Extend, configure, or build custom

This is the build-versus-buy decision, expressed as a list you work down until
something fits. The order is the answer — the exam rewards stopping early far
more often than it rewards building something impressive.

Work down the list, stop at the first row that fits

The capability already exists in an app you own → use the prebuilt. Zero build,
and the vendor maintains it.

It nearly fits → configure and extend it, which keeps you on the upgrade path.
That last point is the one people underrate: a customisation that keeps you
upgradeable is worth far more than one that does not.

Users live in Office and only need knowledge → extend Microsoft 365 Copilot.
It meets them where they already are.

It must act, face customers, or own its lifecycle → build in Copilot Studio.
Actions, channels, application lifecycle management.

It needs a specific model, evaluation or isolation → build in Foundry. Depth
and control.

When a custom model is genuinely justified

Good reasons. A narrow, high-volume, repeatable task where a smaller model is
cheaper to run. A domain where off-the-shelf output is stylistically wrong. A
regulatory requirement for a model you control end to end.

Bad reasons, and these are the ones that appear in scenarios:

"It doesn't know our products" — that is grounding.
"It gets facts wrong" — grounding again, or retrieval.
"We want it to sound like us" — try instructions and examples first.

The ladder, in order

Before anybody proposes training a model, walk up this:

Better prompting → better grounding → a different off-the-shelf model →
fine-tuning.

And the sentence that resolves most of these questions on its own: fine-tuning
teaches form, not facts. If the complaint is about what the model knows,
fine-tuning is the wrong instrument regardless of how sophisticated it sounds.

Exam trap. Extend before you build is one of the judgements the whole exam
rewards. An option proposing a custom build, when a prebuilt or an extension
would do, is wrong even when it is technically excellent.

The one line to carry: work down the list and stop at the first row that fits —
and if the problem is what the model knows, no amount of training fixes it.

---

Currency (19 Aug 2026). "Does it need to act?" still decides this, but the line has
moved slightly: Agent Builder cannot reach external services — that remains Copilot
Studio or Foundry — while declarative agents gained first-party email and meeting
actions in July 2026. See card 44.