CD Consulting R&D

KB-AI · Entry 9 · Published 2026-10-26 · Version 1

Integrated or independent? Microsoft 365 Copilot against build-your-own agents for enterprise knowledge (October 2026)

What Microsoft 365 Copilot's integration with SharePoint, OneDrive, Teams and Outlook really saves a company, what it does not, what an independent agent must build in its place, and what the independent vendors now offer through connectors and the Model Context Protocol: a comparison on retrieval, permissions, sovereignty and price, as of October 2026.

If you are deciding whether to let Copilot answer from your SharePoint, to build an agent around your own search index, or to do both, and want to know what Microsoft's integration really saves you, what it does not, and what the independent vendors now offer in its place, this is the comparison, with prices and the conditions under which each route works.

Method: desk research on 26 October 2026 across Microsoft Learn and Microsoft pricing pages, the documentation of Anthropic, Google, Mistral, OpenAI and vector-database providers, the Model Context Protocol specification and analyst publications; 56 sources typed and dated; figures read only through search summaries are marked to be confirmed; limits declared in section 12. Nothing was installed or executed.

Contents

1. 1. The comparison on one page
2. 2. What the question is, and what it is not
3. 3. What Microsoft provides natively
4. 4. What the native advantage costs
5. 5. What an independent agent must build
6. 6. What the independent vendors now provide
7. 7. The comparison grid
8. 8. Sovereignty and the law
9. 9. Money, on one table
10. 10. What the analysts say, and how far to trust it
11. 11. The decision, in five questions
12. 12. Limits of this entry
13. 13. Sources

1. The comparison on one page

Six findings organise this entry.

1. Microsoft's advantage is real and it is not the model. For content already in Microsoft 365, Copilot grounds its answers on Microsoft Graph and a tenant-wide "semantic index" that Microsoft builds and refreshes without any pipeline on the company's side, and it only returns what the user already has permission to see [S1][S2][S6]. An independent agent must reproduce that plumbing: ingestion, chunking, embeddings, a vector store, hybrid search, and the propagation of access rights to every fragment [S31][S33].
2. The advantage has a price that is not on the invoice. Copilot reveals existing over-sharing — "indexing data doesn't change access permissions to content" [S2] — and Microsoft's own deployment guidance makes remediating permissions the first of three pillars before rollout [S22]. Gartner's 2025 survey found the effort to mitigate over-sharing "significantly greater" than expected, with 40% of respondents in pilot and few scaling [S25][S26].
3. The "disconnected inputs" picture of independent agents is out of date. Since 2025-2026 the independent vendors reach Microsoft 365 through connectors that query SharePoint, OneDrive, Outlook and Teams live, with the user's rights, without building an index: Claude's Microsoft 365 connector (read by default, write only when an administrator enables it) [S34][S35], Gemini's federated SharePoint connector [S38], ChatGPT's SharePoint connectors (to be confirmed) [S39]. The Model Context Protocol, donated to a foundation in December 2025, is becoming the common plug [S36][S37].
4. Microsoft itself has become multi-model and multi-protocol. Anthropic models were enabled by default in many Copilot experiences in January 2026 (not in the EU, EFTA and the UK, where an administrator must opt in, and outside the EU Data Boundary) [S11][S41]; Copilot Studio agents extend through MCP; Microsoft publishes MCP servers of its own [S20][S40]. "Copilot or an independent agent" is no longer a choice between two closed worlds.
5. Where a custom pipeline still wins is unchanged and narrower: sources that Microsoft 365 does not hold (databases, line-of-business systems, code, large document archives), content whose chunking and cleaning must be controlled, sovereign or disconnected deployments, and cost at very large scale. Even there, Microsoft's own Azure AI Search offers the managed middle road, with SharePoint access-control lists still in preview [S33].
6. The honest verdict is hybrid, and the hybrid has changed shape. In 2024 "hybrid" meant Copilot for office work plus a second index for the rest. In October 2026 it means Copilot for the Microsoft estate, independent agents reaching that same estate through connectors when their models or their governance are preferred, and a custom index only for what lives outside. No published case study gives a verified cost for either half [S45].

