A Framework for Microsoft Copilot, Cowork, Agents & Notebooks

The four layers that make Microsoft Copilot useful (and why most teams skip three of them)

Microsoft Copilot has a reputation problem. Teams roll it out, ask it a few questions, get bland answers, and quietly stop using it. The licences keep getting paid for. The promised productivity lift never arrives.

The diagnosis is simple. Copilot isn't Claude. It isn't ChatGPT. Treating it like one of those tools, then judging it on the same criteria, is what destroys the value before you start.

If your organisation is well licensed for M365 and the Microsoft Graph holds most of your data but you can't work out how to make Copilot work for you and you keep falling back to Claude or Chat GPT - read on.

What Copilot actually is

Claude and ChatGPT are general-purpose reasoning tools. You give them a hard problem, they think about it, they web-search for fresh material, they write a sharp answer. They get smarter the harder you push them.

Copilot is something different. It's a grounded assistant that talks to your work data inside Microsoft 365. Its job is to read your emails, files, meetings, chats, and calendar, then act on that context. Where it falls short is exactly where Claude and ChatGPT shine: open-ended reasoning, deep web research, and complex multi-step thinking from a blank page. Independent reviewers have been clear about this for over a year. Copilot "may struggle when tasks require long-form, multi-step reasoning, abstract problem solving, or conceptual exploration without a document on which to anchor the generative exercise", and its creative and analytical capabilities are "generally a step behind ChatGPT and Claude for complex reasoning tasks".

That's not a flaw. It's the design. Microsoft built Copilot to ground its answers in your data. The trade-off was that Copilot, on its own, has been weaker at the reasoning-from-scratch work that ChatGPT and Claude do well.

The good news: a recent shift in the platform has closed the biggest of those gaps. The other gaps are gaps in your data and your workflows, not gaps in the tool. We've spent the last year quietly fixing this at MomentumIQ, both for our own consulting work and inside several client tenants. The change has been substantial enough to be worth writing up.

Here's the four-layer system.

Layer 1: External context, finally solved

The single biggest weakness of Copilot has been deep web research. Where Claude and ChatGPT browse, reason, cross-check, and write a long structured report, Copilot has historically returned thin summaries off Bing.

In April 2026, that changed. Microsoft added Anthropic's Claude Opus 4.7 as a model option inside Copilot Cowork, available through the Frontier program. Per Microsoft's announcement, Opus 4.7 "handles complex, multi-step work with rigor, following your instructions more closely, checking its own outputs before responding, and reading images at higher resolution". Microsoft also confirmed that "working closely with Anthropic, we have integrated the technology behind Claude Cowork into Microsoft 365 Copilot".

For non-technical readers, here is what that means in practice.

Copilot Cowork is a new agent inside Microsoft 365 Copilot that runs long, reasoning-heavy work in the background. You describe what you want, and Cowork builds a plan, runs across your tools and files, holds context across many steps, and creates folders and documents inside your OneDrive or SharePoint as it goes. Each action gets a preview and an approval before it runs.

Frontier is Microsoft's early-access programme for new Copilot capabilities. Your IT admin enrols the tenant, then enables Cowork for users who need it.

Opus 4.7 is Anthropic's most recent flagship model. With it selected as the engine for Cowork, the reasoning quality steps up considerably. The same underlying technology that powers Claude's deep research is now running inside your Microsoft tenant, with Microsoft's identity, compliance, and data protection wrapped around it. Cowork is also able to use SKILL.md files like Claude.

We use Cowork with Opus 4.7 for two recurring jobs:

• Weekly market intelligence. A standing prompt that researches our sector, our clients' sectors, regulatory updates, and competitor moves. The output drops as a structured report straight into a SharePoint folder.

• Client deep-dive briefings. Before any major client conversation, a deep research run pulls together a full picture of the company, leadership, recent news, financial signals, and likely AI maturity questions.

This is the external context layer. Quality external information, captured inside your tenant, on a schedule.

One critical discipline: anything generated by AI without a human verification step gets named 'Market Research - Unverified' before it goes into any shared folder. This protects you from quietly relying on a hallucinated fact later. AI-assisted research is a starting point, not a finished product.

Layer 2: Get your internal data in order

This is the unglamorous work that makes everything else possible.

Copilot is only as smart as the data it can find. If your SharePoint is a mess of duplicated folders, half-named files, and ten different versions of the same document, Copilot will return vague, unhelpful answers because the underlying data is vague and unhelpful.

By the time you're considering this, your data probably already lives in SharePoint. The job isn't to move it. The job is to make it findable.

