Field note · AI architecture
What should remain research, and what is ready to become engineering?
A useful boundary for teams moving technically ambitious AI work toward a product programme.
Research and engineering create different kinds of certainty.
Research reduces uncertainty about what is possible, which representations are promising and where an approach fails. Engineering creates repeatability: defined inputs, interfaces, tests, ownership, operating conditions and recovery paths.
Problems emerge when organisations ask one mode of work to behave like the other.
Signals that a question still belongs in research
- The team is comparing fundamentally different problem formulations.
- Success criteria are still changing as the problem becomes better understood.
- The main value lies in learning whether an approach is possible.
- Failure modes are not yet sufficiently characterised to define an operating boundary.
Signals that the work is ready for engineering
- Inputs, outputs and system responsibilities can be specified.
- Evaluation represents the intended operating context.
- Known failure modes have an owner and a response.
- The team can define what must remain stable while individual models change.
- Delivery risk now matters more than scientific novelty.
The transition is not a ceremony. It is an architectural decision. Teams need to separate what they know about the problem from what they merely observed in a prototype.