When is a custom AI model justified?
Why A is the answer
When existing models cannot meet a measured requirement
Custom models add cost and upkeep, so they need a clear gap.
In detail
- A custom model needs a measured requirement that existing models miss.
- Keep the test results as the case for it.
Use case
An illustrative scenario, not a real incident.
Testing shows that available models miss rare defects in a firm's images. That measured gap justifies a custom model.
See it in different ways
flowchart TD
A{"Does a prebuilt option meet the need?"} -->|Yes| B["Buy: Microsoft 365 Copilot or a prebuilt agent"]
A -->|"Mostly"| C["Extend with an agent or connector"]
A -->|No| D["Build a custom agent"]stateDiagram-v2 [*] --> Requirement Requirement --> Copilot: Copilot already meets it Requirement --> Extend: Copilot needs your data or tools Requirement --> Build: needs its own logic Build --> CustomModel: existing models fail a measured need
flowchart TD
Q{"Did existing models fall short in testing?"}
Q -->|"Yes"| A0["Consider a custom model"]
Q -->|"No"| A1["Use an existing model"]Key point: Custom models need a measured gap.
Why the other options are wrong
- B. When a new model would be interesting for the team to build
- Interest is not a business reason.
- C. When a prebuilt agent has met all of the requirements
- If a prebuilt agent meets the need, stop there.
- D. When the project needs a more impressive demonstration
- A demo is not a requirement.