Copilot: What you do &
don't know:
Where does Copilot get its data?

Hi all, welcome back to episode two of Copilot, the enforceway as it was. This one we're talking about where does Copilot get its data from. Again I've put in kind of Microsoft graph which is essentially its foundation but Microsoft recently released its Work IQ and I just want to take an opportunity to really work through this delivery understand for you guys how best Copilot manages stuff where does it get its information from how it works.

So, I want to give you guys an opportunity to understand the use of Microsoft Graph, its permissions, and how Copilot sees your organizational information. And hopefully, this is going to be our hot topic. It gives you guys a bit of an understanding of where things sit. So, the first thing I want to cover is why this matters. It sounds really cheesy and probably self-explanatory, I guess, but we don't always see the back end. We well in fact we never see the back end of how Copilot works, how it looks for information, what it's sourcing and actually this information is critical for us.

We need to be able to see how that function is what or at least understand it not necessarily see it but understand how it's using that information. The best way that I find to describe this in reality Copilot is tightly governed. So, it's managed by graphs. It's managed by permissions. And I think understanding it's key to explaining AI safety or at least for you guys as MSPs or anyone watching this to be able to explain how AI keeps itself safe particularly Copilot when within a system like this.

So, things kind of highlighting and being aware of graph is essentially the central nervous system of the Microsoft 365 environment. The APIs that aggregate all the organizational data, the permissions and the context, everything you do and change within 365 essentially triggers an API or a change and it runs on the Microsoft graph back end, whether that's using the UI prompts within Copilot. It's essentially the gatekeeper that enforces the access rules. How do we process that?

You would also kind of see it as the translator between your kind of Microsoft's apps and the AI conversion or the generation of that prompt information. That safety layer essentially prevents Copilot from seeing anything a user shouldn't in other words be able to see. So graph is that brain Copilot is the assistant that talks to the brain would probably be the best way to describe that. So, Microsoft graph the brain behind Copilot the image on the bottom right hand corner with the incorrect eye under Copilot kind of clearly states that this has been an image generated by Copilot itself believe it or not but talking around how things run graph is Microsoft's unified data API as I said before it's that central nervous system it runs and connects all of those 365 services the data collection is key here.

It holds the signals from emails, files, meetings, everything that you do, touch or run within Microsoft 365 is connected in some way using Microsoft Graph. And with Work IQ, which is this new gen process within Copilot, it builds those layers of conversations and creates those connections and strings together so that when we're looking for information and enrichment, let me just move myself across. We've got the ability to kind of see the information that's being presented.

Well, actually, there you go. Enrichment is point number three is what I was going to say. And graph enriches the prompts from the with the relevant context is foundationally correct. What this means Copilot never bypasses access controls. And I think that's really vital. I said this in episode one and I'll reiterate this throughout the series. It's never going to bypass what you have access to. So if you have messy permissions, then it's going to have access to that messy permissions.

It's going to be able to source anything. It looks at your information fundamentally through Microsoft Graph, through your permissions, through the access controls granted. It's going to highlight areas of vulnerability. That's how Copilot works. That's how AI works. When you give it kind of permissions to something, it's going to be seeing that information. And it's permission enforced. So graph enforces permissions at every single query. Whenever you make a request, it's going to go through graph or Work IQ and it's going to identify those changes and it's going to provide those remediations.

So talking about those data sources in a bit more layman's terms, we're talking things like Outlook, SharePoint, Calendars, Planner or To-Do, even Microsoft Loop, Viva, OneDrive, and Teams. All those kind of Microsoft 365 productivity applications that we use, Copiler has access to, particularly the ones that you're using. So it will have those permissions based on the information you're presenting to it. And it sits on that foundational layer in graph with its own little application, the Copilot app.

And then we when we're sifting through information, it's going back in, it's going through those that information piece, identifying areas and providing us with those outcomes or that context that we've requested. And I think a critical point which I've put at the bottom, Copilot uses whatever the user already has access to inside these sources. Nothing more and nothing less. It uses what we give it access to. Copilot's powerful to proportion proportional to how well your organization uses these tools.