For us, that meant:

• One folder per active client, with a fixed subfolder structure (signed contracts, deliverables, working drafts, transcripts, supporting docs).

• Permission control at the folder level, so the right people see the right things.

• No use of personal OneDrive as a place where client work hides.

• A naming convention that includes the date and version, so the most recent file is obvious.

• A "deliverables" folder per client that becomes the single source of truth for finalised work.

  
  

This isn't AI work. It's information architecture. It's also the layer that nobody wants to do, which is why most Copilot rollouts stall before they start. There is no clever way around it. You need to know where your data is, and your data needs to be clean.

If you can't honestly answer the question "where is the latest version of [X] for [client]?" in under thirty seconds, this is your starting point.

Layer 3: A living collaboration layer

Files in SharePoint are static. The current state of an active engagement is not.

You need a place where people working on a client today can capture meeting notes, decisions, action items, status updates, and contact changes. That place needs to be editable, shareable, and easy to write into.

We use Microsoft OneNote, with one page per client. We split active clients and prospects into separate notebooks, which makes it trivial to drag a prospect across to the active notebook the day a contract gets signed. Beyond client notebooks, we run a separate notebook with our SOPs, new-starter information, and key facts about MomentumIQ itself. That notebook grounds Copilot in the context of our own business, the same way the client notebooks ground it in the client.

For project work that spans multiple clients or is purely internal, we use Microsoft Loop. Loop is built for fast, real-time collaboration on living documents and works well for short-lived project pages. We link Loop pages back to the relevant client OneNote page when they belong to a specific account.

The discipline that makes this work: every client OneNote page starts with quick links to the critical files and folders for that customer. Signed SoW. SharePoint folder. Market research folder. The OneNote page is the first place anyone joining the engagement looks.

Layer 4: Orchestration with Copilot Notebooks

This is where everything connects.

Copilot Notebooks let you point Copilot at a curated set of sources and give it instructions on how to use them. The important thing about a Copilot Notebook is that it is not a static knowledge base. It's a set of links to live sources. You add the OneNote page, the signed legal documents, the critical SharePoint folder, and the external research (clearly named as unverified if it has not been checked). Copilot then reasons across that source set whenever you ask it a question about that client.

Three things make this useful:

• You only set it up once. The notebook references live sources, so as your OneNote and SharePoint update, the notebook stays current without you having to re-import anything.

• You give Copilot instructions on source hierarchy. Signed SoW above OneNote above SharePoint above market research, with clear rules on what each source is for. Scope, fees, dates: SoW only. Current state: OneNote first. Sector framing: market research, never to override scope.

• Citations come back with answers. Anyone using the notebook can see exactly which file or page the answer came from, which means you can trust it for actual decisions, not just background context.

We set up the SharePoint plus OneNote plus Copilot Notebook flow for every active client. We do the same for major internal operations workflows. Once the pattern clicks, you start spotting opportunities for it everywhere.

Why this works at any size

Three things become true once this is in place.

The shape stays the same as you grow. You just need to design it for your business whether it's services, products, projects or clients you build your structure around. New starters know exactly where to look because every client looks identical structurally. Cowork can help you build that new structure the same way every time too.

It's secure by default. Everything sits inside your Microsoft 365 tenant. SharePoint permissions control who sees what. OneNote and Copilot Notebooks share to specific people. Nothing is exported. No third-party tool holds your client data.

The workflow stays when people don't. This is the one most organisations quietly worry about. When an employee leaves, the workflow they built typically walks out with them. Personal OneDrive folders. Notes in their own apps. Knowledge in their head. With this system, the workflow lives in the business. Anyone with the right permissions picks up where the previous person left off, with full context, the same day.

For most businesses, that's the difference between growing the business and constantly rebuilding it.

What's coming next, and why your structure is the bet that pays off

Microsoft is no longer a single-model platform. As Microsoft put it in the Cowork announcement: "it is this multi-model advantage that makes Copilot different. Your work is not limited by one brand of models. Copilot hosts the best innovation from across the industry and chooses the right model for the job regardless of who built it". Anthropic's Claude is now embedded across Copilot Cowork, Copilot Studio, Copilot in Excel, and Researcher.

The four-layer system above doesn't depend on any single AI model. Whether the reasoning engine inside Copilot is GPT, Claude, or whatever comes next, your SharePoint stays organised, your OneNote stays current, and your Copilot Notebooks stay pointed at the right sources. The AI gets better. Your structure doesn't need to change.

That is the right way to think about all of this. Your AI tools will keep changing. Your data discipline shouldn't have to. The people change and embedding these workflows and good data hygiene will be your biggest hurdle.

