A working guide · For Jodi Schafer

A personalized Copilot team starter guide

Prepared by
Carl Roose, Lead AI Advisor
Updated
May 20, 2026
Shelf life
3–6 months

A note on dates: Microsoft renames things constantly, so feature names here are current as of May 20, 2026. If something looks different in your tenant in a few months, the underlying concepts will still apply.

If you read nothing else

There’s no master switch—you build four small layers, in order.

Don’t try to “load HRM into Copilot” as one master upload—there isn’t a single switch for that. What’s actually doing the work in the background is the Microsoft Graph: SharePoint, OneDrive, Outlook, Teams, meetings. Copilot can already see everything you have permission to see.

What you’re really doing is building four small layers on top of that: (1) a clean, AI-ready SharePoint source of truth, (2) custom instructions for yourself, (3) Notebooks for project-bound work, and (4) Agents for repeatable team workflows. Do them in that order and the rest takes care of itself.

1Orientation

What each thing in your Copilot menu actually is

Your sidebar lists Agents, Notebooks, Library, Create, and Apps. Here’s the plain-English version, in roughly the order you’ll use them.

One thing in your top bar worth understanding

The Work/Web toggle and the Quick / Deep / Auto thinking control. Keep Work selected for almost everything—that’s what grounds Copilot in SharePoint and your other internal content. Switch to Web only when you genuinely need outside-world information.

On thinking depth: your instinct to start in Deep was right. Quality drops with each follow-up turn if you start in Quick, and starting in Deep mostly avoids that problem. Quick is fine for “rewrite this in a warmer tone.” Deep is for proposals, investigation framing, anything with real judgment in it.

2Foundations

How to load company-level context

Several layers, in order of how much they’ll help you.

Your SharePoint grounding is already doing the heavy lifting

When you saw Copilot pull from your HR-outsourcing materials and cite them—that’s the most important “context” mechanism, and it’s already working because your client files, templates, and prior work live in SharePoint. The Microsoft Graph stitches it together. You don’t need to do anything else for Copilot to use those sources, as long as the user has permission to see them.

Custom instructions

The next layer. In Copilot Chat, open settings (the gear icon) → Custom instructions. This is where you tell Copilot persistent things about how to respond to you:

“Default to a warm, direct, plainspoken tone; never consultantese.” · “When drafting external communications, sign as ‘Jodi Schafer, HRM.’” · “Always ask before recommending termination language.”

These follow you across every conversation until you change them. They’re per-user, not per-tenant—so each person on your team sets their own.

Brand Kits

The tenant-wide layer for visual identity—logos, color codes, fonts, approved imagery, brand voice. You create them in the Microsoft 365 Copilot app under Create → Brand kits. Once published as an “Official” brand kit, anything generated through Create or Copilot in PowerPoint follows them. Caveat: full Brand Kit functionality typically requires a Microsoft 365 Copilot Premium license, not just the standard add-on. Check what tier you’re on before going down this path.

Prompt Gallery

The tenant-level home for shared prompts. In Copilot Chat you can save a strong prompt and share it with individuals or groups; over time you build a library of “the way HRM asks Copilot for an inquiry response,” “the way we ask for a job description,” and so on. It’s a lighter-weight standardization tool than full agents—good for the team while you’re still figuring out which workflows deserve their own agent.

Work IQ

Microsoft’s name for the underlying intelligence layer that connects all your tenant data, memory, and custom instructions and feeds them into every Copilot interaction. Not something you “configure” directly—it’s the plumbing. But it’s why grounding in your work content works as well as it does.

One honest caveat: there isn’t one simple “load HRM’ brand voice once and every Copilot answer from every team member now follows it” switch. The reliable way to get consistency is the four-layer stack above—clean source documents, custom instructions, Notebooks, Agents—plus shared prompts in Prompt Gallery where it helps.

3Source of truth

What to put in SharePoint first

Before any of the agent or notebook work, take an afternoon and build a small, deliberate “AI-ready” SharePoint area. Keep it cleaner than your general client archive. Permissions matter—Copilot only surfaces what each user can see, but it will happily surface anything they can see, so check before you publish broadly.

A starter set of folders worth having:

  1. 1
    Brand and voice

    Your voice guide (two paragraphs is plenty—who you write for, what you sound like, what you avoid), preferred and avoided phrases, two or three sample on-brand paragraphs, logo/color/font rules.

