# SR&ED Examples: 19 Eligible Projects in 5 Industries

Source: https://grantops.ai/en/blog/sred-eligible-project-examples/
Last updated: 2026-11-13

## What Makes a Project Eligible for SR&ED?

The **Scientific Research and Experimental Development (SR&ED)** program is Canada's largest R&D tax incentive, delivering [approximately $4.5 billion annually](/en/blog/what-is-sred-guide/) in tax credits to businesses conducting eligible research and development. But one question dominates every conversation between R&D teams and finance departments: *does our work actually qualify?*

SR&ED eligibility comes down to three requirements that must all be present:

1. **Technological uncertainty** -- You attempted something that could not be accomplished using standard practice, known techniques, or publicly available knowledge. The outcome was genuinely uncertain when the work began.
2. **Systematic investigation** -- You followed a structured approach: formulating hypotheses, designing experiments or prototypes, testing, analyzing results, and iterating. Trial and error alone does not qualify unless it follows a methodical process.
3. **Technological advancement** -- The work aimed to generate new knowledge or push the boundaries of existing technology. The advancement does not need to succeed, failed experiments that produced new understanding are eligible.

These three criteria come directly from the [Income Tax Act, subsections 248(1)](https://laws-lois.justice.gc.ca/eng/acts/I-3.3/page-174.html) and are reinforced by decades of CRA policy and Tax Court decisions. Understanding them in practice, through real examples, is the fastest way to determine whether your projects qualify.

This guide provides **19 concrete SR&ED project examples** across five industries (software and IT, manufacturing and engineering, life sciences, AI and machine learning, clean technology), each describing the specific technological uncertainty, the systematic approach taken, and the advancement achieved.

---

## The CRA's Three-Question Eligibility Test

Before diving into examples, it helps to understand the framework CRA's Research Technology Advisors (RTAs) use when evaluating claims. The CRA applies three sequential questions to every SR&ED project:

### Question 1: Was there a technological uncertainty?

Could the problem be solved using standard engineering, available information, or known methods? If the answer is yes, the work is routine and does not qualify. CRA looks for situations where the outcome was uncertain at the outset, where competent professionals in the field could not predict whether a given approach would work.

**Key distinction:** Technological uncertainty is not the same as business uncertainty. Not knowing whether customers will buy a product is a market question. Not knowing whether a particular algorithm can process 10 million events per second with sub-100ms latency on commodity hardware is a technological question.

### Question 2: Was there a systematic investigation?

Did the company follow a structured process to resolve the uncertainty? CRA expects to see evidence of hypothesis formulation, planned experimentation, observation, testing, analysis, and conclusions. Ad hoc troubleshooting or undocumented trial and error typically does not meet the threshold.

### Question 3: Was there a technological advancement?

Did the work attempt to generate new technological knowledge or capability that goes beyond what was publicly available? The advancement can be incremental, and the project does not need to succeed. What matters is that the work was designed to push the boundary of what is currently known or achievable.

A project must satisfy **all three questions** to be eligible. If any one is missing, the claim will be denied. Now, let us see what this looks like in practice across specific industries.

---

## Software and Technology SR&ED Examples

[Software companies](/en/sred/software/) are among the largest claimants of SR&ED tax credits in Canada, yet many underestimate what qualifies. Any development work that involves resolving genuine technical uncertainties, not just building features, is potentially eligible.

### Example 1: Real-Time Stream Processing Under Extreme Throughput Constraints

**The uncertainty:** A fintech company needed to process 5 million financial transactions per second with end-to-end latency under 50 milliseconds, while maintaining exactly-once delivery guarantees. Existing stream processing frameworks (Kafka Streams, Apache Flink) could not meet the combined throughput and latency requirements at this scale without violating consistency guarantees. No published literature or vendor documentation addressed the specific combination of constraints.

**The systematic approach:** The team formulated three hypotheses for resolving the bottleneck: a custom partitioning strategy based on transaction graph topology, a novel checkpoint algorithm that decoupled state persistence from the processing pipeline, and a hybrid approach combining both. Each approach was implemented as a proof of concept, instrumented for latency measurement at the 99th percentile, and subjected to load testing with synthetic transaction data replicating production patterns. Results were analyzed to identify the failure modes of each approach.

**The advancement:** The team developed a new checkpoint coordination protocol that reduced state synchronization overhead by 73% compared to Flink's default algorithm, enabling the target throughput-latency combination. This represented new knowledge about distributed state management under high-throughput, low-latency constraints that was not available from existing sources.

### Example 2: Multilingual NLP Pipeline for Ambiguous Legal Text

**The uncertainty:** A legal technology company needed to extract structured data from Canadian legal documents written in mixed English-French text with domain-specific terminology, inconsistent formatting, and frequent use of ambiguous references. Standard NLP models (including fine-tuned BERT and GPT-based extractors) achieved only 61% accuracy on the company's benchmark dataset because of code-switching between languages within single paragraphs and the highly specialized legal vocabulary that was poorly represented in general training corpora.

