Budgeting Data Labelling in a Research Grant: A Guide for Australian Universities

Why Annotation Is an Increasingly Common Grant Line Item

Australian research investment in AI and machine learning has grown substantially. The Australian Research Council (ARC) awarded over AUD $850 million in grants in 2024–25, with a significant and growing proportion supporting AI-adjacent research in health, environment, and social sciences. The National Health and Medical Research Council (NHMRC) similarly funds projects across clinical AI, medical imaging, and health informatics — all areas where labelled training data is as fundamental to the research methodology as participant recruitment or laboratory reagents.

Yet annotation costs are frequently under-budgeted or omitted from grant applications. The reason is usually unfamiliarity: investigators who have built models using publicly available benchmark datasets underestimate what it costs to create a labelled dataset from study-specific data. When the gap emerges mid-project — when a research team discovers that the 40,000 fundus images they collected need expert ophthalmologist annotation, not a student annotator — the options are to seek a budget variation, reduce scope, or cut quality. None of these is good.

Experienced AI researchers now budget annotation as a first-class line item alongside participant payments, equipment, and computational resources. The AI Taggers university and research annotation service exists specifically to support this category of work, providing the IAA documentation, ethics-compliant data handling, and publication-grade quality controls that academic projects require.

How ARC and NHMRC Budgets Classify Annotation Costs

Neither the ARC Funding Rules nor the NHMRC Funding Agreement explicitly names “data annotation” as a budget category — which is why researchers sometimes assume it is ineligible. In practice, annotation is classified under two existing categories depending on how it is delivered.

Research Services (or Contract Research): When annotation is provided by an external commercial vendor under a service agreement, it is classified as a research service. This is the most common arrangement and the most straightforward to justify: the vendor provides a specific, bounded deliverable (labelled data) that the research team could not produce in-house within the project timeline or to the required quality standard.

Personnel (Casual/Sessional): When researchers hire individuals to annotate data directly — for instance, engaging casual research assistants as annotators — the cost sits in personnel. This arrangement gives the research team direct management control but introduces payroll obligations, training costs, and variable quality risks. It also becomes difficult to scale quickly without triggering enterprise agreement constraints at most universities.

The vendor-as-research-service model is generally simpler to budget, more scalable, and produces the external QA documentation that ethics committees increasingly require. For budgets over AUD $10,000, most universities require three competitive quotes; below that, a single quote with a brief justification is usually sufficient.

What Data Labelling Actually Costs: A Range by Task Type

The biggest source of budget error is applying a single cost assumption across annotation task types that have very different price structures. The following ranges are indicative for Australian research-grade annotation with documented IAA (i.e., not crowdsourced). See the AI Taggers pricing page for current quotes.

Ranges are indicative for research-grade annotation with documented IAA. Actual quotes depend on schema complexity, volume, timeline, and annotator expertise required.

How to Write the Budget Justification

Grant reviewers and research office administrators need to understand three things from a data labelling budget line: what the annotation task is, why it cannot be done by the research team alone (or more cheaply), and how the cost was estimated. A budget justification paragraph covering these three points will satisfy most ARC and NHMRC review panels.

An example for a medical imaging project:

Key elements: specific task, specific volume, specific unit price, qualifications required, and why the research team cannot do it themselves. This is what separates a credible budget line from a round-number estimate that raises flags.

Ethics Approval and Annotation: Getting the Sequence Right

A common sequencing mistake is waiting for ethics approval before engaging with annotation vendors. This delays budget quotes, which delays the grant application, which delays the project. The correct sequence is:

1. Pre-ethics: Obtain budget quotes using a synthetic or publicly available sample dataset that mirrors your study data. Describe the annotation task to the vendor in general terms without sharing identifiable participant data. This gives you a realistic budget estimate for the grant application.
2. Grant application: Include the annotation line item with justification and the quote as an attachment. Note in the budget narrative that data handling will occur under an ethics-approved data processing agreement.
3. Ethics application: Include the vendor's data processing agreement template and describe the annotation workflow in the data management plan. Most Human Research Ethics Committees (HRECs) require a data processing agreement that specifies data residency, retention limits, access controls, and breach notification.
4. Post-ethics approval: Execute the data processing agreement, then share study data with the vendor for annotation. Do not share identifiable data before the HREC has approved the data handling arrangements.

