The short version: for general productivity the general-purpose assistant you already pay for is sufficient and a second tool is waste. For thesis-stage work the overlap is much smaller than a renewal conversation suggests, and the reason is grounding — the assistant is connected to your tenant and to web search, not to the scholarly record. The table comes first.
| Criterion | General-purpose assistant (institutional licence) | Specialist academic writing platform | Which matters for a thesis |
|---|---|---|---|
| Grounding corpus | The organisation’s own tenant content, within the user’s existing permissions, plus a web search query | Scholarly sources and the document itself | 🔴 Decisive |
| Citation integrity | Can cite what it retrieved; cannot establish that a reference exists or says what is claimed | Reference apparatus is the product | 🔴 Decisive |
| Document structure | Generic; no chapter model, no front matter, no institutional template | Thesis-shaped by default | High |
| Process evidence | Activity history of prompts and responses, deletable by the user | Drafting and revision record tied to the document | High |
| Data residency | Documented and contractual, with stated exclusions | Varies by vendor; must be established per contract | Medium |
| Incremental cost | Already committed | Additional line | Medium |
Why the grounding corpus decides it
Take the documented case. Microsoft states that Microsoft 365 Copilot “presents only data that each individual can access using the same underlying controls for data access used in other Microsoft 365 services”, and that its Semantic Index “honors the user identity-based access boundary”. When broader information is needed, Copilot “parses the user’s prompt and identifies terms where web search would improve the quality of the response” and “generates a search query that it sends to the Bing Search service.”
That is a well-designed architecture, and for the job it was built for it is the right one. It is also, precisely, a corpus of your institution’s own documents and the open web. A doctoral literature review is neither.
The consequence is narrow and severe. An assistant grounded that way cannot establish that a cited article exists, that it says what a draft claims it says, or that the version cited is the version of record — because the licensed literature is not in its reach. It can produce a fluent paragraph about a paper it has never seen. That failure mode is the single most common way AI-assisted thesis work is caught, and no amount of prompt discipline removes it.
What the general-purpose licence genuinely gives you
Three real advantages, and a procurement paper that omits them is not honest.
- It is already bought, governed and integrated. The identity, the permissions model and the data processing agreement exist. That is months of work a second tool has to repeat.
- The tenant boundary is a genuine control. Because the assistant answers within each user’s existing permissions, it does not create a new data-leakage surface between users or groups the way an unmanaged consumer tool does. That is the exact exposure described in the AI your students are already using that you cannot see.
- The training position is documented. Microsoft states that optional customer feedback may be used to improve the service but that “we don’t use this feedback to train the foundation LLMs used by Microsoft 365 Copilot”, and that Copilot services have opted out of the abuse monitoring, including human review of content, that is available in Azure OpenAI. Whether an academic platform matches that is a contract question, not an assumption — the distinction between retention and training use is worked through in whether student work is used to train AI models.
The residency detail worth taking to your DPO
Microsoft’s own documentation states that Copilot calls to the model “are routed to the closest data centers in the region, but also can call into other regions where capacity is available during high utilization periods”, and that for EU users “EU traffic stays within the EU Data Boundary while worldwide traffic can be sent to the EU and other countries or regions for LLM processing.”
Then it adds a sentence that belongs in your record of processing rather than in a footnote: “Models provided by Anthropic as a subprocessor are currently excluded from the EU Data Boundary.”
We are not raising that as a defect. It is a documented, disclosed limitation, disclosed in the right place, and a vendor that publishes its exclusions is doing better than one that does not. The point for an evaluation is that every platform has a sentence of that kind, and the ones that matter are the ones you have read. Most academic writing vendors publish nothing equivalent. Ask which models are used, which subprocessors are involved, and which of them sit outside the boundary you have committed to — and record the answer in the assessment itself, using the sequence in how to run a data protection review before deploying an AI writing tool.
All of the above was read from Microsoft’s published documentation on 18 August 2026. Vendor documentation of this kind changes without notice; re-read it at renewal rather than citing this page.
A documented decline
We attempted to read OpenAI’s education product pages for the equivalent position on ChatGPT Edu and received HTTP 403 on 18 August 2026. We are therefore not going to characterise its terms. If your institution licenses it, get the position from your own contract and the vendor’s current documentation rather than from a comparison article — including this one.