2. What the question is, and what it is not

Retrieval-augmented generation (RAG) is the technique by which a language model answers from documents it did not learn in training: the relevant passages are retrieved first and placed in the prompt, so that the answer is "grounded". Every enterprise assistant, integrated or independent, does this. The comparison in this entry is therefore not about models — Copilot routes between several, including Anthropic's since 2026 [S1][S11] — but about who builds and operates the retrieval layer, who enforces permissions, and who pays for what.

The scope is the consolidation of unstructured enterprise knowledge: documents, mail, chats, meetings, pages. Structured data (ERP, CRM, warehouses) and software code are mentioned where they change the comparison, but they are not its object. The entry does not cover prompt injection, which has its own treatment in this chapter, beyond noting that the one documented incident of the period, EchoLeak, hit the integrated product [S32][S44].

3. What Microsoft provides natively

The request flow. Microsoft describes five steps: the prompt leaves a Microsoft 365 application for Copilot; Copilot queries Microsoft Graph and the semantic index to ground it; the enriched prompt goes to the language model; the answer returns; Copilot goes back to Graph and the index for post-processing, then returns the answer and any application command [S2]. Nothing in this chain is installed, configured or hosted by the customer.

The semantic index. It is "generated from the content of Microsoft Graph", lexical and vectorial at once, and lets an organisation "search through billions of vectors" [S2]. It covers Word, PowerPoint, PDF, OneNote and SharePoint pages up to 512 MB per file, plus mail at the user level and connector data at the tenant level; shared and delegated mailboxes are not covered. New SharePoint files accessible to at least two people are indexed daily, modified files immediately; every Copilot customer has a tenant-level index, and the page is inconsistent about whether the user-level index exists today or "in time" [S2]. It cannot be disabled; a site can be excluded through Purview data-loss prevention or by removing it from search results. The index lives in the tenant's region, inside the EU Data Boundary for European tenants, and does not count against storage quotas [S2]. Microsoft does not publish the embedding model, the vector dimension or the chunk size.

Work IQ. Since 2026 the term is official: "a workplace intelligence layer that enables agents to access and reason over organizational and business data, context, and tools", with four components (chat, context, tools, workspaces) exposed through agent-to-agent calls, a remote MCP server and a REST API [S3]. Its APIs reached general availability on 16 June 2026, billed by usage for third-party or custom agents, and included for Copilot licence holders [S3][S4]. Microsoft does not describe how Work IQ and the semantic index combine internally.

Permissions. "Microsoft Copilot only surfaces organizational data to which individual users have at least view permissions" [S6]; the index "honors the user identity-based access boundary" and checks access in real time [S2][S6]. Encrypted content requires the user to hold the EXTRACT and VIEW rights; sensitivity labels are inherited by generated content; prompts, responses and referenced content are audited and reachable by eDiscovery [S5]. Connector items carry their own access-control list from the source system [S12].

Connectors and agents. More than a hundred ready-made Copilot connectors cover SAP, Salesforce, ServiceNow, Confluence, network shares and others; federated connectors use MCP and require a Copilot licence per user; a quota of 50 million items per tenant and 5 million per connection is reported but could only be read through a search summary [S12][S13][S14]. Copilot Studio builds agents on knowledge sources (SharePoint, uploaded files resynchronised every four to six hours, connectors, MCP servers) and bills them in Copilot Credits: 2 for a generative answer, 10 for grounding on the tenant graph, free for licensed employees within fair-use limits, with agents switched off at 125% of prepaid capacity [S15][S17][S18][S19][S20]; the extensibility documentation of late 2026 also speaks of "plugins" for Copilot itself, without a published rule for choosing between declarative agents, custom-engine agents and plugins [S24]. Copilot Tuning, the fine-tuning of models on company data, was still in early access in June 2026 [S21].

