Mike Tholfsen, a senior Microsoft 365 PM, just published a tour of five Copilot Chat features most people miss: the model picker, Pages, the Researcher agent, Notebooks and the Prompt Gallery. All five are real productivity wins. None of them are obvious from the default Copilot screen. This is the implementer's view from someone who deploys Copilot for a living: how to use each, and the compliance, data-scoping and governance question that needs to be asked before the team comes to rely on them. Pairs with our [post on the 5 Copilot agents most UK SMEs aren't using](/resources/blog/microsoft-copilot-chat-agents-sme-governance/) from yesterday.
The 5 tips in 60 seconds
Tholfsen's full video runs about 9 minutes and is worth a watch. Quick summary so you can decide which sections of this post to skip.
- Mode switcher and model picker. Top-right of the Copilot Chat window: switch between Auto, Quick Response and Think Deeper. Crucially, also switch the underlying model between OpenAI (the GPT family) and Anthropic (Claude). Most users never touch this control.
- Copilot Pages. Turn a Copilot Chat response into a live, AI-collaborative document. Talk to chat on the left, the page updates on the right. The page is a Loop component under the hood, so it embeds in Outlook and Teams.
- Researcher agent. Deep research with structured prompts, citations and export options (Word, PowerPoint, infographic). Uses Microsoft's "critique and council" approach: OpenAI drafts, Claude refines.
- Copilot Notebooks. A focused workspace built on specific files. Generates an audio podcast overview, mind maps, study guides, and can answer questions only about the notebook's scoped contents.
- Copilot Prompt Gallery. Hundreds of pre-built prompts filterable by job, industry and task. Three-dot menu in the new chat screen.
Source: Mike Tholfsen on YouTube. Credit to him for surfacing all five.
Now what every UK SME owner should know before turning these on for the team.
Tip 1: The model picker, and the compliance question Microsoft's PM didn't need to answer
This is the one we get most questions about, and the one Tholfsen, as a Microsoft PM, didn't need to address.
When you flip the model picker to "Claude", a lot of compliance officers, data protection leads and risk-conscious MDs see "Anthropic" and worry that their data has just left Microsoft's perimeter for a third-party AI provider in the United States.
It hasn't.
Microsoft hosts the Claude models inside Azure. When you pick Claude inside Copilot Chat, your prompts and any files you attach are processed under Microsoft's existing M365 Copilot data boundary. That includes EU and UK data residency commitments, the same DLP and audit logging you already have, and the same contractual position on training (Microsoft does not use your enterprise prompts to train the foundation models).
What this means in practice for a UK SME:
- Picking Claude rather than GPT inside Copilot Chat is a model choice, not a vendor change. Your compliance posture does not change.
- "Think Deeper" routes through a more capable model and takes longer. It also tends to cost more compute under the hood. Use it deliberately, not for every prompt.
- Prompt and response logging is identical across the model choices. Microsoft Purview Audit logs the interaction under `CopilotInteraction` regardless of which model you picked.
The one thing that does change with model choice is output style and quality. Claude tends to write longer, more structured responses; GPT tends to be punchier and faster. Mike's recommendation, and ours, is to put the model switcher into your daily workflow rather than leaving it on Auto. Some prompts are better served by one model than the other.
Practical rule: any prompt where the answer goes into a regulator-facing document, a board paper or a public statement is worth running through Think Deeper. Routine emails and meeting summaries are fine on Quick Response.
Tip 2: Copilot Pages, and the Loop governance the video skips
Pages are powerful. They are also Loop components, which most SME IT teams have not configured properly because Loop arrived quietly and was on by default in most M365 tenants.
What you can do with a Page: take a Copilot Chat response, click "edit in pages", talk to chat on the left, the page updates on the right in real time. Style shortcuts let you add emojis, change tone, convert to email or blog post format. Share the page with a click. Embed it inside Outlook or Teams as a live component that anyone with access can edit.
What most SMEs have not done:
- Confirmed default sharing scope. Loop pages default to organisation-wide visibility in most tenants unless an admin has tightened it. That means a "private notes" page that someone shares "with the team" can be discoverable by the rest of the company depending on how the link is shared. Pull up the SharePoint admin centre, find the Loop settings, decide what your default should be.