**The systematic approach:** The team systematically evaluated five approaches: fine-tuning a multilingual model on a curated legal corpus, developing a two-stage pipeline with language detection followed by language-specific extraction, creating a custom tokenizer trained on bilingual legal text, combining rule-based pattern matching with neural extraction, and training a domain-adapted model from scratch on 50,000 annotated legal documents. Each approach was evaluated against the same 2,000-document test set with precision, recall, and F1 scores tracked across 14 entity categories.

**The advancement:** The final architecture, a two-stage pipeline using a custom bilingual tokenizer feeding into a domain-adapted transformer model, achieved 89% accuracy, a 28-point improvement. The custom tokenizer's handling of bilingual legal compounds represented new knowledge about tokenization strategies for code-switching domains.

### Example 3: Distributed System Architecture for Unpredictable Load Patterns

**The uncertainty:** A SaaS platform experienced traffic spikes of 200x baseline during periodic events, with spike timing that was unpredictable. Standard auto-scaling mechanisms (AWS Auto Scaling, Kubernetes HPA) had 3-5 minute cold-start delays, causing cascading failures during rapid traffic increases. The challenge was to design an architecture that could absorb a 200x load increase within 15 seconds without over-provisioning resources during normal operation.

**The systematic approach:** The team investigated predictive pre-scaling using time-series analysis of historical traffic patterns, a warm pool architecture with novel resource-sharing mechanisms between tenants, and a serverless-hybrid approach using pre-warmed function containers. Each approach was prototyped, tested under simulated spike conditions matching historical patterns, and evaluated on four metrics: response time during spike onset, time to full capacity, cost during baseline operation, and request failure rate during transition.

**The advancement:** The team developed a predictive pre-scaling system using a custom LSTM model trained on historical traffic data combined with a novel warm-pool sharing algorithm. The system reduced spike response time from 3-5 minutes to 8 seconds while keeping baseline cost within 12% of the non-resilient architecture. The warm-pool sharing algorithm and the approach to combining predictive scaling with reactive overflow represented novel contributions not found in published cloud architecture literature.

### Example 4: Novel Approach to Database Query Optimization for Graph Traversals

**The uncertainty:** A cybersecurity company needed to execute complex multi-hop graph traversals across a network topology dataset of 500 million edges within 2-second response windows. Existing graph databases (Neo4j, Amazon Neptune, JanusGraph) exceeded 30 seconds for the required traversal patterns because of the combination of traversal depth (6+ hops), filter complexity, and dataset scale. No known query optimization technique addressed this specific combination.

**The systematic approach:** The team developed and tested four optimization strategies: a custom graph indexing structure based on bloom filter hierarchies, a query decomposition algorithm that breaks complex traversals into parallelizable sub-queries, a caching layer using locality-sensitive hashing to identify similar traversal patterns, and a hybrid storage model combining adjacency lists with column-oriented edge attributes. Each strategy was benchmarked against 200 representative production queries with response time, memory utilization, and result accuracy as evaluation criteria.

**The advancement:** The bloom filter hierarchy index reduced traversal time from 30+ seconds to 1.4 seconds for 95% of query patterns. The underlying indexing technique, using probabilistic data structures to prune graph traversal branches before they are explored, represented a novel approach to graph query optimization that extended beyond existing published methods.

### Example 5: Computer Vision for Automated Defect Detection in Variable Environments

**The uncertainty:** A construction technology company developed a drone-based system to detect structural defects (cracks, spalling, delamination) in concrete infrastructure. Standard computer vision models (ResNet, YOLO variants) achieved high accuracy under controlled conditions but degraded severely (accuracy dropping from 94% to 52%) when deployed in real-world conditions with variable lighting, weathering, surface coatings, and camera angles. No published method addressed reliable defect detection across the full range of environmental variables encountered in field inspections.

**The systematic approach:** The team investigated domain adaptation techniques, synthetic data augmentation using physics-based rendering, multi-spectral image fusion (combining visible, thermal, and near-infrared imagery), a novel data augmentation pipeline replicating real-world environmental variance, and an ensemble approach combining specialized models for different environmental conditions. Each approach was tested against a field-collected dataset of 15,000 images from 47 different inspection sites captured across four seasons.

**The advancement:** The multi-spectral ensemble approach, combined with a custom augmentation pipeline that used physically accurate lighting simulation, achieved 91% accuracy across all field conditions. The physics-based augmentation approach for generating training data that accurately represented real-world environmental variance was a novel contribution to the applied computer vision domain.