  2. 2
    Services and offerings

    HR outsourcing overview, workplace investigations overview, training and facilitation, compliance consulting, recurring deliverables, standard exclusions and caveats.

  3. 3
    Proposal system

    Proposal template, approved scope language, two or three sample winning proposals, discovery questions, assumptions and dependencies, standard next-steps language.

  4. 4
    Inquiry and email language

    Standard inquiry response, interest-form language, “let’s get on a call” language, scheduling-link rules, polite-decline language, follow-up sequences.

  5. 5
    Job descriptions and postings

    JD template, posting template, EEO/accommodation language, pay-transparency guidance for the jurisdictions you cover, inclusive-language checklist.

  6. 6
    Investigations

    Workflow and SOP, interview-plan template, report template, evidence-log template, confidentiality and privilege guidance, human-review requirements.

  7. 7
    Thought leadership

    Prior video transcripts, approved blogs and articles, partner-publication examples, your content calendar, audience definitions, sample social posts by platform.

  8. 8
    Quality and guardrails

    Your “must verify” checklist, citation expectations, what Copilot may draft versus what a human must decide, legal and compliance review triggers, client-confidentiality rules.

While you’re at it: archive or delete obviously stale templates, lock down sensitive client and investigation folders, and check for overbroad sharing. SharePoint hygiene is the single biggest predictor of how good Copilot’s output will be once you start pointing agents at it.

4Working pattern

A prompt pattern for the whole team

Microsoft’s own recommended structure is four ingredients: goal, context, source, expectations. It’s worth making this HRM’s default and putting it in your shared prompts.

The template
Goal
What I want you to create or answer.
Context
Who it’s for, why we need it, constraints, tone, decision stage.
Source
Use these files, emails, notebook references, SharePoint folders. Don’t use other sources unless I ask for web research.
Expectations
Format, length, sections, citations, what to flag, what not to invent.
A real example · inquiry response
Goal
Draft a first response to this inbound inquiry.
Context
Prospect is asking about HR outsourcing for a 75-person professional services firm. Tone: warm, confident, practical.
Source
Use HRM’ inquiry-response templates and the HR outsourcing service overview. Use the email below as the prospect’s ask.
Expectations
Under 175 words. Include a placeholder for our scheduling link. Don’t quote pricing or promise availability. End with two optional clarifying questions.

The point isn’t bureaucracy—it’s that “draft an inquiry response” gives Copilot 90% less to work with than the version above, and the difference shows up in the output every time.

5Build

Building reusable agents

You have two tools for this. For everything you described on the call, you want the simpler one.

Agent Builder

Lives inside Copilot itself (Agents → New agent). No code, no separate portal. You describe what you want, point it at SharePoint folders or specific files for knowledge, write instructions in plain language, and share it with named people or groups. This is the right tool for proposal drafting, inquiry responses, job descriptions, and any of the other repeatable work you mentioned.

Copilot Studio

The heavier sibling—a separate web portal at copilotstudio.microsoft.com. It exists for agents that need to do things outside Copilot itself: talk to a CRM or HRIS, trigger approvals, route inquiries by service line, publish to external audiences, multi-step workflow logic. You don’t need it yet. If a use case ever outgrows Agent Builder, Microsoft lets you copy the agent into Studio without starting over.

Three agents worth building first, in this order.

Agent 1

Proposal drafting assistant

Purpose
Turn discovery notes or an RFP into a first-draft proposal in HR Matte rs’ structure and voice.
Knowledge sources
Proposal template, services and offerings overviews, two or three approved sample proposals, discovery and intake questions, brand voice guide, standard assumptions and caveats.
Instructions (sketch)
Draft from approved HRM source language first. If the client ask is missing scope, audience, timeline, deliverables, buyer, or decision criteria, ask before drafting. Keep visible placeholders—don’t invent fees, dates, client names, legal claims, or staffing commitments. Match HRM’ voice: warm, direct, practical, never consultantese. Cite or link source files when possible. End with a short “human review needed before sending” section.
Starter prompts
  • “Create a proposal outline from these discovery notes.”
  • “Draft a first proposal using HRM’s template and approved service language.”
  • “Rewrite this scope section in our voice, keeping the commitments unchanged.”
  • “List what’s missing before this proposal can go to the client.”
Agent 2