- Set retention. Loop pages live in the user's OneDrive by default. When that user leaves, the OneDrive lifecycle policy applies. If you have a 30-day "delete leaver OneDrive" policy and your team has been collaborating on a Loop page that lives in that user's OneDrive, you have a problem.
- Audited who has edit access. Loop's permissions model has its own quirks. A page can be edited by people who have not been explicitly invited if it has been shared via Teams.
- Configured DLP. DLP on Loop is separate from DLP on email and Teams. It needs to be enabled.
None of this means avoid Pages. They are genuinely useful. It means run the 30-minute governance review before encouraging the team to put their working notes there.
Tip 3: Researcher, and what to do with citations the model is "sure" about
Researcher is the closest thing in Copilot to a competitor to ChatGPT Deep Research and Perplexity. Pick a structured prompt (market trend analysis, competitive analysis, status report), attach your own files for internal context, and Researcher runs a long parallel job that produces a detailed, cited report. You can export it to Word, PowerPoint or even an infographic.
It is impressive. It is also wrong in subtle ways that matter for a UK SME using it on regulator-facing or client-facing work.
Three things we tell clients:
- Cited does not mean correct. Researcher cites the URLs it consulted, but the link between a specific claim and a specific source is generated, not verified. We have seen Researcher cite a paywalled academic paper for a statistic that does not appear in the paper. The cited URL exists. The cited statistic does not. Anyone shipping Researcher output to a regulator without a manual fact-check is taking a risk that the output itself does not flag.
- Critique and council is interesting, but read both. Microsoft's research found that having OpenAI draft and Claude refine produces measurably better answers. Useful default. But for any prompt where the answer matters, switch to "Model Council" mode (visible in the same model picker) and read both responses side by side. Where they agree, you are probably safe. Where they disagree, you have something to think about.
- Internal files plus public research is the most useful pattern. Researcher's competitive analysis prompt is genuinely strong when you give it your own product specs, GTM plan and competitor list as attached files, and let it overlay public market data. Without your internal context, the output is generic; with it, the output is sharper than most £5,000-day analyst reports we have seen ordered by SME marketing directors.
Practical rule: Researcher is a first draft, not a final answer. Treat it like a junior analyst's brief, not the conclusion.
Tip 4: Notebooks, and the data-scoping question that actually matters
Notebooks are the most underused of the five tips and, in our view, the most useful for an SME.
A Notebook is a focused workspace bound to a specific set of files. Anything you ask the Notebook is answered using only those files (plus a configurable amount of public web context). The audio podcast overview is genuinely good. The mind map is a fast way to onboard a new joiner. The chat lets you ask "what are the common themes" or "where does the contract terminate" and get answers grounded in the documents you put in.
The implementation question is: what's in the Notebook is what the AI sees, and what anyone with access to the Notebook can ask about.
A few examples we have actually walked clients through:
- A finance team built a Notebook over the management accounts and a folder of contracts. Useful for quick "what does the renewal terms look like for client X" lookups. The Notebook also had visibility of an HR memo about a salary review for three named staff because it was in the same folder. Shareable link. Solvable problem, but only if someone notices.
- A sales team built a Notebook over the deal pipeline spreadsheet, the meeting notes from the last six months and the customer email threads. Useful for "what's the latest position on Account X". Also means a junior sales person can ask "what was discussed in the leadership pricing call last month" if that meeting note is in the same folder.
The fix is not "don't use Notebooks". It is:
- Build a Notebook around a deliberately scoped folder, not "everything I have access to".
- Audit the folder before adding it to the Notebook.
- Treat the Notebook's share permissions as data permissions, not document permissions.
- For client work, build the Notebook per-client. Re-using one Notebook across clients leaks context.
This is the same scoping discipline that worked for SharePoint sites for years. The mistake is assuming a Notebook is "just a smart chat", when it is actually a new data surface with its own access pattern.
