Model development, perception and autonomous systems: connect technical experiments to documented research expenses.
Estimate your credit →AI and robotics teams may conduct qualified research when they evaluate technical alternatives to improve a model, software product or robotic system. The work must address uncertainty about capability, method or appropriate design, rely on technical principles and satisfy the other Section 41 requirements. Buying an AI tool or deploying a known model is not enough.
Only documented qualified research expenses count. Eligible costs may include research wages, supplies used in research, qualifying computer rental and generally 65% of qualifying contract research. Production hosting, depreciable equipment and routine operations are not automatically eligible.
This guide covers AI model development as well as embodied AI and robotics. General platforms, APIs and non-AI software development remain in the software guide. Explore Software / SaaS
These activities may qualify only when the project meets the four-part test and applicable exclusions. A job title, new product or use of AI does not establish eligibility.
| Business component | Example | Technical uncertainty |
|---|---|---|
| Model architecture | Compare architectures for a recommendation or anomaly-detection model | Which architecture can meet accuracy and latency constraints with available data? |
| Training and evaluation | Test training strategies and evaluation harnesses | Can the model generalize to unseen conditions rather than memorize training examples? |
| AI agents and retrieval | Evaluate retrieval and verification designs for an assistant | Which design achieves reliable grounded outputs under realistic workloads? |
| Inference optimization | Compare quantization and serving architectures | Can memory and latency be reduced without unacceptable loss of accuracy? |
| Perception and sensor fusion | Combine camera, lidar and inertial measurements | Which fusion approach remains reliable under occlusion and sensor noise? |
| Motion planning | Prototype collision-aware planning for a mobile robot | Can the robot navigate changing obstacles within its compute budget? |
| Manipulation | Test grasp planning for objects with variable geometry | Which perception and actuation design achieves repeatable handling? |
| Simulation to hardware | Compare simulation assumptions against physical prototypes | Which model changes address the performance gap on real hardware? |
Find your own job titles below. Each row shows the wage category a role usually falls under and the kinds of work that typically do and do not count. Titles alone never make wages eligible; the activity must meet the four-part test.
6 of 6 roles shown
| Role | What they do | Wage category | Typically qualifies | Typically does not |
|---|---|---|---|---|
| Builds, trains and evaluates models for the product. | Direct research | Comparing model architectures to resolve uncertainty, Experimental training and inference optimization | Calling a vendor API without technical experimentation, Routine prompt edits | |
Typically qualifies
Typically does not
|
||||
| Builds the planning and motion software inside the robot. | Direct research | Testing uncertain motion-planning alternatives, Prototype manipulation algorithms | Installing a standard robot, Routine production calibration | |
Typically qualifies
Typically does not
|
||||
| Helps machines interpret cameras and other sensors. | Direct research | Sensor-fusion experiments, Evaluating vision models under uncertain operating conditions | Routine dataset labeling, Running established inspection models | |
Typically qualifies
Typically does not
|
||||
| Builds data and compute systems for model experiments. | Direct research | Experimental training-pipeline architecture, Testing uncertain distributed inference designs | Routine data exports, Production hosting administration | |
Typically qualifies
Typically does not
|
||||
| Sets up and records controlled prototype trials. | Direct support | Preparing rigs for qualified experiments, Collecting prototype test results directly supporting research | Routine production quality checks, Customer maintenance | |
Typically qualifies
Typically does not
|
||||
| Directly supervises the engineers doing the research. | Direct supervision | First-line technical supervision of qualifying experiments, Reviewing experimental design alternatives | General management, Fundraising and recruiting | |
Typically qualifies
Typically does not
|
||||
Wage categories follow IRC Section 41. Only the portion of time spent on qualified research counts, and the percentage must be documented.
A team cannot determine which perception and grasp-planning approach will handle reflective parts. Engineers compare model architectures in simulation, test prototypes on a controlled rig and record success rates and failure modes. Research wages and eligible prototype supplies may be reviewed when the work meets all requirements.
Separate routine labeling, production deployment, customer training and standard equipment purchases from the experiments. A failed prototype can still support the research record; success is not the eligibility test.
Illustrative project, not a credit estimate or eligibility determination. Document the uncertainty, alternatives, results and eligible expenses before calculating a credit.
The IRS requires qualifying research to satisfy four tests. Here's how they typically map for this industry:
Seek improved function, performance, reliability or quality of a model, software product or robot used in the business.
Use computer science, mathematics as applied to technical development, or engineering principles to evaluate the design.
Identify what was unknown at the outset about model capability, training method, sensor fusion or robot design.
Evaluate alternatives through documented training runs, benchmarks, simulation and physical prototype tests.
Take these into a 15-minute conversation with your team.
The last question matters most. Customer-funded work and work performed outside the US are generally excluded. Funded work is excluded to the extent your business does not bear the financial risk or does not retain substantial rights in the results.
No. Buying a model, adding a chatbot or using vendor APIs does not by itself establish qualified research. Identify the technical uncertainty and the alternatives you experimentally evaluated.
Potentially, when the work evaluates technical alternatives to resolve uncertainty and meets the four-part test. Routine retraining with established methods and ordinary data preparation should be separated from qualifying experiments.
Nondepreciable supplies used in qualified research may count, subject to the expense rules. Depreciable robots, machinery and other capital equipment are not qualified research supplies merely because they are used for experiments.
Amounts paid for the right to use computers in qualified research may qualify as computer rental under Section 41. Separate experimental workloads from production inference and general hosting, and retain usage and billing records.
Qualifying contract research generally uses 65 percent of eligible payments for US research, subject to the requirements. Review who bears economic risk and retains substantial rights; customer or grant funding can affect eligibility rather than automatically deciding it.
Keep model versions, experiment plans, evaluation results, simulation logs and prototype test records that connect alternatives to technical uncertainty. Pair them with supported time allocations and expense records; internal-use AI software may face additional requirements. Review the IRS rules
Most companies underestimate what they can claim. Our study covers every qualifying role, every component, and every year still open for amendment.
Get your free estimate →