4. What the native advantage costs

Over-sharing. The cleanest statement is Microsoft's: indexing does not change permissions [S2]. What Copilot changes is the efficiency of search. A budget site that was open to the wrong people was always open; Copilot makes the fact visible in an answer [S23]. Microsoft's "Secure and governed data foundation" guidance therefore puts "remediate oversharing" first, before guardrails and regulatory compliance, with Purview and SharePoint Advanced Management as the tools; the latter is stated to be included in the Copilot licence [S22]. The stopgap of the first wave, Restricted SharePoint Search — an allow-list of at most one hundred sites plus the user's last two thousand items — "isn't a security boundary", and new enablement has been blocked since 31 July 2026 in favour of Restricted Content Discovery [S23].

Gartner's 2025 survey of 187 IT leaders gives the measure of the problem: 40% of organisations were in pilot, 5% of those that had finished a pilot moved to a wider deployment, return on investment was "elusive", and the effort to mitigate over-sharing was "significantly greater" than anticipated [S25][S26]. No Microsoft page quantifies how many customers are affected or how long the clean-up takes.

Content hygiene. The index "works best when the connector content is text rich" [S2]; stale, duplicated and unmaintained documents are retrieved like any other. The note that prompted this entry calls this the "garbage in, garbage out" trap, and it is the one point on which the integrated and the custom route do not differ: a custom pipeline lets engineers tune chunking and filtering, but it does not clean the documents either.

Boundaries. Copilot is covered by Microsoft's data-residency commitments; EU traffic "stays within the EU Data Boundary while worldwide traffic can be sent to the EU and other countries or regions" in periods of high load [S6]. Two nuances matter for a European buyer. First, "models provided by Anthropic as a subprocessor are currently excluded from the EU Data Boundary" and are off by default in the EU, EFTA and the UK [S6][S11]. Second, the "in-country processing" announced in November 2025 for fifteen countries was redefined in April 2026 as "local data inferencing", with a revised calendar — end of 2026 for Australia, India, the United Arab Emirates, the United Kingdom and the United States, 2027 for Canada, 2028 for Japan — and regional inference for the EU and EFTA [S10]. Belgium was never on the list; the relevant guarantee for a Belgian tenant is the EU Data Boundary, not an in-country promise. Prompts, responses and Graph data are not used to train foundation models, and Copilot has opted out of Azure OpenAI's abuse monitoring [S6].

Lock-in and price. The integration is the lock. The enterprise add-on is listed at 30 dollars per user per month (read through a search summary of Microsoft's pricing page; to be confirmed) [S9]; the Copilot Business add-on for up to 300 users is listed at 21 dollars, promoted at 18 dollars for the first year between 1 July and 31 December 2026, or 25.20 dollars monthly [S7][S8]; Microsoft 365 E7, which bundles Copilot, is reported at 99 dollars [S9][S30]. Work IQ APIs, Cowork and Code are billed by usage on top [S1][S3].

5. What an independent agent must build

The independent route starts from a model and an empty retrieval layer. Microsoft's own Azure AI Search documentation lists the bricks, because it sells them: an indexer, a text-split skill, integrated vectorisation, an index, retrieval, and "agentic retrieval" over "knowledge sources" [S33]. In full, a company must choose or build:

1. ingestion — connectors to each source, parsers per format (PDF, Office, HTML, mail, images with OCR), handling of updates and deletions, synchronisation of access rights;
2. chunking that respects headings, tables and lists, with overlap;
3. vectorisation — an embedding model, and its versioning, since changing the model means re-indexing;
4. a vector store or search engine — Azure AI Search, Pinecone, Weaviate, Qdrant, Elasticsearch, pgvector;
5. hybrid search (lexical and vector), re-ranking, citations and grounding;
6. security — identity filters per document;
7. evaluation — reference question sets, recall and precision, regression tracking;
8. operation — logging, cost, freshness monitoring.

Two of these deserve a closer look, because they are where the integrated product earns its keep.

Permissions at the fragment. When a skillset splits documents into chunks, the access-control fields must be projected onto every chunk, otherwise permission filtering does not apply to fragments [S33]. Azure AI Search offers four ways to filter by document: string-comparison security filters, available in production; POSIX ACLs and role-based scopes for Azure storage, in preview; Purview sensitivity labels, in preview; and SharePoint ACLs, in preview, where changes inherited from a site, library or folder require an explicit refresh [S33]. The vendor's own warning is that preview features are not recommended for production workloads. There is also a structural point, not an incident: an index copies permissions at indexing time, so a right withdrawn in SharePoint stays honoured by the index until the next refresh. Live connectors (section 6), which query Microsoft 365 at request time with the user's rights, do not have that window.