### Example 6: Cryptographic Protocol for Privacy-Preserving Data Sharing

**The uncertainty:** A healthcare data company needed to enable multiple hospitals to collaboratively analyze patient outcome data without any party exposing raw patient records. Existing approaches (federated learning, homomorphic encryption, secure multi-party computation) each presented disqualifying limitations: federated learning leaked gradient information, fully homomorphic encryption was computationally infeasible at the required data scale, and generic MPC protocols had communication overhead that made them impractical across wide-area networks.

**The systematic approach:** The team designed and evaluated a hybrid protocol combining differential privacy with a custom secure aggregation scheme, a modified homomorphic encryption approach using domain-specific optimizations to reduce computational overhead, and a novel protocol combining secret sharing with function-specific computation that exploited the restricted set of statistical operations required. Each protocol variant was formally analyzed for privacy guarantees and benchmarked for computational and communication cost on datasets of increasing scale.

**The advancement:** The team developed a custom secure aggregation protocol that reduced communication overhead by 85% compared to generic MPC by exploiting the specific structure of the required statistical computations. The protocol maintained provable privacy guarantees while achieving practical performance for datasets up to 10 million records across 12 participating institutions.

---

## Manufacturing and Engineering SR&ED Examples

[Manufacturing SR&ED claims](/en/sred/manufacturing/) are among the most underutilized. Many manufacturers assume their process optimization or material experimentation is "just engineering," but if the work involves genuine uncertainty that could not be resolved using known techniques, it qualifies.

### Example 7: Welding Dissimilar Metal Alloys for Lightweight Automotive Components

**The uncertainty:** An automotive parts manufacturer needed to join aluminum alloy (6061-T6) to ultra-high-strength steel (UHSS, 1500 MPa tensile strength) for a structural vehicle component. The vast difference in melting points (660C vs 1,510C), thermal expansion coefficients, and the formation of brittle intermetallic compounds (Fe-Al IMCs) at the joint interface made conventional fusion welding methods produce joints that failed under cyclic loading at 40% of the target fatigue life. No published welding procedure specification (WPS) existed for this specific alloy combination at the required strength and fatigue performance.

**The systematic approach:** The team systematically evaluated friction stir welding (FSW) with tool geometry variations, laser-assisted friction stir welding with controlled heat input profiles, ultrasonic-assisted resistance spot welding, and a novel hybrid approach combining mechanical interlocking with adhesive bonding and selective laser ablation surface pre-treatment. For each approach, the team fabricated 30 test specimens, performed tensile testing, fatigue testing (10^6 cycles), cross-sectional metallographic analysis, and electron microscopy to characterize the intermetallic layer thickness and morphology.

**The advancement:** The hybrid mechanical interlocking approach with laser ablation surface pre-treatment produced joints achieving 112% of the target fatigue life by limiting the intermetallic compound layer to under 2 micrometers, significantly below the 10-micrometer threshold at which brittleness dominates. The specific combination of surface pre-treatment parameters and interlocking geometry for this alloy combination represented new knowledge in dissimilar metal joining.

### Example 8: Adaptive Automation System for High-Mix, Low-Volume Production

**The uncertainty:** A contract manufacturer needed to automate assembly of 340 different product variants running in batch sizes of 5-50 units. Traditional hard automation required dedicated tooling per variant (economically infeasible at these volumes), while collaborative robots with standard gripper systems could not handle the range of part geometries, weights (50g to 12kg), and surface characteristics (polished metal, textured plastic, porous ceramic). No commercially available automation system could accommodate this product variety without manual changeover times exceeding the production time saved.

**The systematic approach:** The team developed and tested a modular robotic cell with interchangeable end-effectors and automatic tool changing, a vision-guided system using 3D point cloud matching for part localization without part-specific fixtures, a soft robotic gripper using granular jamming with variable stiffness, and an AI-driven motion planning system that generated assembly paths from CAD models without manual programming. Each approach was tested against 50 representative product variants, measuring cycle time, defect rate, changeover time, and the percentage of variants handled without human intervention.

**The advancement:** The final system combined vision-guided localization with a novel variable-stiffness gripper and automated path generation from CAD data. It handled 94% of product variants with zero changeover time and achieved defect rates below 0.3%. The gripper design, which used a combination of granular jamming and pneumatic finger articulation to adapt to unknown geometries, represented a novel contribution to the robotic grasping problem for industrial applications.

### Example 9: Novel Biodegradable Polymer Formulation for Food Packaging

**The uncertainty:** A packaging manufacturer needed to develop a compostable food packaging material that met three simultaneous requirements: oxygen barrier performance within 15% of conventional PET, heat resistance to 100C for hot-fill applications, and complete composting within 180 days under industrial conditions. No existing biopolymer (PLA, PHA, starch blends) met all three requirements simultaneously. Blending approaches described in published literature achieved two of three targets but consistently failed on the third.

