Where to start

Approved tools and getting started

Requesting a license

I want Claude access

Confirm Copilot doesn't already cover it, then get the exact next step.

New to this

I need the vocabulary first

What an LLM, an agent, a skill, and a connector actually mean, briefly.

Have access already

I want to use it well

What data rules apply, how to get a good result, and what to verify.

Going deeper

I want to go further with it

Working sessions, department by department, and where the record lives.

Approved today

Microsoft 365 Copilot, Claude (chat, Projects, Cowork, and Code), and a small set of approved connectors into company systems. Anything else isn't cleared for company or customer data yet.

Before you request a license

Do you need Claude, or does Copilot already cover this?

Microsoft 365 Copilot

Quick drafting in Word and Outlook, meeting summaries, spreadsheet analysis, searching across your files and email. Already included with your Microsoft license.

Try this first

Claude

Reasoning through a long or complex document, comparing several sources, sustained back-and-forth analysis, writing or reviewing code.

Requires a license

A simple way to tell

If the task lives inside Outlook, Word, Excel, Teams, or SharePoint and mostly needs drafting, summarizing, or search, Copilot alone likely covers it. If it needs deep analysis, long-context reasoning, or code, that's Claude territory.

First, the policy

Employee Experience is sending everyone Hayden's AI Acceptable Use Policy directly. Confirm where you stand before requesting access.

I've received or will sign the AI Acceptable Use Policy from Employee Experience.

Now, tell your manager

Access runs through a short business case. Email your manager and include:

What you'd use it for
How often: daily, weekly, or occasional
Roughly how much time it takes today without it

That's it

Once your manager approves, IT will follow up to get you set up.

Shared vocabulary

The words worth knowing first

Terminology matters more early on than mastering any one feature. This is the short version.

Computer systems built to do things that normally take human thinking: understanding language, recognizing patterns, making a recommendation, or creating something new.

LLM (large language model)

The specific kind of AI behind Claude and Copilot. It's trained on huge amounts of text to predict and generate language, which is why it's good at writing, summarizing, and reasoning through problems.

Agent

An LLM equipped with tools and a goal, working in a loop: it takes an action, checks what happened, and decides the next step on its own, continuing until the task is done instead of stopping after one answer. Claude Cowork is Anthropic's product built around this, point it at a task and it works through the steps itself.

Skill

A saved, repeatable procedure for a specific task, so the same request gets handled the same reliable way every time instead of starting from scratch.

Project

A saved workspace that holds files, instructions, and conversations for one ongoing piece of work, a client, a report, a recurring process, so the background is already there every time instead of restated.

Artifact

A document, piece of code, or visual Claude creates as its own separate, viewable object in the conversation, something you can open, edit, and reuse directly.

Project

the ongoing workspace

Skill

the reusable recipe

Artifact

the one-time output

Context

Everything the model can actually see at once when it responds: the conversation so far, any documents you've shared, any instructions given. Quality matters here far more than quantity.

Connector

A secure, permissioned link between Claude and an approved company system, using your existing access. It only does what it's built for, a document connector searches and reads files, nothing more.

Prompt

The instruction you give the model. A specific prompt with real context produces a noticeably better result than a vague one.

Hallucination

A confident, well-written answer that is simply wrong. Every LLM does this, including Claude, which is why verification matters more than the tool you're using.

Model

The actual AI system doing the work. Claude and Copilot run on different models, and Claude itself comes in a few sizes, Haiku is fastest, Sonnet is the everyday default, Opus handles harder problems. Bigger and slower isn't automatically better, it depends on the task.

Token

The small chunks of text a model reads and writes, roughly three-quarters of a word each. Length limits and usage are usually measured in tokens, not words or pages.

Reasoning

On harder questions, Claude can work through a problem in visible steps before giving a final answer, instead of answering immediately. Slower, but more reliable on genuinely difficult problems.

Multimodal

The ability to work with more than typed text: images, screenshots, PDFs, spreadsheets. Both Claude and Copilot can read a photo of a whiteboard or a scanned document directly.

Custom instructions

Standing preferences you set once, tone, format, background about your role, that Claude then applies automatically to new conversations without you repeating them.