Cost and delay. No serious published study gives the average cost or lead time of an enterprise RAG project; the figures in circulation come from consultancies' blogs without method and are not reproduced here [S46]. What is published is the price of the infrastructure. Pinecone lists a 20-dollar monthly Builder tier, a Standard tier with a 50-dollar minimum, an Enterprise tier with a 500-dollar minimum, storage at 0.33 dollars per gigabyte per month and reads at 16 to 18 dollars per million on the Standard tier [S42]; pgvector costs the Postgres instance it runs on; the rest varies with traffic, not with the number of vectors. The causes of failure that recur in the literature — data not ready, permissions leaking, chunking that destroys structure, no systematic evaluation — are the same that Gartner's press releases cite when predicting that 60% of AI projects without AI-ready data would be abandoned by 2026 and that over 40% of agentic AI projects would be cancelled by the end of 2027 (both to be confirmed on the original releases, which could not be opened) [S47][S48].

6. What the independent vendors now provide

The sharpest change since the question was first asked is that the independent vendors no longer ask the company to rebuild Microsoft 365 in a vector store. They connect to it.

Claude. Anthropic's Microsoft 365 connector is available on every Claude plan and searches SharePoint, OneDrive, Outlook mail and calendar, and the Teams chats and channels the user belongs to; the page is dated 8 September 2026 [S34]. Reading is the default; writing (sending mail, creating files, posting in Teams) exists only if an administrator enables each write tool, with user confirmation required for Teams [S34][S35]. The prerequisites are the ones a security team would expect: a Claude organisation owner enables the connector on Team and Enterprise plans, and in every tenant a Microsoft Entra global administrator grants a tenant-wide consent to a documented list of delegated Graph permissions (Files.Read.All, Sites.Read.All, Mail.Read and others) [S35]. "The connector only retrieves data on-demand during active queries and doesn't cache file content"; calls appear in the Microsoft 365 audit log and Microsoft's data-loss prevention applies [S35]. The limits are equally documented: SharePoint search is tenant-wide, per-site permissions are not supported, and the default permission set is broad [S35].

Gemini. Google's SharePoint Online connector works in two modes: federation, which queries SharePoint live with the user's rights, returns around twenty results and cannot search inside attachments; or ingestion, which indexes the content and is recommended for complex PDFs, with data locations selectable between global, US and EU [S38].

ChatGPT and others. OpenAI documents SharePoint connectors and synchronised "company knowledge" for its business plans, but its help and pricing pages returned access errors on the day of consultation; the details are therefore to be confirmed [S39]. Mistral's enterprise offer lists private deployments and an opt-out from training, with connectors by tier not published on the page read [S43].

The common plug. The Model Context Protocol was donated by Anthropic to the Agentic AI Foundation under the Linux Foundation on 9 December 2025, with Block and OpenAI as co-founders; Anthropic reported more than ten thousand public MCP servers and over 97 million monthly SDK downloads [S36]. Microsoft publishes MCP servers for Azure, for Dynamics 365 and for its documentation, and Copilot Studio agents consume MCP servers [S20][S40]. One line of the specification should be read by every architect: "Authorization is OPTIONAL for MCP implementations" — when present it is built on OAuth 2.1, resource indicators and audience-bound tokens, but the security of a connector depends on its implementation, not on the protocol [S37].

And Microsoft went the other way. Anthropic models were enabled by default in many Microsoft 365 Copilot experiences on 7 January 2026, routed per request, with administrators able to opt out — and, in the EU, EFTA and the UK, required to opt in [S11][S41]. The company that chooses Copilot may therefore be using Claude inside it; the company that chooses Claude may read SharePoint through it. The two routes share models and share data; what separates them is who runs the retrieval layer and whose governance applies.