**The systematic approach:** The team designed a structured experimental program evaluating 23 polymer blend formulations based on PHA/PLA/cellulose nanocrystal combinations, with systematic variation of blend ratios, plasticizer types, and processing conditions. Each formulation was characterized for oxygen transmission rate (ASTM D3985), heat deflection temperature (ASTM D648), tensile properties (ASTM D638), and composting rate (ASTM D6400). A design-of-experiments approach was used to efficiently explore the formulation space.

**The advancement:** A PHA-PLA blend with cellulose nanocrystal reinforcement (8% by weight) and a proprietary reactive compatibilizer achieved all three targets. The oxygen transmission rate was within 11% of PET, heat deflection temperature reached 105C, and 92% disintegration occurred within 160 days. The specific compatibilization chemistry that enabled simultaneous barrier performance and compostability at this level was a novel contribution to biopolymer formulation.

### Example 10: Process Optimization to Reduce Scrap Below Industry Benchmarks

**The uncertainty:** A precision CNC machining shop needed to reduce scrap rate on a complex titanium aerospace component from 8.2% (already below the 12% industry average) to under 2%. The scrap was caused by thermal distortion during multi-axis machining of thin-walled features, and existing compensation strategies (thermal pre-compensation models, adaptive feed rate control) had already been implemented. Achieving sub-2% scrap required resolving residual thermal effects that were not accounted for by any published machining model for this alloy and geometry combination.

**The systematic approach:** The team developed a real-time thermal monitoring system using embedded thermocouples and infrared imaging, created a finite element thermal model calibrated to in-process measurements, designed a new machining sequence that distributed thermal load more uniformly across the part, and implemented a closed-loop system that adjusted cutting parameters in real time based on thermal deviation from the model prediction. Each intervention was tested on batches of 100 parts with full dimensional inspection.

**The advancement:** The closed-loop thermal compensation system reduced scrap rate to 1.1%, well below the 2% target. The thermal model, specifically calibrated for thin-walled titanium machining with real-time correction factors, represented new knowledge about thermal behavior in this specific manufacturing context that went beyond published machining science.

---

## Life Sciences and Biotech SR&ED Examples

[Life sciences and biotech companies](/en/sred/biotech/) frequently undertake work that qualifies for SR&ED, from novel drug delivery research to diagnostic assay development. The key is distinguishing experimental development from routine product testing.

### Example 11: Targeted Drug Delivery Using Engineered Lipid Nanoparticles

**The uncertainty:** A pharmaceutical company was developing a lipid nanoparticle (LNP) delivery system to transport an mRNA payload to hepatocytes (liver cells) with at least 70% targeting specificity. Standard LNP formulations (similar to those used in mRNA vaccines) showed less than 20% hepatocyte specificity, with the majority of nanoparticles accumulating in non-target tissues. Published literature described various surface-modified LNP approaches but none achieved the required targeting specificity without also reducing payload encapsulation efficiency below viable thresholds.

**The systematic approach:** The team designed a systematic screen of 48 LNP formulations varying ionizable lipid structure, PEG-lipid anchor length, targeting ligand density, and particle size. Each formulation was evaluated in vitro for encapsulation efficiency, stability, and cell-type-specific uptake using flow cytometry with fluorescently labeled payloads. The top 8 candidates were advanced to biodistribution studies in animal models, with tissue-specific quantification of mRNA expression using bioluminescence imaging and RT-qPCR.

**The advancement:** A formulation using a novel ionizable lipid with a branched tail structure and an optimized GalNAc targeting ligand density of 5 mol% achieved 76% hepatocyte specificity while maintaining 92% encapsulation efficiency. The relationship between lipid tail branching geometry and tissue-specific uptake represented new knowledge about LNP design principles.

### Example 12: Development of a Rapid Point-of-Care Diagnostic Assay

**The uncertainty:** A diagnostics company needed to develop a lateral flow assay that could simultaneously detect three respiratory pathogen biomarkers from a single nasal swab sample with sensitivity exceeding 95% and results delivered in under 15 minutes. Existing multiplex lateral flow designs suffered from cross-reactivity between detection lines and reduced sensitivity when multiple analytes competed for the same sample volume. No published multiplex lateral flow platform achieved the required combination of sensitivity and speed for this biomarker panel.

**The systematic approach:** The team investigated four approaches: a sequential flow architecture with timed release valves, parallel independent detection strips fed from a common sample pad, a novel conjugate particle design using spectrally distinct quantum dot labels to eliminate cross-reactivity, and a microfluidic-enhanced lateral flow hybrid that used capillary-driven metering to deliver optimized sample volumes to each detection zone. Each design was fabricated in prototype quantities and tested against a panel of 200 clinical samples with known pathogen status.

**The advancement:** The microfluidic-enhanced design with quantum dot conjugates achieved 96.2% sensitivity across all three analytes with a 12-minute time to result. The capillary metering system that precisely controlled sample volume delivery to each detection zone within a lateral flow format was a novel integration of microfluidic and lateral flow technologies.