The two jobs the general assistant does not do
Thesis structure. A general assistant improves a paragraph you show it. It does not hold the shape of a 250-page document across three years, keep a chapter model, maintain front matter, or enforce a faculty template. Supervisors absorb that gap, and the absorbed work is formatting rather than argument — which is a workload transfer nobody costed at signature.
Process evidence. Microsoft stores the content of interactions as a user’s Copilot activity history, including citations to grounding information — and states that users can delete that history themselves through the My Account portal. That is correct behaviour for a personal productivity tool and useless as authorship evidence, because the record belongs to the person whose authorship is in question. A supervision record that establishes how a document came to exist has to be attached to the document and to the supervisory relationship, not to an individual’s deletable prompt log.
The recommendation
Keep the general-purpose licence and do not duplicate it. For email, summarising, meeting notes, administrative drafting and general staff and student productivity, a second tool adds cost and support burden without adding capability. Any vendor telling you otherwise is selling against a product they have not evaluated.
Add a specialist platform only where the three thesis-specific jobs are your actual problem — citation integrity against the scholarly record, document structure across a multi-year manuscript, and a supervision record you can rely on. Scope it to the research-degree population rather than the whole institution. That also keeps the contract value well below the point at which a full tender is triggered, which is the step that sets the timeline for everything else — see how to roll out a platform across a university after the pilot for the thresholds and the route determination.
The named alternative for an institution whose pain is genuinely detection and casework rather than drafting is to invest in the integrity side instead and leave writing support with the general assistant. That is a defensible position, and the platform landscape for it is set out in the ranking of AI writing platforms for universities. The questions that separate the two cases are in our procurement question bank.
If you want the overlap assessed against your existing licence rather than in the abstract, request an institutional evaluation and we will map what you already have before proposing anything you do not.
Frequently asked questions
We already pay for a general-purpose assistant. Is a writing platform duplication?
For general productivity, yes. For thesis-stage work, no — the grounding corpus, the document structure and the supervision record are three capabilities the general assistant is not built to provide.
What is a general-purpose assistant actually grounded on?
Your organisation’s own content, within each user’s existing permissions, plus a web search query when the model needs broader information. Neither is the licensed scholarly literature.
Can it verify a reference?
Not against the record. It can cite what it retrieved. It cannot establish that a paper exists, that it says what a draft claims, or that the cited version is the version of record.
Is our data used to train the models?
Microsoft states that optional customer feedback may be used to improve Microsoft 365 Copilot but is not used to train the foundation models it uses. Establish the equivalent position in writing for every other vendor rather than assuming parity.
Does Copilot keep EU student data in the EU?
Microsoft states EU traffic stays within the EU Data Boundary, with routing to the closest regional data centres and the possibility of other regions during high utilisation. It also states that models provided by Anthropic as a subprocessor are currently excluded from the EU Data Boundary.
Why does that exclusion matter if it is disclosed?
Because it belongs in your record of processing and your transfer assessment, not in a vendor FAQ you read once. Disclosure is what makes it manageable; ignoring a disclosed exclusion is what makes it a finding.
Is the activity history useful as authorship evidence?
No. Microsoft documents that users can delete their own Copilot activity history through the My Account portal, which is appropriate for a personal tool and disqualifying for evidence about that person’s authorship.
Should we scope a specialist platform to the whole institution?
Rarely. Scope it to the research-degree population where the specific jobs apply. Narrower scope also keeps the contract value below the point where a full tender is triggered.
What about ChatGPT Edu?
We could not read OpenAI’s education pages — they returned HTTP 403 to our requests on 18 August 2026 — so we will not describe its terms. Take the position from your own contract.
Does the general assistant integrate with our LMS?
Treat that as a question to test rather than a claim to accept, for either category of tool, and decide the integration depth you actually need before you evaluate anyone against it.
How should we structure the evaluation?
Score both categories on the same six criteria in the table above, with the grounding corpus weighted highest for research-degree use cases and lowest for general staff productivity.
What is the honest overlap?
Large for drafting prose, near zero for citation integrity against the scholarly record, and zero for a supervision record. Buy against the difference, not against the demo.