Inquiry response assistant

Purpose
Draft consistent responses to inbound inquiries and route people to the right next step.
Knowledge sources
Inquiry email templates, interest-form language, scheduling and Bookings instructions, service-line descriptions, polite-decline language.
Instructions (sketch)
Start by classifying the inquiry by likely service line. If it’s promising, draft a warm, concise reply that invites a call and includes the right scheduling-link placeholder. If details are missing, ask only for the minimum needed. Don’t promise availability, quote pricing, offer legal conclusions, or predict outcomes. Match HRM’s tone: warm, credible, plainspoken.
Starter prompts
  • “Draft a reply to this inquiry with the right next step.”
  • “Turn this messy inbound email into a warm response with a scheduling-link placeholder.”
  • “Write a follow-up if they haven’t responded after a week.”
Agent 3

Content repurposing assistant

Purpose
Turn your monthly video transcript into blogs, partner articles, and social posts while preserving your original thinking and voice.
Knowledge sources
Prior video transcripts, approved blogs, partner-publication articles, sample social posts by platform, brand voice guide, audience definitions.
Instructions (sketch)
Treat the transcript as the source of truth. Preserve your point of view, but clean up spoken-language artifacts. Don’t add new factual or legal claims without flagging them for verification. Generate derivative content by channel: blog, LinkedIn, partner publication, newsletter, short social. Always end with a “fact-check needed” list of anything that should be verified before publishing.
Starter prompts
  • “Create a blog post from this transcript in HRM’s voice.”
  • “Give me five LinkedIn post options from this transcript.”
  • “Turn this transcript into a 900-word partner-publication article. Preserve my POV and flag any added claims.”

A note on the video-to-content workflow: it can also live as a Notebook (one per video, with the transcript plus your style guide plus an example article you liked). Both work. The agent version scales better once you’ve nailed the prompts; the notebook version gives you more room to iterate on a single piece. Try both on the next video and see which feels right.

Don’t try to build all three agents at once. Pick one—probably proposal drafting—and get it actually good before moving to the next. Test on three real recent examples, refine the instructions, share with one person on the team, and watch them use it. The SOPs Carl mentioned on our call belong inside the agent’s instructions; the level of granularity that feels almost obsessive for humans is exactly right for an agent.

6Day-to-day

Getting Copilot drafts into the actual email

The frustration you described—Copilot drafting in the right-hand chat pane but no clean way to get it into the email—was because you were using the wrong entry point.

The right way

In-line in the email itself

Open a new email, reply, or forward. Inside the email body you’ll see a “Draft with Copilot” link, or click the Copilot icon in the email’s toolbar and choose Draft. A prompt box appears inside the email. Type what you want, hit Generate. Review the result, optionally adjust length or tone, then click Keep it—the draft is in your email body, ready to edit and send.

The wrong way

The right-side chat pane

That panel is Copilot Chat. It’s great for asking questions, but it doesn’t have a direct “insert into the email I have open” action. If you draft an email there, you’re copy-pasting.

Two more useful things while you’re in Outlook: typing / inside an email body opens a quick menu including “Draft with Copilot.” And Coaching by Copilot, on the same icon, does the opposite of drafting—you write your own message, then ask Copilot to critique it for tone, clarity, and how it might land. Often more useful than having Copilot write it for you in the first place.

7Under the hood

Models: what the picker actually means

Microsoft has gotten more model-pluralistic over the last six months. As of mid-May 2026, here’s what you’ll typically see.

OptionWhat it isWhen to reach for it
Auto · Quick · Think deeper Thinking-depth toggle Not a model choice. Auto picks for you, Quick is for routine answers, Think deeper uses slower reasoning. Default to Think deeper for anything with judgment in it.
GPT-5.5 (Instant / Thinking) OpenAI flagship Released April 23, 2026. Thinking rolled into Copilot April 27; Instant May 7. Strong across the board—analysis, writing, reasoning, math.
Claude Opus 4.7 Anthropic flagship Added to Copilot April 16, 2026. Lives in Copilot Cowork, Copilot Studio (early release), Copilot in Excel, and now the Chat model picker. Often a slightly warmer, more compositional sensibility on writing tasks.

Both are at the top of the field. The differences are real but subtle, and any honest comparison says try both, pick whichever sounds more like you on a given task. For a high-stakes deliverable—a proposal you really care about, a sensitive client communication—running it through both and picking the better starting point takes 30 extra seconds and is worth it.

One administrative note: Anthropic models in Copilot Chat roll out through Microsoft’s Frontier Program. As of January 6, 2026, they’re enabled by default in US tenants (EU, UK, and EFTA tenants still require admin opt-in). You likely already have access. If Claude doesn’t appear in your model picker, that’s the box to check, under Microsoft 365 Admin Center → Copilot → Settings → User Access → Copilot Frontier.

