day?: They leave with a working template that delegates one routine task to AI (e.g., triaging unread email into a spreadsheet). Day 2 onward puts it into real workflows.
Can you run this for our company?: Yes. We adapt the demos and curriculum to your industry and roles. For subsidies and class size, reach out via the white-paper form at the end of this article.