### Example 13: Bioinformatics Algorithm for Rare Variant Detection in Low-Coverage Sequencing

**The uncertainty:** A genomics company needed to reliably detect rare somatic variants (allele frequency below 1%) from low-coverage whole genome sequencing data (average 10x coverage). Existing variant callers (GATK HaplotypeCaller, DeepVariant, Strelka2) had false-negative rates exceeding 60% for variants below 1% allele frequency at 10x coverage. Increasing sequencing depth to solve the problem was cost-prohibitive for the company's clinical screening application. No published algorithm addressed reliable rare variant calling at this combination of allele frequency and coverage depth.

**The systematic approach:** The team developed and benchmarked four algorithmic approaches: a deep learning model trained on synthetic variant data generated by in-silico mutagenesis, an ensemble method combining multiple existing callers with a trained meta-classifier, a novel error model incorporating instrument-specific and sequence-context-specific error profiles, and a Bayesian approach that integrated prior knowledge from population variant databases with sample-specific error characteristics. Each approach was evaluated against a truth set of 3,000 known variants spiked into real sequencing data at controlled allele frequencies.

**The advancement:** The Bayesian approach with context-specific error modeling reduced the false-negative rate to 18% for variants at 0.5-1% allele frequency, a dramatic improvement over the 60%+ baseline. The context-specific error model, which characterized error rates as a function of trinucleotide context, GC content, and instrument-specific parameters, represented new knowledge about computational approaches to variant detection under data-limited conditions.

---

## AI and Machine Learning SR&ED Examples

[AI and machine learning projects](/en/sred/ai-ml/) are among the fastest-growing sources of SR&ED claims in Canada. Developing novel model architectures, training approaches, or deployment methods all frequently involve genuine technological uncertainty.

### Example 14: Few-Shot Learning for Industrial Quality Inspection

**The uncertainty:** A manufacturing AI company needed to train defect detection models for new product lines using only 15-30 labeled defect images per class (versus the thousands typically required). Transfer learning from ImageNet-pretrained models achieved only 67% accuracy with this limited data. Data augmentation (rotation, flipping, color jitter) improved accuracy to 74%, still far below the 95% production requirement. No published few-shot learning method had demonstrated reliable industrial defect detection at these accuracy levels with such limited labeled data.

**The systematic approach:** The team developed and compared five approaches: meta-learning (Prototypical Networks, MAML) adapted for defect detection, a custom contrastive learning framework trained on unlabeled production images to build domain-specific feature representations, a generative approach using a conditional GAN to synthesize realistic defect images from few examples, a physics-informed augmentation pipeline that modeled the physical processes that create each defect type, and a hybrid combining contrastive pre-training with meta-learning fine-tuning. Each approach was evaluated across 8 product lines with leave-one-out cross-validation.

**The advancement:** The hybrid approach, contrastive pre-training on unlabeled production images followed by meta-learning fine-tuning on 20 labeled examples, achieved 96.1% accuracy across all product lines. The contrastive pre-training strategy, which learned defect-relevant features from the structure of normal production images without any labels, represented a novel approach to industrial few-shot learning.

### Example 15: Novel Reinforcement Learning for Robotic Manipulation in Unstructured Environments

**The uncertainty:** A robotics company needed to train a robotic arm to grasp and manipulate irregularly shaped objects in cluttered, unstructured environments, such as sorting recyclable materials from a mixed waste stream. Standard reinforcement learning approaches (PPO, SAC) trained in simulation failed to transfer to the real world due to the sim-to-real gap in contact dynamics and object interactions. Domain randomization improved transfer but still achieved only a 43% grasp success rate on novel objects, far below the 85% required for commercial viability. No published sim-to-real transfer method had achieved the required success rate for this class of manipulation tasks.

**The systematic approach:** The team developed and tested a progressive domain randomization strategy that adapted randomization parameters based on real-world performance feedback, a residual reinforcement learning approach that fine-tuned a simulation-trained policy with limited real-world interaction data, a novel tactile simulation model that more accurately modeled soft-body contact dynamics, and a hierarchical policy architecture that separated high-level grasp planning from low-level contact-rich manipulation. Each approach was evaluated on a standardized test set of 100 objects across 10 material categories with 50 grasp attempts per object.

**The advancement:** The hierarchical policy with residual real-world fine-tuning achieved an 89% grasp success rate on novel objects. The key innovation was the hierarchical decomposition that allowed the high-level policy to transfer from simulation while the low-level contact policy was efficiently fine-tuned with only 500 real-world interactions, representing a novel approach to sim-to-real transfer for contact-rich manipulation.