7. The comparison grid

8. Sovereignty and the law

For a European company three facts frame the choice. Microsoft's EU Data Boundary covers Copilot, with the overflow and the Anthropic exclusion noted above [S6][S11], and Microsoft announced in June 2025 and extended in October 2026 a set of "sovereign" options — external oversight of engineer access, customer-managed keys, a sovereign private cloud on Azure Local — whose October details come from the press and are to be confirmed [S27][S28]. Mistral, with OVHcloud, Scaleway and T-Systems, offers models and hosting under European control, and its open-weight models can be self-hosted, some under a commercial licence [S43]. A fully disconnected deployment is possible in bricks — an open model, a self-hosted vector store — but no vendor documentation read for this entry describes an end-to-end air-gapped document agent; the assembly and the evaluation remain the company's.

On the legal side this entry cites texts, not advice. The General Data Protection Regulation (Regulation 2016/679) governs the processing, the processor contract, security and impact assessment, and transfers outside the Union [S50]; the NIS2 directive (2022/2555) adds supply-chain security for the entities it covers [S51]; the AI Act (Regulation 2024/1689) applies to the company as a deployer, at minimum through AI literacy and transparency towards the people who interact with the system, while "high-risk" classification depends on the use case, not on the retrieval technique [S52]. A 2026 omnibus regulation reported by law firms to have postponed the high-risk obligations to December 2027 could not be checked on the Official Journal for this entry and is to be confirmed [S53]. None of this favours one route over the other; all of it favours the route whose data flows the company can describe.

9. Money, on one table

Two remarks. Copilot's price is a seat; the independent vendors increasingly combine a seat with metered usage, so the total depends on adoption. And neither column includes the cost that both routes share and that no source measures: the clean-up of rights and content.

10. What the analysts say, and how far to trust it

Gartner's 2025 Microsoft 365 and Copilot survey (187 IT leaders): 40% in pilot, 5% of completed pilots scaling, "elusive" return, over-sharing mitigation "significantly greater" than expected [S25][S26]. Forrester's Total Economic Impact study of March 2025, commissioned by Microsoft, reports a 116% three-year return for a composite organisation of 25,000 employees [S29]. The United Kingdom's government trial with twenty thousand licences reported 26 minutes saved per day, self-declared [S54]. On the build side, MIT's "GenAI Divide" report of 2025 is cited for a purchase success rate of about two thirds against about one third for internal builds, with figures that vary between summaries and a primary report not read for this entry [S55]; Gartner predicted in September 2026 that 70% of enterprises would abandon agentic AI built by vendors' forward-deployed engineers by 2028, in a release that could not be opened [S56]. The sources point the same way — value depends on the data, the governance and the cost, rarely on the model — but none of them is an independent, reproducible measure of a document agent built against one bought, and the most favourable Copilot figure is vendor-commissioned.

11. The decision, in five questions

1. Where does the knowledge live? If it is in Microsoft 365, Copilot's retrieval layer already exists and an independent agent can reach the same content through a connector; building a second index for it is the one option that is now hard to justify. If it is elsewhere — databases, line-of-business systems, archives, code — some pipeline will have to be built or bought, and Microsoft's connectors compete with everyone else's.
2. How clean are the rights? Both routes reveal over-sharing; only the custom index adds a copy of the rights that can go stale. Whatever the route, the remediation comes first, and its cost is the one the surveys say is underestimated.
3. Which models, under whose governance? Copilot routes among models, including Anthropic's, under Microsoft's terms; an independent agent uses its vendor's models under the vendor's terms, with the company's own harness around it. The choice is about whose controls the company wants to configure, audit and answer for.
4. Where may the data go? The EU Data Boundary, the Anthropic exclusion and the revised local-inference calendar define Copilot's answer; the independent vendors' answers are less documented, and a sovereign or disconnected deployment is only available to those who build.
5. Who will build and run it? The integrated route needs administrators and a clean estate; the connector route needs the same plus a vendor relationship; the custom route needs engineers, an evaluation discipline and an operations budget, and the analysts' failure rates are about that route.