8Licensing

What you’re actually paying

Your hunch that pricing has come down was right.

Probably you

Microsoft 365 Copilot Business

$18/ user / month

Annual billing, through June 30, 2026. Then $21/mo. Month-to-month: $25.20.

For organizations up to 300 users. Requires a qualifying Microsoft 365 Business plan as the base license.

Larger orgs

Microsoft 365 Copilot Enterprise

$30/ user / month

Annual commitment.

Requires E3, E5, Business Standard, or Business Premium as the base.

So you’re probably paying $18 or $21 per Copilot-licensed user, depending on when you signed up. Worth confirming in the admin center under Billing → Your products, or with your Microsoft reseller. If you signed up before SMB pricing existed and you’re still on the old $30 rate, ask about switching to Copilot Business—you qualify.

The promotional $18 rate locks in if you renew before June 30, 2026, so that’s a real deadline worth knowing for your renewal timing.

9Caveats

A few small things worth flagging

Marketplace agents and plugins

Microsoft has an AppSource marketplace where third parties publish agents and connectors. Some are excellent and save real time; some are random, some are sketchy. Treat anything you adopt like any other vendor: who made it, what data does it touch, what does it cost. Your admin (you) controls which agents are enabled in the tenant via Microsoft 365 Admin Center → Copilot → Settings → Agents. Don’t enable broadly without reviewing first.

The SurveyMonkey hallucination you caught

A perfect example of why your “I want to understand it so I can explain it to the team and figure out what the guardrails are” instinct is exactly right. Embedded AI in third-party tools (SurveyMonkey, Notion, whatever) often uses some background model the vendor doesn’t disclose, with wildly varying reliability on math and numbers. Anything quantitative coming out of an AI feature should be spot-checked against the raw data until you’ve calibrated how trustworthy that particular tool is. You did the right thing catching it before the report went out.

Prompt injection risk

When you start pointing agents at content from outside HRM—client files, downloaded templates, documents from prospects—there’s a small but real risk that hidden instructions in those files will try to manipulate the agent. For now: stick to your own SharePoint content as agent knowledge sources, and be a little suspicious of templates from the open internet.

10For the team

Guardrails worth sharing with the team

A short version you can hand to the team verbatim:

  1. Copilot is a drafting and analysis assistant, not the final authority.
  2. It can be wrong, and the same prompt can produce different answers.
  3. It must not invent client facts, fees, deadlines, legal claims, investigation findings, or HR recommendations.
  4. Work mode is for internal context; Web mode is for outside or current information—and Web results should be verified.
  5. Sensitive investigation and client files stay in properly permissioned folders.
  6. If a task carries legal, compliance, employment, privacy, or reputational risk, a qualified human reviews before use.
  7. Start a new chat (or type “new topic”) when switching tasks—Copilot is limited to the current conversation’s context, and stale context muddies the output.
11Plan

A 30-day rollout

Week 1

Clean the source

Build the AI-ready SharePoint folder structure. Add brand, service, proposal, email, content, and guardrail documents. Archive obviously stale templates. Permission-check the client and investigation folders.

Week 2

Standardize prompts & lift everyday quality

Set your own custom instructions (15 minutes, immediate quality lift). Create 6–10 reusable team prompts using the goal / context / source / expectations pattern. Save to Prompt Gallery or a shared Loop page. Show the team Work vs. Web, Think Deeper, and the Outlook in-line drafting flow.

Week 3

Build two agents

Proposal Drafting Assistant first, then Inquiry Response Assistant. Test each on three real, non-sensitive recent examples. Revise the instructions based on what goes wrong, not what goes right. Share with one person on the team and have them try it cold.

Week 4

Add content workflows

Build the Content Repurposing Assistant. Also try one Notebook for the next monthly video—same source, different surface—and decide which feels better for that work going forward.

12Onwards

Where to go from here

Skim Microsoft’s documentation for Agent Builder and Notebooks when you have a quiet hour—both are short and current. Search “Microsoft 365 Copilot Agent Builder” and “Microsoft 365 Copilot Notebooks” on learn.microsoft.com.

Holler with questions as you go—happy to look at an agent’s instructions and tell you where they could be tighter, or to talk through whether a particular workflow belongs in an agent versus a notebook versus a quick chat. And let’s get something on the calendar to walk through what you’ve built once you’ve got a couple of agents running.