### Example 16: Deploying Transformer Models on Edge Devices with Strict Resource Constraints

**The uncertainty:** An IoT company needed to deploy a natural language understanding model on a microcontroller with 512KB RAM and no floating-point unit for a voice-controlled industrial device. Standard model compression techniques (pruning, quantization, knowledge distillation) applied to state-of-the-art transformer models could not reduce model size below 3MB while maintaining acceptable accuracy (target: 90% intent classification accuracy on the company's 47-class industrial command set). No published edge NLP deployment had achieved this accuracy level within these hardware constraints.

**The systematic approach:** The team investigated aggressive quantization to 4-bit and 2-bit integer representations, neural architecture search (NAS) to discover architectures optimized for the target hardware, a novel approach using binary neural networks for the attention mechanism combined with 4-bit representations for feed-forward layers, and a hybrid approach combining a compressed on-device model with selective cloud offloading for ambiguous inputs. Each approach was evaluated on the target hardware for accuracy, inference latency, memory footprint, and power consumption.

**The advancement:** The binary attention architecture with 4-bit feed-forward layers, discovered through a custom NAS procedure, achieved 91.3% accuracy with a 380KB model size and 45ms inference latency. The binary attention mechanism, which replaced standard floating-point attention weights with learned binary codes and a compact codebook, was a novel architecture contribution that enabled transformer-class models to run within microcontroller memory constraints.

---

## Clean Technology SR&ED Examples

[Clean technology](/en/sred/cleantech/) is a priority area for both the SR&ED program and broader Canadian innovation policy. Projects that involve developing new environmental technologies, improving energy efficiency, or reducing emissions frequently qualify.

### Example 17: Advanced Battery Management System for Second-Life EV Batteries

**The uncertainty:** A clean energy company needed to develop a battery management system (BMS) for repurposing retired electric vehicle batteries (with 70-80% remaining capacity) for stationary grid storage. The core challenge was that individual cells within retired battery packs had diverged significantly in capacity, internal resistance, and degradation trajectory during automotive use. Standard BMS algorithms designed for uniform cell populations could not safely manage packs where cell characteristics varied by 30-40%, leading to under-utilization of aggregate capacity (only 45% of theoretical capacity was usable) and premature shutdown. No published BMS algorithm addressed this degree of cell-to-cell variability.

**The systematic approach:** The team developed and tested an individual cell characterization protocol using electrochemical impedance spectroscopy to classify cells by degradation state, a dynamic cell grouping algorithm that reconfigured series-parallel connections based on real-time cell state, a predictive degradation model trained on aging data from 2,000 retired cells to forecast individual cell end-of-life, and a multi-objective optimization algorithm that balanced energy extraction against cell safety margins. Each approach was validated on a 100kWh pilot system assembled from retired EV modules, with 6 months of operational data collection.

**The advancement:** The dynamic reconfiguration system increased usable capacity from 45% to 78% of theoretical aggregate capacity while maintaining all cells within safe operating parameters. The predictive degradation model, which accurately forecast individual cell failure within a 10% time window, represented new knowledge about second-life battery degradation patterns that was not available from existing literature focused on first-life automotive applications.

### Example 18: Carbon Capture Using Novel Sorbent Materials

**The uncertainty:** A cleantech startup was developing a direct air capture (DAC) system targeting a cost below $200 per tonne of CO2 captured. Existing solid sorbent materials (amine-functionalized silica, metal-organic frameworks) required regeneration temperatures above 100C, making the energy cost of the capture-release cycle the dominant cost factor. The company hypothesized that a new class of moisture-swing sorbent materials could reduce regeneration energy by 60%, but no moisture-swing sorbent had demonstrated the required CO2 capture capacity (above 1.5 mmol/g) with sufficient cycle stability (less than 10% degradation over 1,000 cycles).

**The systematic approach:** The team synthesized and characterized 34 sorbent formulations based on quaternary ammonium functionalized polymer resins with systematic variations in polymer backbone chemistry, functional group density, pore structure, and hydration management. Each formulation was evaluated for CO2 capture capacity under ambient conditions, regeneration energy requirement, kinetics (capture and release rates), and cycle stability over 500 accelerated aging cycles. Materials meeting initial screening criteria were advanced to 1,000-cycle testing.

**The advancement:** A formulation using a cross-linked polystyrene backbone with optimized quaternary ammonium loading achieved 1.8 mmol/g capture capacity with only 6% degradation over 1,000 cycles and a regeneration energy 58% lower than thermal-swing alternatives. The specific cross-linking and functionalization chemistry that enabled both high capacity and cycle stability represented a novel sorbent design contribution.

### Example 19: AI-Optimized Building Energy Management

**The uncertainty:** An energy management company needed to reduce total building energy consumption by 30% beyond what conventional building automation systems (BAS) achieved, without occupant comfort degradation, in a commercial building portfolio with varied HVAC systems, building envelopes, and occupancy patterns. Standard model predictive control (MPC) approaches required building-specific physics models that took 3-6 months to calibrate per building and degraded in accuracy as building conditions changed. No published approach achieved the target savings with a practical deployment timeline across diverse building types.

