Registered Research Field · ANZSRC 4602
Artificial intelligence: where implementation ends and research begins
Connecting an established model to a workflow is usually implementation. Research may begin when vendor claims and public benchmarks cannot determine whether it will perform on your data, under your constraints, at an error rate the business can accept.
How this field connects
A project can draw on more than one registered field. The links below make those technical intersections visible without forcing a business problem into a single industry label.
Built Environment and Design
Engineering
Environmental Sciences
Information and Computing Sciences
01
Where implementation ends and research may begin
A project is not research merely because it is difficult, new to the team or commercially important. We first examine standards, published knowledge, supplier evidence and comparable work. Research begins only where the required technical outcome still cannot be determined with confidence.
Usually implementation
- integrating a documented model API
- reproducing published benchmark results
- configuring a standard workflow
May require research
- establishing a performance ceiling on sparse proprietary data
- constructing metrics for asymmetric error costs
- testing performance inside edge-device constraints
02
Open questions in this field
Can the model reach the operational threshold on the data we can realistically produce?
Evidence needs a pre-defined holdout set, a threshold tied to an operating consequence, configuration records and error analysis by class.
Which errors should the system optimise against?
Accuracy can hide the errors that matter most. The work may need a cost-sensitive metric agreed before optimisation.
Can human review make the system reliable without removing its value?
Review rate, override pattern, escalation time and residual risk must be observed over time.
Can the system remain useful as data changes?
Where ground truth arrives late, drift detection may require proxy signals and prospective alert thresholds.
Can accuracy, latency, memory and power requirements be met together?
Model variants must be compared on a fixed device and workload profile.
03
Taking a question from this field forward
We start from the operating decision rather than a preferred technology: isolate what is already known, define the outcome that remains unresolved, set thresholds before the work begins, and stage it so the highest-consequence assumption is tested first.
The Ignition research value chain
One methodology, six steps — each with a named output
- 01FrameA testable research question
- 02DesignExperiment plan & controls
- 03ExecuteRuns & recorded findings
- 04IntegrateSpecialists, coordinated
- 05GovernEvidence log & audit trail
- 06ReturnValidated, claim-ready records
General information only. This page does not determine R&D Tax Incentive eligibility and is not tax, legal, financial or grant advice. Whether any government support pathway may be relevant depends on the project's specific activities, structure, evidence and the current program rules.
Next step
A researcher reads your description — not a salesperson.
Bring the operating problem, the result you cannot explain or the decision your team cannot make with confidence. You will get a reply within two business days — including when the honest answer is that a research pathway is not the right tool.