Tip 5: The Prompt Gallery, and why quality varies
The Prompt Gallery is the easiest of the five tips to adopt. Three-dot menu, filter by job or industry or task, find a prompt that fits what you are trying to do, hit suggest, edit the placeholder, run.
It is also the tip where Microsoft's curation runs out at the edges. Some prompts are excellent. Some are written for enterprise-scale teams and read as bizarre when adopted by a 30-person SME. A few read as if they were drafted by the product team rather than the customer.
A few patterns we use:
- Use the industry filter to find the closest match, then strip the assumptions. A "financial services" prompt that asks for "your investment committee briefing pack" can be adapted to "your monthly client review" without losing the structure that makes the prompt work.
- Read the prompt before running it. The Gallery prompts are visible. Most of the time, they are doing two or three useful things you can lift into your own prompt library.
- Keep your own gallery. Once you find five or ten prompts that work for your business, save them in a OneNote or a SharePoint List. The Gallery rotates; your prompts shouldn't.
The KPI that ties this back to yesterday
Yesterday's post set a hard productivity target: 4 hours saved per user per month on a £28 monthly Copilot licence. If you are getting less, we said the problem is the rollout, not the technology.
These five tips, together, are the most direct route to that 4-hour KPI for a typical SME team:
- Model picker saves 5 to 10 minutes per Think Deeper prompt that would otherwise have taken three quick prompts in a row to refine.
- Pages save the "Copilot answer to Word document to share via email" round trip that used to take 4 to 8 minutes per task.
- Researcher replaces the 30 to 90 minutes a marketing manager or analyst would otherwise spend doing the same desk research.
- Notebooks save the "where was that file again, what did we agree" lookups that quietly eat 15 minutes per person per day.
- Prompt Gallery saves the "how do I phrase this" thinking time for prompts the team writes monthly.
We have measured the median user goes from 1 to 2 hours per month saved (chat only) to 5 to 7 hours per month saved (chat + 2 to 3 of these tips properly deployed). That is the licence paying for itself, with margin.
The catch is that none of this happens by accident. It happens because someone shows the team how to use the controls, the team agrees on a few use cases, and the IT team sets up the governance bundle (DLP, retention, audit) so the productive use is also safe use.
Bottom line
Tholfsen's tour is good. Microsoft's product team has done a lot of work to make these features accessible. The five tips are real, the demos work, the productivity gains are credible.
The piece Microsoft cannot put into a 9-minute YouTube video is the compliance posture, the data-scoping conventions and the team-level adoption pattern that turn the features from interesting demos into a measurable hours-saved-per-user-per-month figure. That work falls to whoever runs IT in your business.
If that's an internal team with time to read the Microsoft Purview docs and configure DLP for Loop, you have everything you need to make this work. If it isn't, that's the conversation we have most weeks.
About this article
Source video: Mike Tholfsen: "My 5 Favorite Microsoft Copilot Tips to Get More Done". Watched and synthesised on 2026-06-13.
Companion post: The 5 Copilot Chat agents most UK SMEs aren't using, and the governance reality to fix first (published 2026-06-12).
Further reading:
- Microsoft 365 Copilot data, privacy and security
- Microsoft Purview Audit (Copilot Interaction events)
- Loop component administration in Microsoft 365
- Microsoft 365 Copilot data residency and EU Data Boundary
About Inflection Point
Inflection Point is a UK managed-IT and cyber-security firm. 200+ active clients across the UK, 16+ years EOS-run, founder-led with 25 years industry experience. ISO 27001 certified, Cyber Essentials Plus, Microsoft Solutions Partner, rated 4.9/5 across 150+ Trustpilot reviews. We help UK SMEs deploy Microsoft 365 Copilot productively and safely, with under-15-minute remote response and 1-hour on-site SLA. From £39 per user per month for managed IT.
If you have Copilot licences and want a structured rollout that hits 4+ hours saved per user per month, book a 30-minute discovery call. We will tell you which of these five tips to deploy first, what to lock down before turn-on, and what the rollout looks like for your specific team.
Iain Godding is the founder of Inflection Point. He has 25 years of UK IT and cyber security industry experience.
