What HMRC Means by an Advance in Science or Technology in EdTech
EdTech sits in a category that HMRC's technical assessors find genuinely difficult: the product delivers an educational outcome (not a scientific or technological one), but the engineering behind it may involve real technical challenges. The BIS Guidelines on the Meaning of R&D for Tax Purposes apply without modification: the work must seek an advance in science or technology, and face genuine uncertainty that a competent professional could not resolve from existing knowledge.
For EdTech this means HMRC draws a sharp line between the educational content or pedagogical design of a product (not qualifying) and the technical systems that deliver, assess, or personalise it (potentially qualifying). A new online course, a curated video library, and a teacher-facing dashboard built on standard components do not qualify. A novel adaptive sequencing algorithm, a machine-learning-based diagnostic engine, or a technically uncertain natural-language marking system may qualify strongly, provided the technical challenge is articulated clearly and is genuinely separate from curriculum design.
The relevant SIC codes for EdTech companies are typically 62010 (computer programming) or 85590 (other education). The SIC code does not determine eligibility; the qualifying activity test does.
What Qualifies as R&D in EdTech
Adaptive Learning and Knowledge Modelling
Adaptive learning systems are one of the strongest qualifying categories in EdTech, but only where the adaptation itself required genuine technical development. Qualifying examples include: building a Bayesian knowledge tracing or item response theory (IRT) model to estimate learner mastery at scale, where existing open-source implementations were insufficient for the required latency, accuracy, or item-bank size; developing a reinforcement-learning policy for instructional sequencing that was experimentally derived because published policies underperformed on the target learner population; engineering a real-time knowledge-graph update system that maintains learner state across sessions and devices without latency that degrades the learning experience; and developing novel cold-start approaches for adaptive systems where a new learner's state must be inferred with minimal data.
Automated Assessment and Marking Engines
Automated assessment is a strong claiming category where the technical challenge is in the marking or scoring algorithm, not just the interface. Qualifying examples include: building a natural language processing system to score short-answer or essay responses against a rubric with validated inter-rater reliability comparable to trained human markers; developing computerised adaptive testing (CAT) engines that select items in real time from large item banks to minimise test length while maintaining measurement precision; and engineering automated feedback systems that diagnose specific misconceptions or error types from learner responses with greater granularity than pattern-matching on common errors.
Speech and Language Technology for Education
Speech recognition and synthesis systems customised for educational use, particularly for early literacy, language learning, or accessibility, frequently qualify. Qualifying examples include: adapting or fine-tuning speech recognition for non-native accents or child speech where standard commercial engines underperformed to a degree that made the product undeliverable; developing pronunciation assessment models that identify specific phonemic errors with actionable feedback; and building low-latency audio pipelines for real-time conversation practice in a learning context.
Educational Data Infrastructure and Privacy-Preserving ML
Data infrastructure qualifies where it required novel technical work. Privacy-preserving machine learning (federated learning, differential privacy) applied to educational data to enable model training without centralising sensitive learner data is a genuine area of technical uncertainty. Novel data models for representing learning interactions, assessment data, and competency frameworks in ways that existing learning record store (LRS) specifications did not accommodate may also qualify.
What Does NOT Qualify: EdTech Anti-Patterns
These are the most common non-qualifying items in EdTech claims:
- Curriculum and content development. Writing lesson plans, producing video content, authoring question banks, and designing learning pathways are pedagogical work, not R&D, regardless of the effort involved.
- Standard LMS configuration and integration. Setting up Moodle, Canvas, or a third-party learning management system, including building integrations via documented APIs, is not qualifying R&D.
- UI/UX design and A/B testing. Choosing button colours, testing copy variants, and optimising onboarding flows are product and marketing work, not technical advance.
- Applying existing ML models without modification. Using a pre-trained language model via an API for content recommendation or chatbot responses, without novel fine-tuning or architectural work, is not qualifying R&D.
- Teacher or learner analytics dashboards built on standard BI tools. Connecting a database to a charting library to display engagement metrics is routine development, not an advance in science or technology.
- Compliance with accessibility or data protection standards. Implementing WCAG 2.1 compliance or GDPR controls is not R&D unless it required technically novel approaches beyond documented best practice.
Qualifying Costs for EdTech Under the Merged Scheme
Staffing costs. UK-payroll engineers, machine learning scientists, and NLP researchers who work directly on qualifying technical projects. EdTech product managers are typically excluded unless they are working directly on the technical design of a qualifying system. Content producers, teachers, and learning designers are excluded entirely.
Subcontractors and EPWs. UK-based machine learning contractors, academic NLP researchers engaged on qualifying work, and data-annotation services engaged as part of a qualifying ML development project are potentially claimable. Overseas contractors are excluded under the merged scheme's UK-only rule. The 65% haircut applies to unconnected subcontractors.
Cloud and data costs. Cloud compute for ML model training, data storage for learner interaction datasets used directly in qualifying R&D, and data licences for training corpora all became claimable from April 2023. Production serving of the live EdTech platform does not qualify. See our qualifying expenditure guide for full cost-category details.
Software licences. Development tool licences, ML framework licences, and annotation platform costs used directly in qualifying R&D are claimable, apportioned to R&D use.
ERIS for EdTech
Many EdTech companies are loss-making in their early years, particularly those building technically complex platforms before they have scaled subscription revenue. Where qualifying R&D expenditure exceeds 30% of total expenditure and the company is loss-making, ERIS applies at 27%. For a loss-making EdTech company with £700,000 of qualifying R&D spend out of £2m total (intensity 35%), the ERIS credit is £189,000 versus £140,000 under the standard merged-scheme rate. The ERIS credit is payable in cash, typically within 40 working days of a complete CT600.
Worked Example: An Adaptive Assessment Platform
A UK EdTech company (SIC 62010) builds a computerised adaptive testing platform for secondary-school maths. Total expenditure for the year to 31 March 2026 is £1.9m. The company is loss-making. It employs 9 software engineers, 2 machine learning researchers, 3 curriculum specialists (excluded from claim), and 4 commercial staff (excluded). A specialist review identifies:
- £420,000: salaries for qualifying technical staff at 70-100% R&D time apportionment.
- £85,000: cloud compute for IRT model training and adaptive engine testing.
- £38,000: UK academic subcontract for psychometric validation (65% of £58k invoiced).
Total qualifying spend: £543,000. R&D intensity: £543k / £1.9m = 28.6%. Just below the 30% ERIS threshold, so the standard merged scheme applies.
Credit at 20%: £543,000 x 20% = £108,600. Net of corporation tax (nil for loss-making company): most of this is payable as cash via surrender of losses, subject to the PAYE cap calculation. Had the company crossed 30%, the ERIS credit would have been £146,610 under the 27% rate.
HMRC Enquiry Risks for EdTech
HMRC's main concern in EdTech claims is the inclusion of non-technical staff and non-technical work. Claims that include curriculum designers at high R&D time apportionments, or that describe the educational impact of a feature rather than the technical uncertainty behind it, are the primary enquiry triggers. A second common issue is cloud cost allocation: claiming all compute spend without separating R&D development environments from production serving. A third is the treatment of ML model costs: HMRC expects to see that the model development involved genuine uncertainty, not just calling an API. An experienced adviser will frame the technical narrative at the correct level and exclude non-technical cost lines.
What to Do Next
If your EdTech company has engineers working on genuinely novel technical systems, the claim is worth assessing. The eligibility checker takes five minutes. For a full review, request a free assessment from an HMRC-registered specialist. Related guides: SaaS R&D tax credits, SaaS platform engineering R&D, merged scheme overview, and qualifying expenditure categories.