MCP

Short for Model Context Protocol, the technical standard connectors are built on. You don't need to understand how it works, just that it's the plumbing behind every connector.

Using it well

What data rules apply, and how to get a good result

Applies the same way whether you're in land, finance, plans, estimating, purchasing, sales, or customer service.

Claude can assist with work. It cannot own the work.

Managing context: the habit that matters most

Context is everything Claude can see when it answers: your message, the conversation so far, anything you've shared. A cluttered chat buries what matters, so this habit affects answer quality more than which model you use or how you phrase things.

A long, drifting chat

Old messages and pasted material crowd out your actual question.

A fresh, focused chat

Just what's needed, so your question gets full attention.

Start a new chat for a new topic. Don't keep adding onto one long conversation once you've moved on.
Share only what's needed. If you only need one section of a long document, paste that section, or ask for a summary first.
Put the important part first or last in whatever you paste, not buried in the middle.
Use a Project for anything recurring (a client, a report, an ongoing task), so the background is already loaded every time instead of re-explaining it.

Worth knowing

Claude automatically summarizes older messages once a chat gets long, you don't have to manage that by hand. Starting fresh is still simpler and more reliable, and that habit holds up even as the mechanics change.

What can and cannot go into an AI tool

Public

Published marketing copy, public pricing. Safe with any approved tool.

Internal

Meeting notes, non-sensitive reports. Safe with approved enterprise tools.

Confidential

Business strategy, financial plans. Needs case-by-case approval.

Restricted

SSNs, bank details, passwords, API keys. Never goes into any AI tool.

A simple rule of thumb

Enterprise Claude doesn't train on our data, but "enterprise" isn't the same as "cleared for anything." If you wouldn't casually email it or post it publicly, don't paste it into any AI tool.

You still own the final answer

Claude can draft. Claude cannot approve.

Verify before you act on or send out anything involving financial numbers, contracts or compliance claims, estimates or technical specifications, vendor pricing, customer-facing commitments, or anything you're reporting up as fact. Route it through your manager or a power user in that area first.

Getting a good result

Give it context

Paste in the document, data, or background it actually needs.

Describe the goal

Explain the underlying problem you're solving.

Format, if you have one

A table, a summary, bullet points. Nice to have, not required.

Good prompts versus bad prompts

Too openIs this okay to move forward with?

BetterSummarize the key points in this document. Quote anything important. Flag anything that seems unusual or unclear so I can look closer.

Too openWhich of these is the better option?

BetterCompare these two options side by side. What's different between them, and what's missing from either one?

Personal use is fine

Using Claude for things outside work is genuinely fine, as long as it doesn't crowd out your work usage. Hayden's policy supports this.

Questions or concerns

Something feels off, or you're not sure about a use case, talk to your manager, the Technology team, or Employee Experience.

Going further

Working sessions, proposed structure

Each department gets recurring time to look at what's useful for their own work. Here's the proposed shape of it.

What a session covers

Rather than one long training, each session looks at one or two of these applied to something real in that department's work:

Skills & Projects

Turning a recurring task into something reusable instead of rebuilding it each time.

Connectors

What company systems are already reachable, and what a good question to ask them looks like.

Web search & memory

When Claude should look something up versus rely on what's already been shared with it.

Who this is for

Whole department

Everyone gets exposure to what's relevant to their day-to-day, in plain terms.

Champions (1–2 per department)

Go a level deeper, following the three-step track below.

Learn

Build real skill with the tools, well past the basics.

→

Apply

Use it on one real workflow in your own department.

→

Show the team what worked, and coach the next person.

Topics to pull from

A working list, sorted by how much depth the topic needs. Everyday items fit into any AI Fundamentals session. The rest is where a champion track earns its time. Some of this will shift as Claude changes, so plan to revisit it.