It really focuses on what and how we're using it. If we're using Copilot, but all our day-to-day running sits in Google, for example, it's just mental. Doesn't work as you'd expect. So, it's things like this that you have to be aware of. It's not going to be reading Google Suite Docs if it's not plugged into it. And it's not. Is using the Microsoft 365 tenant and it's the information in there. So going back through the grounding layer, I kind of mentioned how this works previously in the last slide deck, but I want to reiterate the kind of filtering that it uses.

So again our user prompt using Work IQ Microsoft graph to sift through the relevant data the permissions utilizing our filter contexting and structuring the data that's relevant ready to be processed back throws it into the LLM for processing so working through a large language model that kind of generates a better genai context or reply goes through that security filter so it will go and validate and verify everything it's meant to be sending this is where data loss prevention comes in.

Is there anything sensitive it's picked up that needs to be restricted before it goes to us? Really good question. The answer is that security filter is going to look at that and monitor it. And I want to focus on the key takeaway here. Copilot doesn't roam your tenant freely. It just it just doesn't. It uses the information and permissions you have. It retrieves just enough context to answer the prompt all whilst respecting the permissions that are in that environment.

So don't think and at any point is it going above and beyond is it doing too much? The answer is no, it's not. It's using what you have permissions to already. So, what Copilot cannot see, I think, is something that people need to be aware of.

It cannot read content the user cannot access. I've said this again and I'll always repeat it. It cannot elevate permissions. It can't take ownership of a tenant and start controlling it. It can't view blocked SharePoint libraries. If you don't have permissions to it, guess what? It's not going to see it. It can't bypass data loss prevention policies. And this is really important to be aware of. It's not going to take ownership of a sensitivity in piece of information and produce it to you if a DLP policy is blocking it.

It follows that security filter model. It cannot read private teams chats. It's not part of it's not part of already. So if you are having that communication, it's going to see it. But if there's a conversation between your line manager and another colleague, it's not going to be able to sift through that information. It's protected. It's there. It's grounded. It's based on our own visibility.

It cannot see restricted sensitivity labeling and I like to include this particularly around data loss prevention and sensitivity labeling where we're at a governance point or a road map around data governance as MSPs in fact as a world using technology where data has always been important data is being filtered manipulated whether that's cookies are being harvested from websites data is completely relevant to the here and now and we as businesses need to make sure that we're making sure it's protected where possible and AI is opening up a can of worms of data permissions over permissioning.

It's showing us and highlighting areas that we're vulnerable and we need to make sure we're acting on it and sensitivity labeling data loss prevention is a serious step towards protecting data from day one. We want to focus on access controls. So delivering access controls restricts those areas that Copilot can access AI can access. Sharing settings absolutely business critical file and folder permissions external sharing. Should we be sharing externally? Probably not. And if we are we need to make sure we put the right controls in.

Entitlement management access packages can really help to define how people access the organizational data. Sensitivity labeling is another hot topic. Again I've said this before. Or we sit with what used to be as your information protection policies. Now we can customize strict labeling whether that's by group, department, label identity, general, public, confidential, highly confidential. We can start those customizations, build out the right model and push this to customers. And we could do all of that utilizing inforcer.

We have that capability to push these policies out, predefine them in advance. Same with Purview data loss prevention. We need to have policies and rules in place to classify and protect data. It's basic 101 now. We have got to make sure data governance is in place. And as MSPs, managed service providers or managed intelligence providers, if you read in other terms, we need to be on top of that game, ahead of the game more than anything.

How do we get to that point? What do steps do we need to take? And then finally, conditional access, identity based policies is absolutely crucial. Now we need to make sure that we are implementing strong foundations on conditions to access the environment. Whether that's conditions to access data, conditions to access devices, conditional access policies allow us to fine-tune and refine those processes. It's identity based policies delivery. So data types Copilot interacts with I think is really important.

It's structured data and unstructured data and it's hybrid data. It's information throughout the environment. So when we talk about structured, we're talking SharePoint lists and tables, specific tasks, predefined information in calendars.

If it's connected to things like ServiceNow, which guess what it can be, then it's going to be focusing on that structured data, ordered information, what's unstructured, the content within documents, the content within emails, even the content within PDF context, content, teams, meeting transcripts. This is classed as unstructured data. It's data that isn't defined with metadata columns, predefined outputs. This is all information that's been fine-tuned and drafted by ourselves in a document. It can read that information.