The verdict of the note that prompted this entry — hybrid — stands, with one correction. The hybrid of 2026 is not "Copilot plus a second index". It is Copilot for the Microsoft estate, independent agents reaching that estate through connectors when their models or their governance are preferred, and a custom index only for what Microsoft 365 does not hold. The company that reinvents the wheel today is the one that re-indexes its SharePoint.

12. Limits of this entry

• Prices are list prices in dollars, read on 26 October 2026; euro prices were not found; the 30-dollar enterprise price, the 99-dollar E7 price, the Copilot Studio package price and the connector quotas were read through search summaries of Microsoft pages and are marked to be confirmed. OpenAI's help, privacy and pricing pages returned access errors; Gemini Enterprise prices were not published on the page read.
• Analyst sources are typed as such; Gartner's documents are paywalled and were read through press relays; Forrester's study is commissioned by Microsoft; MIT's report was not read in its primary form.
• No published, verifiable cost or lead time exists for an enterprise RAG project or for a hybrid architecture; the entry says so rather than citing blogs.
• The legal section cites regulations by number and is not legal advice; the 2026 omnibus regulation is reported by law firms and was not checked on the Official Journal.
• Microsoft's documentation changes monthly; page dates are given in section 13. Work IQ's internal articulation with the semantic index, the embedding model and the chunk size are not published. The entry is based on documentation, not on a test of either product.

13. Sources

All sources consulted on 26 October 2026. Page dates are those shown by the publisher.