**The systematic approach:** The team developed and tested a transfer learning approach that used data from previously optimized buildings to accelerate model training for new buildings, a model-free reinforcement learning controller that learned optimal HVAC strategies through interaction without requiring a physics model, a hybrid approach combining a lightweight physics model with a learned residual correction term, and an ensemble approach that maintained multiple control strategies and selected among them based on real-time performance. Each approach was deployed in a controlled A/B test across 12 buildings over 4 months, with energy consumption, thermal comfort (predicted mean vote), and indoor air quality as evaluation metrics.

**The advancement:** The hybrid physics-residual approach reduced deployment time from months to 2 weeks per building while achieving 33% average energy reduction across 12 buildings with no comfort complaints. The transfer learning mechanism that adapted the residual model across building types using only 2 weeks of site-specific data was a novel contribution to data-efficient building energy optimization.

---

## What Does NOT Qualify for SR&ED

Understanding what **does not** qualify is just as important as knowing what does. CRA rejects claims every year for including activities that fall outside the program's scope. These exclusions apply regardless of industry.

### Routine Engineering and Software Development

Building a product using established frameworks, languages, and patterns is not SR&ED, even if the product is new. Configuring a cloud infrastructure on AWS, building a CRUD application with React and Node.js, or integrating a third-party API are standard engineering tasks. The fact that your team has not done something before does not make it technologically uncertain if competent professionals in the field would know how to accomplish it.

### Market Research and Business Analysis

Conducting user interviews, analyzing market trends, performing competitive analysis, building financial models, or running A/B tests on pricing are business activities. They do not involve technological uncertainty in the SR&ED sense, even if they inform product development decisions.

### Style, Aesthetic, and UX Design Changes

Redesigning a user interface for better usability, changing color schemes, updating layouts, or improving visual design are not eligible. These are design decisions, not technological investigations. However, if the UX change requires solving a genuine technical problem (for example, rendering complex data visualizations in real time on mobile devices), the underlying technical work may qualify.

### Standard Quality Assurance and Testing

Running test suites, performing QA, conducting user acceptance testing, and fixing bugs are routine development activities. However, developing a fundamentally new testing methodology to address a testing problem that cannot be solved with existing approaches could qualify as SR&ED.

### Data Collection and Surveys

Gathering data, running surveys, and compiling statistics are not SR&ED activities on their own. However, developing a novel data collection instrument or a new analytical method to extract insights from data that resists analysis by existing methods could qualify.

### Commercial Production and Tooling

Manufacturing products for sale, setting up production lines, and building manufacturing tooling are not SR&ED, even if the product itself was developed through SR&ED-eligible work. The eligibility applies to the development phase, not the production phase.

> **The bright line:** If a competent professional in your field could accomplish the work by applying known techniques and existing knowledge, it is not SR&ED. The uncertainty must be real and technological, not merely a matter of effort, scale, or time. When in doubt, ask: *did we attempt something where the outcome was genuinely unpredictable using available knowledge?*

---

## How to Document SR&ED Projects

Strong documentation is the difference between a successful claim and a denied one. CRA expects contemporaneous evidence, records created during the project, not reconstructed after the fact.

### What CRA Wants to See

For each SR&ED project, maintain documentation that addresses:

- **The technological uncertainty** at the outset, what did you not know, and why could it not be resolved using standard methods?
- **The hypotheses formulated**, what approaches did you consider and why?
- **The work performed**, experiments, prototypes, iterations, and the technical details of each
- **The results and analysis**, what happened, what did you learn, and how did each result inform the next step?
- **The conclusions**, what advancement was achieved, or what knowledge was gained from the investigation?

### Practical Documentation Strategies

- **Commit messages and pull requests** -- For software projects, detailed commit messages and PR descriptions create contemporaneous technical records. Learn how [AI automates SR&ED claim preparation](/en/blog/ai-automates-sred-claims/) by extracting this evidence from your development tools automatically.
- **Lab notebooks and experiment logs** -- For physical science projects, dated lab notebooks remain the gold standard.
- **Design documents and technical specifications** -- Written before work begins, these capture the uncertainty and planned approach.
- **Meeting notes** -- Technical design meetings, sprint retrospectives, and architecture review notes all serve as evidence.
- **Test results and data** -- Raw experimental data, benchmark results, and analysis reports demonstrate systematic investigation.

For a comprehensive guide to the T661 form and filing process, see our [T661 form guide](/en/blog/t661-form-guide/) and the [complete SR&ED guide](/en/blog/what-is-sred-guide/).

---

## Frequently Asked Questions

### Does my project need to succeed to claim SR&ED?

No. SR&ED eligibility is based on the attempt to resolve technological uncertainty through systematic investigation, not on the outcome. If your team spent six months developing a new algorithm that ultimately did not meet performance targets, the work is still eligible as long as genuine technological uncertainty existed, a systematic approach was followed, and the goal was to achieve a technological advancement. Failed experiments that generate new knowledge about what does not work are explicitly eligible under the program.