It can interact with it. It can manipulate it if you request it to. And then you've got your hybrid data. So these are the things that sit outside of unstructured and structured data like SharePoint pages. It's able to read a news post in SharePoint. Fundamentally and the thing I really like is loop components. It can read, it can collect loop information. This is like unstructured, structured. It's all over the place. Loop, I have no idea.

It's stored content types. I actually have no idea how loop really truly interacts. It'll be interesting to see people's comments on that. And then the other thing is Power BI insights being able to read and manipulate the information provided in a Power BI report or a dashboard. We can plug into these things using Copilot and generate some really useful content.

So I want to talk about tenant boundary rules here. Copilot does not use external tenant data, other tenant data, Microsoft staff data, public internet content unless using Bing as said already and GPT training data sets. There are five fundamentals it doesn't do. It doesn't train on the information that we have and then reuse it elsewhere. It's locked down to our tenant. There's rules. Microsoft have built this to be rulebased. It's going to be refined to our platform.

It's not going to be free to roam elsewhere to provide information elsewhere. It's locked down to the internal data that we use. And it's really important that we understand that model of delivery. So permissions model I think is a hot topic. Copilot safety model is based entirely on user permissions. If I can see it, Copilot can see it. If I can't see it, Copilot can't see it. Again, reiterating these same models, it's what we have access to already.

Real world examples of this, limited SharePoint access. So, user with access to three SharePoint sites and the restricted folders like HR is not going to see that information. And we can take one step further. If we're not restructuring SharePoint to make sure our data is safe on a permissioning level, we can enhance this with data loss prevention. And so block AI actions, Copilot will refuse to request entirely. So we can go in and be creating trainable classifiers around HR information and we can lock that down utilizing data loss prevention policies.

And again, as MSPs, we can predefine these in advance in our baseline for inforcer in our baseline tenant and then push this out to our customers.

And I think that adds huge value when we talk about structuring data. So I want to talk about examples, use cases and data flow. Summarizing a team's channel is really useful. It's a function that we should be using every day. The ability to kind of summarize information going on in a team's chat or a channel really helps us as individuals to be efficient when we've got other co colleagues, sorry, discussing things with one another. Like yesterday is a good example for me.

My solutions architect team that I work with were chatting away and they'd probably done well over 200 messages in the space of two hours and I just got straight into Copart and said summarize the information. It turned out most of it was jargon so I could ignore most of it which was great. But there's other elements that it benefits on pulling information on meeting transcripts and notes really useful if you're running late to a meeting and you need to catch up or you've had to drop from the meeting early and focus on something else.

This is how people can use Copilot. So, we're starting to become efficient, starting to streamline the information that we're getting, what's relevant, what's not. List me five bullet points from this meeting that are relevant to me and my job role. Copilot's going to look at what your job title is, your job role, and identify and then pull that information out based on that meeting. So, if it's a all hands meeting, this produces some really good information for you.

You might miss here and there a few things, but the bulk of what you're after is absolutely there for you to see. Summarizing email threads, I think, is one of the biggest things I find really useful. When I'm being thrown into a meeting to have a conversation, I'm pulled an email thread that's perhaps got 60 messages on. Summarizing it really helps me to get an understanding, get on the front foot of being able to analyze that information without having to sit for an hour reading through these conversations.

Sometimes you do depending on the severity and sometimes you don't. Sometimes it's just that summarize this email and off we go. And then the other thing I always use it for is drafting proposals. So I've said this before, you can link it to multiple contexts and information. You can draft Copilot agents that point to a repository in SharePoint as an example around change controls and all that information. And you can start getting Copilot to draft new change controls for new information you're producing.

As an MSP, we're constantly making changes to customers environments. Change controls are inevitable depending on the industry and regulations you have to adhere to like 27,0001. We would be creating those change controls. It's a requirement and as a result we can create a bespoke agent that specializes in just helping for us to reuse and recreate change control documentation following the same methodology. All that can be done with Copilot agents using what we have already with Copilot itself for 365.

We don't have to re redesign and redo everything we've done in the past. We can utilize what's already available. So talking about data hygiene, I think that's really important.