1. S1 Official. Microsoft Learn, "What is Microsoft Copilot?", page dated 23 October 2026. https://learn.microsoft.com/en-us/microsoft-365/copilot/copilot-overview
2. S2 Official. Microsoft Learn, "Semantic indexing for Microsoft Copilot", page dated 23 April 2026, updated 18 August 2026. https://learn.microsoft.com/en-us/microsoftsearch/semantic-index-for-copilot
3. S3 Official. Microsoft Learn, "Work IQ overview", page dated 25 September 2026. https://learn.microsoft.com/en-us/microsoft-365/copilot/extensibility/work-iq/
4. S4 Official. Microsoft 365 Blog, "Announcing the new Work IQ APIs", 2 June 2026. https://www.microsoft.com/en-us/microsoft-365/blog/2026/06/02/announcing-the-new-work-iq-apis/
5. S5 Official. Microsoft Learn, "Microsoft Copilot data protection architecture", page dated 7 April 2026, updated 21 October 2026. https://learn.microsoft.com/en-us/microsoft-365/copilot/architecture-data-protection-auditing
6. S6 Official. Microsoft Learn, "Data, privacy, and security for Microsoft Copilot", page dated 9 July 2026. https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-privacy
7. S7 Official. Microsoft Learn, "License options for Microsoft Copilot", page dated 19 October 2026. https://learn.microsoft.com/en-us/microsoft-365/copilot/copilot-license-options
8. S8 Official. Microsoft, Microsoft 365 Copilot Business pricing page. https://www.microsoft.com/en-us/microsoft-365-copilot/business
9. S9 Official (read through a search summary; to be confirmed). Microsoft, Microsoft 365 enterprise plans and pricing. https://www.microsoft.com/en-us/microsoft-365/enterprise/microsoft-365-plans-and-pricing
10. S10 Official. Microsoft Copilot Blog, "Microsoft offers in-country data processing to 15 countries", 4 November 2025, with an editorial note of 3 April 2026. https://www.microsoft.com/en-us/copilot/blog/2025/11/04/microsoft-offers-in-country-data-processing-to-15-countries/
11. S11 Official. Microsoft Learn, "Anthropic models in Microsoft online services", page dated 18 September 2026, updated 23 October 2026. https://learn.microsoft.com/en-us/microsoft-365/copilot/connect-to-ai-subprocessor
12. S12 Official. Microsoft Learn, "Copilot connectors overview", page dated 6 October 2026. https://learn.microsoft.com/en-us/microsoft-365/copilot/connectors/overview
13. S13 Official. Microsoft Learn, "Prerequisites for deploying connectors", page dated 1 June 2026. https://learn.microsoft.com/en-us/microsoft-365/copilot/connectors/prerequisites
14. S14 Official (read through a search summary; to be confirmed). Microsoft Learn, connectors licensing page. https://learn.microsoft.com/en-gb/MicrosoftSearch/licensing
15. S15 Official. Microsoft Learn, "Billing rates and management — Copilot Studio", page dated 3 August 2026. https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-messages-management
16. S16 Official (read through a search summary; to be confirmed). Microsoft, Copilot Studio pricing. https://www.microsoft.com/en-us/microsoft-365/copilot/pricing/copilot-studio
17. S17 Official. Microsoft Learn, "Knowledge sources summary — Copilot Studio", page dated 21 July 2026. https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-copilot-studio
18. S18 Official. Microsoft Learn, "Unstructured data as a knowledge source", page dated 26 August 2026. https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-unstructured-data
19. S19 Official. Microsoft Learn, "Quotas and limits — Copilot Studio", page dated 6 October 2026. https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-quotas
20. S20 Official. Microsoft Learn, "Extend your agent with Model Context Protocol", page dated 26 August 2026. https://learn.microsoft.com/en-us/microsoft-copilot-studio/agent-extend-action-mcp
21. S21 Official. Microsoft Learn, "Microsoft Copilot Tuning overview (early access preview)", page dated 3 June 2026. https://learn.microsoft.com/en-us/microsoft-365/copilot/copilot-tuning-overview
22. S22 Official. Microsoft Learn, "Secure and governed data foundation for Microsoft Copilot — foundational deployment guidance", page dated 6 May 2026. https://learn.microsoft.com/en-us/microsoft-365/copilot/secure-govern-copilot-foundational-deployment-guidance
23. S23 Official. Microsoft Learn, "Restricted SharePoint Search", page dated 6 July 2026. https://learn.microsoft.com/en-us/sharepoint/restricted-sharepoint-search
24. S24 Official. Microsoft Learn, "What is a Microsoft 365 Copilot plugin?", page dated 30 September 2026. https://learn.microsoft.com/en-us/microsoft-365/copilot/extensibility/
25. S25 Analyst (paywalled; read through press relays). Gartner, "Key insights from the 2025 Microsoft 365 and Copilot survey", 2025. https://www.gartner.com/en/documents/6548002
26. S26 Press. TechPartner News, "Gartner: Microsoft Copilot hype offset by ROI and readiness realities", 26 June 2025. https://www.techpartner.news/news/gartner-microsoft-copilot-hype-offset-by-roi-and-readiness-realities-618118
27. S27 Official. Microsoft, "Announcing comprehensive sovereign solutions empowering European organizations", 16 June 2025. https://blogs.microsoft.com/blog/2025/06/16/announcing-comprehensive-sovereign-solutions-empowering-european-organizations/