A useful sanity check

If you're not sure whether your Copilot setup is doing real work, three questions:

1. Think about a key person risk across your business. If they left tomorrow, could a new starter be productive on their accounts within a week, using only what's in SharePoint and OneNote?

2. When you ask Copilot a hard question about a client, does it cite the right document, or does it give you a generic summary?

3. Do you have one clear place where every active client's current state lives, or do you go hunting across email, OneDrive, Teams chats, and someone else's notes?

Three nos means you have work to do.

The bigger point

There is no one AI tool that does every job. Copilot, Claude, ChatGPT, Gemini all have strengths and gaps. Copilot's strength is talking to your work data at enterprise scale, inside an environment your IT, security, and compliance teams already trust. Its gap, until April 2026, was deep external reasoning. With Cowork on Opus 4.7, that gap is now closed for organisations enrolled in Frontier.

What's left is the work most teams are skipping: getting their data clean, their living state captured, and their orchestration set up. Once that's in place, every Copilot conversation gets sharper. Every new model release gets easier to plug in. Every new starter gets productive faster. And the workflow you paid to build doesn't walk out of the building when someone resigns.

A visual aid to put it all together…

What this isn't.

This system solves real problems. It also has limits worth naming up front.

Risk	Mitigation
Setup is the easy part. The discipline of keeping OneNote current and SharePoint clean is the hard part.	Build OneNote updates and SharePoint hygiene into team rituals. Weekly status meeting, end-of-engagement checklist, named accountability per client. Without that, the system decays within months.
Cowork on Opus 4.7 requires Frontier enrolment and Anthropic enabled as a subprocessor in your tenant.	Get your IT admin to enable both before committing to Cowork operationally. In some regulated sectors and regions this takes weeks. The rest of the system still works without Cowork.
Long Cowork research runs can be slow. A deep client briefing can take 10 to 20 minutes.	Use Cowork for scheduled or pre-meeting work, not real-time questions. Keep Copilot Notebooks as the fast-answer layer for live conversations.
Copilot Notebook citations are not always accurate. Copilot occasionally cites a source that doesn't contain the claim.	Treat citations as verification prompts, not proof. For decisions of consequence, click through and read the source. The citations make verification fast, not unnecessary.
Copilot Notebooks have source limits. Linking a folder doesn't mean every file is treated equally.	Link the most important sources directly (signed SoW, OneNote page, key deliverables), not entire folders. Keep the source list curated, not exhaustive.
"Inside the Microsoft boundary" is only as secure as your underlying setup. Default M365 settings are not enough for sensitive client data.	Review SharePoint permissions, conditional access, sharing policies, and DLP before granting Copilot access to client folders. Treat the security baseline as a precondition, not an output.
Anthropic-as-subprocessor changes your data flow. Some regulated clients (finance, government, health) require explicit approval of new subprocessors.	Check client contracts before enabling Claude in their workflows. Document the subprocessor chain. Have an alternative path (Copilot without Cowork) for clients who can't approve.
Source hierarchy in your notebook instruction is a soft rule, not an enforced one. Copilot may not always follow the priority order you've written.	Test the notebook with known-answer questions after setup. If hierarchy isn't holding, restructure the instruction. Re-test after any major model update.
You're building on product features Microsoft will keep changing. Notebooks may be replaced by declarative agents within two years.	Keep the architecture (folder, OneNote page, orchestration layer per client) stable, and treat the specific tools as replaceable. The pattern survives a tool change.

Where MomentumIQ comes in

We built this for ourselves first. We've now built versions of it inside client tenants. The work splits into three pieces:

• Information architecture (getting SharePoint, OneNote, and Loop set up properly).

• Workflow design (the per-client or per-project pattern, the SOP notebook, the new-starter onboarding).

• AI orchestration (the Cowork prompts for external market research, Copilot Notebook templates, source hierarchy rules).

• Behaviour change - training, FAQ's, finding edge-cases, AI agents to assist with compliance

If you're paying for Microsoft 365 Copilot and not getting the value back, drop us a message. We would love to discuss your experience and explore if the Four Layers might be a good solution for you. 

Sources

• Microsoft Tech Community: Available today: Anthropic Claude Opus 4.7 in Microsoft 365 Copilot (April 2026)

• Microsoft Tech Community: (Co)work in Progress (April 2026)

• 4sysops: Microsoft Copilot Cowork: New AI features for Microsoft 365 (May 2026)

• Solway Stories: What's the Difference Between Copilot, ChatGPT, and Claude for Business? (March 2026)