R&D tax credits for AI & agentic systems.
From foundation-model teams to applied AI startups, UAE companies are building systems that push past what off-the-shelf tools allow. That experimental work often qualifies for the R&D Tax Credit — the discipline is proving precisely which part.
How qualifying R&D shows up in AI.
The UAE regime tests every project against the five OECD Frascati criteria. Here is what each looks like in a training run, an agent loop, or an evaluation harness.
Novel
A model, method or agent architecture not already available in the field.
Creative
Original approaches to training, retrieval or evaluation.
Uncertain
Whether the approach will actually work is unknown at the outset.
Systematic
Planned experiments with tracked runs and measured results.
Transferable
Methods reproducible beyond one researcher or dataset.
The kind of work that qualifies.
Illustrative projects across model development, agentic systems and applied AI. Eligibility is always confirmed project by project.
Domain-specific model training
Training or adapting a model where existing architectures fail on Arabic-language or region-specific data.
Reliable agentic systems
Building an agent that plans and executes multi-step tasks dependably under real-world uncertainty.
Inference optimisation
Novel techniques to cut inference cost or latency below what standard methods allow.
Grounded RAG at scale
Retrieval and grounding methods that hold accuracy where off-the-shelf pipelines break down.
Additive manufacturing
New materials or additive methods requiring iterative process development.
Materials engineering
Developing or adapting materials to perform where existing ones fail.
Novel evaluation methods
Building evaluation and safety harnesses for behaviours existing benchmarks cannot measure.
Multimodal integration
Fusing text, vision or signal data in ways current models do not readily support.
Genuine R&D, not routine engineering.
The value we add is drawing this line correctly — claiming what qualifies, and defending it, while leaving out what does not.
- Experimental work to overcome a genuine technical uncertainty in a model or method.
- Developing new training, retrieval or evaluation techniques with unknown outcomes.
- Re-engineering systems to reach a performance target thought unreachable.
- Systematic, tracked experimentation to resolve the uncertainty.
- Calling an existing model API or standard fine-tuning with off-the-shelf tooling.
- Building routine application screens or CRUD around a model.
- Prompt engineering that a competent professional could readily deduce.
- Data labelling and pipeline plumbing with no technical uncertainty.
£50M+
claimed across 450+ UK companies with a 100% audit success record — including complex, engineering-heavy claims in energy and industrials. The same chartered method now applies to the UAE.
“We needed a partner who understood deep-tech innovation and could handle complex claims efficiently. RDvault delivered exactly that.”
Hybird — a Techstars company
A UAE AI case study will feature here as claims complete under the new regime.
We mostly use existing foundation models. Can we still claim?
Often, yes — if your team did genuine development to overcome a technical uncertainty (a novel architecture, training method, retrieval approach or evaluation technique). Simply calling an API or standard fine-tuning does not qualify.
Does prompt engineering qualify?
Rarely on its own. If achieving a reliable result required systematic experimentation to resolve a real uncertainty — not tuning a competent professional could readily deduce — the surrounding development work may qualify.
Is data work eligible?
Routine labelling and pipeline plumbing usually is not. But novel methods to generate, augment or evaluate data where the approach itself is uncertain can form part of a qualifying project.
What evidence will we need?
Contemporaneous records: the technical uncertainty, the experiments run to resolve it, and project-level cost and staff-time allocation. We build this as the work happens — and pre-approval is mandatory before claiming.
Building AI? Let’s find what qualifies.
We assess your projects against the five criteria, handle pre-approval, and build the evidence — before a single figure is claimed.