28. S28 Press (to be confirmed). CIO.com, "Microsoft expands sovereign cloud offerings as data control concerns grow", October 2026. https://www.cio.com/article/4238389/microsoft-expands-sovereign-cloud-offerings-as-data-control-concerns-grow.html
29. S29 Analyst (commissioned by Microsoft; read through a search summary). Forrester, "The Total Economic Impact of Microsoft 365 Copilot", March 2025. https://tei.forrester.com/go/microsoft/M365Copilot/
30. S30 Official (search listing only; to be confirmed). Microsoft Tech Community, "Microsoft 365 E7 and Agent 365 are now generally available", 2026. https://techcommunity.microsoft.com/
31. S31 Vendor. Pinecone and the vector-database providers cited in [S42]; and Microsoft Learn, Azure AI Search documentation cited in [S33]. (Bricks of an enterprise RAG pipeline, as named by the vendors.)
32. S32 Vendor (press-analyst). Trend Micro, "Preventing zero-click AI threats: insights from EchoLeak", 2025. https://www.trendmicro.com/en_us/research/25/g/preventing-zero-click-ai-threats-insights-from-echoleak.html
33. S33 Official. Microsoft Learn, "Document-level access control — Azure AI Search", updated 17 September 2026. https://learn.microsoft.com/en-us/azure/search/search-document-level-access-overview
34. S34 Vendor. Anthropic, "Connect to Microsoft 365", page dated 8 September 2026. https://support.claude.com/en/articles/15183774-connect-to-microsoft-365
35. S35 Vendor. Anthropic, "Microsoft 365 connector security guide". https://support.claude.com/en/articles/12684923-microsoft-365-connector-security-guide
36. S36 Vendor. Anthropic, "Donating the Model Context Protocol and establishing the Agentic AI Foundation", 9 December 2025. https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation
37. S37 Official. Model Context Protocol specification, "Authorization", revision of 28 July 2026. https://modelcontextprotocol.io/specification/latest/basic/authorization
38. S38 Vendor. Google Cloud, "Connect Microsoft SharePoint Online with data federation" (Gemini Enterprise). https://docs.cloud.google.com/gemini/enterprise/docs/connect-sharepoint-online
39. S39 Vendor (access error on the day; to be confirmed). OpenAI, "SharePoint connectors on ChatGPT". https://help.openai.com/en/articles/12143177-sharepoint-connectors-on-chatgpt
40. S40 Vendor. Microsoft, Azure SDK Blog, "Announcing Azure MCP Server 2.0"; Microsoft Dynamics 365 Blog, "Dynamics 365 Customer Service MCP Server generally available", 30 July 2026. https://devblogs.microsoft.com/azure-sdk/announcing-azure-mcp-server-2-0-stable-release/
41. S41 Press (to be confirmed). UC Today, "Microsoft 365 Copilot to enable Anthropic models by default", January 2026. https://www.uctoday.com/unified-communications/microsoft-365-copilot-to-enable-anthropic-models-by-default-what-compliance-leads-need-to-know/
42. S42 Vendor. Pinecone, "Pricing". https://www.pinecone.io/pricing/
43. S43 Vendor. Mistral AI, "Pricing". https://mistral.ai/pricing
44. S44 Press-analyst. EchoLeak, CVE-2025-32711, as described in [S32].
45. S45 Author's finding. No published case study with a verifiable cost of a hybrid Copilot-plus-independent-agent architecture was found on 26 October 2026 (searches across vendor, analyst and press sources).
46. S46 Author's finding. No published, methodologically described study of the cost or lead time of an enterprise RAG project, and no contractual EU data-residency guarantee for Claude or ChatGPT, were found on 26 October 2026.
47. S47 Analyst (access error on the day; to be confirmed). Gartner, "Lack of AI-ready data puts AI projects at risk", 26 February 2025. https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk
48. S48 Analyst (read through a search summary; to be confirmed). Gartner, "Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027", 25 June 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
49. S49 Vendor. Anthropic, Claude pricing. https://claude.com/pricing
50. S50 Official. Regulation (EU) 2016/679 (General Data Protection Regulation). https://eur-lex.europa.eu/eli/reg/2016/679/oj
51. S51 Official. Directive (EU) 2022/2555 (NIS2). https://eur-lex.europa.eu/eli/dir/2022/2555/oj
52. S52 Official. Regulation (EU) 2024/1689 (Artificial Intelligence Act). https://eur-lex.europa.eu/eli/reg/2024/1689/oj
53. S53 Press (law firms; to be confirmed on the Official Journal). Cooley, "Digital AI Omnibus delays key deadlines", 2026. https://cdp.cooley.com/digital-ai-omnibus-delays-key-deadlines-introduces-new-rules/
54. S54 Press (to be confirmed). The Register, "What will UK government workers do with an extra 26 minutes a day?", 3 June 2025. https://www.theregister.com/2025/06/03/uk_government_study_ai_time_savings
55. S55 Analyst (secondary; to be confirmed). MIT NANDA, "The GenAI Divide: State of AI in Business 2025", as reported by Fortune and Yahoo Finance. https://finance.yahoo.com/news/mit-report-95-generative-ai-105412686.html
56. S56 Analyst (access error on the day; to be confirmed). Gartner, "Gartner predicts 70% of enterprises will abandon agentic AI built by vendor forward-deployed engineering by 2028", 29 September 2026. https://www.gartner.com/en/newsroom/press-releases/2026-09-29-gartner-predicts-70-percent-of-enterprises-will-abandon-agentic-ai-built-by-vendor-forward-deployed-engineering-by-2028