Everyday, fits in AI Fundamentals

Custom instructions Checking your usage What an Artifact is Claude in Excel Claude in Chrome Editing a prompt in chat To-do lists in a chat Which connectors are approved Memory: what Claude remembers between chats, and how to manage it Using a Skill someone else shared with you

Power user depth, the champion track

Building and editing your own Skill Connectors vs. local MCPs, and what a connector can actually do Slash commands (Claude Code) Model choice and reasoning depth (Haiku, Sonnet, Opus) Chat vs. Cowork vs. Code, when to use which Live vs. static Artifacts Permission modes for Claude Code: ask, accept edits, plan, auto, bypass Background and scheduled tasks Forking a conversation (Claude Code) Claude Design, for decks and quick visuals Plugins Research mode, for multi-source reports Projects at depth: files, memory, and instructions together

What a session produces

Every session, whatever the topic, leaves the department with the same five things:

A few approved use cases

A few uses to avoid or route for review

A few reusable prompts worth saving

One workflow worth turning into a Project or Skill

One named department champion

What persists

The record is the sessions themselves: each of the five things above, plus a short summary, logged right after the session, no separate document to maintain. See .

Where things actually stand

The last 30 days, in numbers

Based on a 30-day usage review. Directional where noted, real where numbers are given.

91.7%

weekly seat utilization (industry average is 40–60%)

16,034

lines of code accepted from Claude Code in one month, 100% suggestion accept rate

Artifacts created company-wide this month, likely just an awareness gap

Three adoption patterns

Mature: chat + connectors, broad and habitual Emerging: Cowork, small but consistent High-output: Claude Code, concentrated and provable ROI

Raw chat volume dropped 60–75% over the period, because people moved from quick questions into deeper workflows like Cowork and connected Microsoft 365 use.

Where the leverage is

Projects usage is concentrated. A handful of employees account for most project usage, and only about 4% of the org has ever created one.

Next: a short internal demo from a current user.

Claude Code has a real ROI story. Two engineers produced meaningful accepted code volume this month.

Next: a show-and-tell for the rest of engineering.

Artifacts usage is zero. Likely an awareness gap, one good example tends to unlock the rest.

Next: share two or three internal examples.

Cowork hasn't spread past early adopters. Most of the remaining value sits outside engineering.

Next: pilot Cowork with one non-engineering team.

Directional capacity estimate

Based on industry benchmarks. Worth confirming with a short internal survey.

Developer capacity (Claude Code)

Roughly 960–1,560 hours a year, about 0.5–0.8 FTE equivalent.

Broader productivity

Roughly 2,900–4,300 hours a year, about 1.5–2 FTE equivalent.

Where this comes from

The research behind this Hub

Each choice below traces back to a specific study or company program, linked directly so anyone can check the source.

Anthropic Academy, free Claude courses

Anthropic's own courses, no cost, an email to enroll. For anyone new to Claude, "Claude 101" (about an hour) is the recommended starting point, covering the interface, Projects, Skills, and Artifacts.

claude.com/resources/courses

Stanford: The Enterprise AI Playbook

An academic review of 51 successful enterprise AI deployments (Pereira, Graylin, Brynjolfsson, March 2026). Used here as an outside check on this Hub's structure.

digitaleconomy.stanford.edu, Enterprise AI Playbook (PDF)

Peer networks drive real adoption

A 2026 study of AI coding-agent rollouts found usage spread through peer networks far more than through top-down mandates. It's why this Hub leans on department champions and working sessions rather than one company-wide push.

arxiv.org/abs/2607.01418

Workday: gains lost to rework

A January 2026 survey of 3,200 employees and leaders found nearly 40% of time saved with AI gets lost to fixing its mistakes, and only 14% of employees see a clear net gain. This is the case for real proficiency training, not just handing out licenses.

newsroom.workday.com

Deloitte's Claude rollout

Deloitte gave 470,000 employees Claude access and built a dedicated Center of Excellence, certifying 15,000 practitioners to support it. The tiered structure here (Terminology, AI Fundamentals, a champion track) follows the same shape at a scale that fits Hayden.

anthropic.com/news/deloitte-anthropic-partnership

Copies a summarize-this prompt for the source, and opens Claude Desktop with it ready to send if you have the app installed. Paste it into claude.ai either way.

Contribute

Ideas and working session notes

What people have actually submitted: a use case worth trying, or a summary from a working session.

Add one

Type

Department (optional)

Title

Details

Added.