We need to make sure that our data is not messy as less messy as possible. Copilot will always become messy if we have unstructured teams channels. Old SharePoint silos that aren't used anymore. But guess what? We have access to files with poor naming conventions, sensitivity, sensitive data in the wrong place, and excessive ex everyone except external permissions. These things really highlight kind of areas of concern and we need to be more defined. Groupbased permission management renaming to structure data that makes sense not just for us when we're looking at it in SharePoint but just generally sensitivity data in the wrong place is absolutely critical you know oversharing and moving data into wrong locations we can remediate those with sensitivity labels but unstructuring this information really makes Copilot's life harder and in turn makes our life harder so we want our data to be governed and that allows Copilot to become more powerful for us.

So that is the strong sensitivity labeling. It is clean permission structures, clear folder hierarchies, active life cycle management. Microsoft have this available for us to use. We absolutely should be using it and regular governance reviews.

We're getting to the point where governance is hot topic and being able to run a journey with a customer around data governance, data hygiene on a regular basis allows us to be at the front foot for proactive support, front foot for being proactive when we talk to our customers. And that shows value as an MSP as to what we're giving to customers. So I want to talk about why this matters specifically for MSPs. So where can we almost monetize and take opportunities as managed service providers or managed intelligence providers?

So data readiness assessments evaluate client environments for Copilot deployment readiness available in inforcer. Get that running now. Get all your customers on board and start running that readiness assessment. Permission cleanup projects. When we look at over permissioning and we start that conversation around Copilot, we can start generating conversations and talking points around being ready for a Copilot with the readiness assessment. And we've got project opportunities for project cleanup, you know, tidiness, restructuring and rearchitecting teams and SharePoint.

So we've, you know, fundamentally it's structured appropriately to make it clean. Another project, another opportunity. Standardizing sensitivity labels for deployment, data loss prevention, conditional access policies, all of these things start to drive commercial opportunities for us as businesses. And this is what's going to help us to kind of refine and be on the front foot of AI delivery, taking that initiative and driving positive change. It starts with us as MSPs and it gives us the opportunity to drive change.

Copilot exposes weak governance. The MSPs can already fix and we can utilize and leverage inforcer to be able to do those Copilot improvements. So data loss prevention foundations sensitivity label foundations shareoint SharePoint sharing permissions all of these things are buttons and tick boxes you can do with inforcer on mass deployments we should be utilizing a multi-tenant management solution to allow us to do so. So I want to kind of cover areas that we're already looking at the advantage of I've already said with inforcer just to kind of cover really quickly permission drift.

We've got drift detection, the ability to kind of identify changes made to conditional access policies, generally data loss prevention policies. We can now push out on mass predefined them GDPR enhanced UK financial data. All these things we can pre-build in a baseline, sell them as a service to our customers and push them out to our customer environments with just a few clicks. Opening, sharing risks, being able to see the differences and the wrong permissions outlined.

We're now adapting and growing that SharePoint understanding and that's measurable and so important when we talk about AI readiness for any business. Conditional access point access alignments I think is just a huge area. We spend our lives with renamed conditional access policies which drive me mad. My OCD levels through the roof. We need to get on top of it. We want standardization. And I say this whenever I talk to any customer, whenever I'm doing a pre-sales conversation, your Microsoft 365 tenant doesn't have to be unique, but it does have to be secure.

And everyone has the exact same secure measures. It's just written differently, designed differently, but the foundations of security is exactly the same. And that should be standardized. As an MSP building your best practice of best security measures you can push out in one go whenever you take on a new customer prospect opportunities. The other thing is your sensitivity labeling and gaps. We've got the opportunity to get on the forefront of this, fill those gaps, resolve those issues.

This is where we're going to start seeing that string of information. AI readiness scoring, another hot topic, enabling quantifiable AI readiness assessments. And then client reports produces client-friendly data governance reports. inforcer has got some incredible reports for alignments, assessments, prospecting opportunities. Those reports really help to drive our sales team, pre-sales teams to have those conversations around Copilot, next steps. And that's it, guys. Hopefully, this little miniseries has been mini episode has been really useful.

We're going to start on our next one a bit later on. But feel free to stay tuned for the next episode. A few more to go. Hopefully you're finding these really useful.