### Can I claim SR&ED for work done by contractors or subcontractors?

Yes, but with important limitations. Payments to arm's-length contractors for SR&ED work performed in Canada can be claimed, but only 80% of the contract amount is eligible (the "proxy" reduction). The contractor must be performing eligible SR&ED work, not routine services. If the contractor is non-arm's-length (for example, a company owned by the same shareholders), the claim is limited to the contractor's actual SR&ED expenditures rather than the amount invoiced. All SR&ED work must be performed in Canada to be eligible.

### How far back can I claim SR&ED retroactively?

You can file SR&ED claims up to **18 months** after the end of the tax year in which the work was performed. For example, if your fiscal year ends December 31, 2025, you have until June 30, 2027 to file your claim. Many companies file their first claim covering the prior year's work after learning about the program. There is no requirement to have pre-registered or notified CRA in advance.

### What is the difference between SR&ED and IRAP?

[SR&ED](/en/sred/) is a retroactive tax credit, you claim it after the work is done, as part of your corporate tax filing. [IRAP](/en/irap/) is a prospective grant, you apply before the work begins and receive funding during the project. SR&ED is administered by CRA; IRAP is administered by the National Research Council. You can and should claim both, but you cannot claim the same expenditure under both programs. The portion of costs funded by IRAP is excluded from your SR&ED claim. [Learn more about how to stack IRAP and SR&ED](/en/blog/irap-vs-sred-comparison/).

### How much can I expect to receive from an SR&ED claim?

For Canadian-controlled private corporations (CCPCs) and eligible Canadian public corporations (ECPCs), the enhanced refundable rate is **35%** on the first $6 million in eligible expenditures, returning up to **$2,100,000** annually. Amounts above $6 million qualify for a 15% non-refundable credit. When combined with provincial SR&ED credits (which range from 3.5% to 30% depending on the province), total effective rates can exceed 60% in provinces like Quebec. [Use GrantOps to estimate your SR&ED refund](/en/application/).

### Do I need a consultant to file an SR&ED claim?

No, but many companies use one, especially for their first claim. SR&ED consultants typically charge 15-25% of the refund amount as a contingency fee, which significantly reduces your net benefit. An alternative is to use [GrantOps to automate SR&ED claim preparation](/en/automate-sred/). GrantOps connects directly to your development tools (GitHub, Jira, Slack) to extract R&D evidence and generate CRA-compliant documentation at a fraction of traditional consultant costs.

### Can I claim SR&ED for incremental improvements to an existing product?

It depends on whether the improvement involves technological uncertainty. Adding a new feature using standard development practices does not qualify. However, if the improvement requires solving a technical problem where the solution is not known in advance, for example, improving the performance of an algorithm by 10x when standard optimization techniques have already been exhausted, the work may qualify. The test is always whether competent professionals in the field would face genuine uncertainty in achieving the desired outcome.

### Are cloud computing and infrastructure costs eligible for SR&ED?

Cloud computing costs (AWS, GCP, Azure) used directly for SR&ED activities are eligible as materials consumed in the prosecution of SR&ED. This includes compute costs for training machine learning models, running simulations, or hosting development environments used for experimental development. However, cloud costs for production systems, staging environments used for routine testing, or general business operations are not eligible. You must be able to attribute specific cloud expenditures to specific SR&ED projects.

---

## Start Identifying Your SR&ED-Eligible Projects

If any of the examples in this guide resemble work your team has done in the past 18 months, there is a strong chance you have unclaimed SR&ED tax credits waiting to be recovered.

The average SR&ED claim for Canadian SMEs is worth **$150,000-$500,000 per year**. Companies with larger R&D teams regularly recover $1 million or more annually.

[GrantOps automates the entire process](/en/automate-sred/), from identifying eligible projects in your codebase and development tools to generating the technical narratives and financial schedules required for your T661 filing.

**[Start free — check your SR&ED eligibility in the app](https://app.grantops.ai/auth?mode=signup)**

You can also explore our industry-specific guides:
- [SR&ED for Software Companies](/en/sred/software/)
- [SR&ED for Manufacturing](/en/sred/manufacturing/)
- [SR&ED for Biotech and Life Sciences](/en/sred/biotech/)
- [SR&ED for AI and Machine Learning](/en/sred/ai-ml/)
- [SR&ED for Clean Technology](/en/sred/cleantech/)
- [What Is SR&ED? The Complete Guide](/en/blog/what-is-sred-guide/)

---

Try GrantOps: https://app.grantops.ai/auth?mode=signup (free to start, no credit card). Plans: Starter $99 CAD/month, Professional $499 CAD/month, plus an 8% success fee. Pricing: https://grantops.ai/en/pricing/