For projects involving sensitive data — health records, children's data, or data from vulnerable populations — the HREC will typically require a specific description of how annotation staff are trained in privacy obligations and how access to the data is controlled and logged. Ask your annotation vendor for their HREC-submission data management documentation before submitting your ethics application.

What Research-Grade Annotation Deliverables Look Like

The quality documentation that a peer-reviewed publication needs differs from what a product team needs. Journals publishing machine learning research increasingly require authors to report inter-annotator agreement metrics, annotator qualifications, and annotation guidelines as part of the methods section or supplementary material. Major venues including NeurIPS, ICML, and MICCAI now have data documentation requirements that effectively mandate this provenance.

For Australian medical AI research, the Therapeutic Goods Administration (TGA) Software as a Medical Device (SaMD) framework and its reference to IEC 62304 and ISO 14971 create additional documentation requirements for datasets used in commercial-pathway research. If your university project has a commercialisation pathway — which many NHMRC Investigator Grant projects do — annotation documentation needs to be audit-ready from the start, not reconstructed later.

A well-specified research annotation engagement should deliver: an annotation guidelines document (publication-ready), per-record annotation with confidence scores or multi-annotator votes, inter-annotator agreement calculation (Cohen's kappa or Krippendorff's alpha depending on task type), a data provenance log at annotator level, and an HREC-format data processing summary. Specify these as contract deliverables before the project begins.

For more on what rigorous annotation quality controls involve and when they matter most, see the guide to validating annotation quality before it reaches your model. For a broader discussion of the build-versus-buy decision for annotation capacity, see build vs buy annotation.

A Representative University Research Project

Consider a three-year NHMRC Investigator Grant for a project developing a deep learning model for early detection of age-related macular degeneration (AMD) from colour fundus photographs. The dataset comprises 25,000 fundus images collected from collaborating ophthalmology clinics across two states.

The annotation requirement is grading each image for AMD severity using a four-level scale (no AMD, early, intermediate, late). This requires credentialed ophthalmologists — the treating clinicians who captured the images are not available for annotation, and research assistants are not qualified to grade AMD severity. The research team estimates 2,500 cases will be ambiguous and require adjudication between two independent graders.

Budget structure: 22,500 standard-complexity images at AUD $5 per image = AUD $112,500. 2,500 adjudicated images (two graders plus adjudicator review) at AUD $18 per image = AUD $45,000. IAA calculation and documentation: included. HREC data handling agreement: provided at no cost. Total annotation budget line: AUD $157,500 across Years 1 and 2, split roughly 60/40.

This budget would be supported by a vendor quote, a brief methodology justification, and the statement that multi-reader adjudication is required to produce the gold standard labels needed for training a clinical-grade model. The NHMRC review panel would see a credible, well-evidenced budget line.

Free Pilot Samples Before Grant Submission

One practical option that many research teams do not know is available: annotation vendors who work with universities can often provide a small free pilot — typically 50–200 records — before a grant is awarded. This serves two purposes. It gives the research team a realistic sense of annotation quality and turnaround time before committing grant funds. And it provides concrete feasibility evidence for the grant application.

A research team that can state in their methodology section “we piloted annotation on 100 images from the target data distribution and achieved Cohen's kappa of 0.82 between independent graders” has materially strengthened their case that the annotation methodology is achievable and the quality targets are realistic. Grant reviewers respond well to teams who have done the legwork rather than simply asserting feasibility.

AI Taggers offers a free pilot for qualified research projects. Contact us with a brief description of your study data, annotation task, and grant timeline — we can provide a pilot dataset and IAA report within five business days, at no cost, to support your application.

Frequently Asked Questions