Your AI platform sets the ceiling

Halfway through one implementation, the chosen platform turned out not to support everything the design needed. The prompts were good and the design was sound. The platform could not do it, and no prompt fixes that. We adjusted the design and the expectations, but those limits should have been on the table before the start.

Since then I check the platform before anything gets built. Copilot Studio, a no-code builder or a chatbot you buy off the shelf: the choice fixes what the system can ever do. In week one you can still move that line. In week three you pay for it.

Why does the platform set the ceiling?

A platform fixes which models you use, which systems it can talk to and which data it can read. A prompt steers what the model does inside that space. It does not change the space.

In practice I see five kinds of limits. Which models are available and who decides. Which external systems the platform is allowed to call. Which files it reads, up to what size and in what format. How it behaves in the channel where people actually use it, which can differ from the test environment. And how much traffic your licence allows.

A demo shows none of the five. It runs on three sample files, with one user, in the builder’s environment. The limits appear once the real work arrives: the long manual, the protected contract, the whole department asking questions at once.

What are the limitations of Copilot Studio?

Microsoft publishes them on its “Quotas and limits” page (Microsoft Learn, accessed September 2026), and some only show up with real files and real users. Four examples from that page:

• If the maker has no Microsoft 365 Copilot licence in the same tenant, the agent only uses SharePoint files under 7 MB for generative answers. Larger files have to be split.
• Uploaded documents with a confidential or highly confidential sensitivity label, or a password, are not indexed. They show as ready for use but return no answers.
• Agents cannot write and run code. Microsoft itself notes that answers to analytical questions over Excel files may not be optimal as a result.
• A trial or developer environment handles 10 generative AI requests per minute. Anyone sending a message above that gets an error.

None of these is a reason to skip Copilot Studio. A help desk over well-kept Word documents in your own Microsoft 365 environment fits it well. If your use case revolves around numbers in spreadsheets or confidential contracts, its core sits right on a limit. Microsoft updated that page as recently as August 2026, so read it again before your kickoff.

Copilot Studio or custom build?

Choose the platform when your use case sits well inside its limits and your organisation already runs it. Choose a custom build when the core of the use case sits on a limit and you have someone to maintain the system after the first release.

That second condition matters more than it sounds. A platform your IT department already runs has an owner, a security model and a vendor shipping updates. A custom build only has those once you arrange them. A good prompt inside an existing product sometimes beats a €15,000 bespoke application; I wrote about that in From building AI to deciding on AI.

There is a third outcome people often miss: the same use case in a different place. Sometimes the work fits better in a system the team already uses, and no new platform is needed at all.

Which ceiling sits outside the AI platform?

Your existing systems, your licences and the people who have to decide. An AI platform that can do everything is no use against a source system without a connector, or a licence your users do not have.

The 7 MB example shows how these connect: that limit disappears with a different licence. Who can request which licence, and how long does that take? Who in security, legal or the works council still has to say yes? A platform that fits technically but gets approved a quarter from now does not fit this project.

So before kickoff I look at four things: the AI platform, your own systems, the licences and who decides. Sometimes the outcome is a different platform, sometimes a different home for the use case.

The pre-kickoff platform check

I ask these seven questions before a design exists. If it takes a day, that day is well spent.

1. Which three things can this platform not do? Ask the vendor or the agency. “None” is a red flag.
2. Which models are available, and who chooses? Can you switch later, or are you tied to what the vendor offers?
3. Which systems does it need to talk to? Is there a connector for each, and does the data volume fit within that connector’s limits?
4. Which files does it need to read? Test with the real ones: the largest, the oldest scan, the document with a sensitivity label. Not with three tidy sample PDFs.
5. Which licences are needed, for whom? And what happens to limits and costs when you move from a trial environment to production?
6. Which channel will people use it in? Test in that channel itself.
7. Who still has to say yes? Security, legal, the works council. Put their names and a date in the plan.

If the core of your use case sits on a limit, change the platform or the use case before you build. I would not hire an agency that does not ask these questions itself; the rest of my criteria are in the guide to choosing an AI agency, and the lesson this comes from is in my memo on 35+ implementations.

You only see the ceiling if you ask about it.