What HMRC Means by an Advance in Science or Technology in Professional Services Tech
LegalTech, AccountingTech, and InsurTech all build software for regulated industries with domain-specific requirements. That domain specificity can create genuine technical challenges, but HMRC's BIS Guidelines on the Meaning of R&D for Tax Purposes require the advance to be in science or technology, not in the application of professional expertise. The test is whether a competent software engineer, NLP researcher, or data scientist could resolve the uncertainty from existing published knowledge.
This is a sector where the non-technical work often dwarfs the technical work in effort and cost. A LegalTech company building a contract review tool spends more on product management, legal expert annotation, and client delivery than on the core NLP model. An InsurTech building a motor pricing platform spends more on actuarial design, data procurement, and regulatory compliance than on the ML pipeline. HMRC's test applies only to the technical layer. The professional services expertise embedded in the product does not itself qualify, but the engineering required to build a system that delivers it at scale or accuracy levels not previously achievable may qualify strongly.
Relevant SIC codes include 62010 (computer programming) for LegalTech and InsurTech product companies, 66290 (other activities auxiliary to insurance) for InsurTech, and 69201 (accounting) or 74909 (other professional services) for AccountingTech. The SIC code does not determine eligibility; the BIS Guidelines test does.
What Qualifies in LegalTech, AccountingTech, and InsurTech
LegalTech: NLP, Contract Analysis, and Knowledge Systems
LegalTech generates strong qualifying claims where the underlying NLP or knowledge engineering involved genuine technical uncertainty. Examples include: developing novel document-understanding models for legal contracts where general-purpose models failed to meet accuracy, recall, or latency requirements; building novel clause-extraction or obligation-identification systems trained on proprietary legal datasets with architecture choices made experimentally; engineering novel graph-based knowledge representation systems for legal reasoning or precedent retrieval at scale; developing novel cross-reference and citation extraction systems for legislation or case law where existing NLP pipelines underperformed; and building novel multi-label classification systems for legal document categorisation where the taxonomy size, label imbalance, or document length created technically uncertain modelling challenges.
InsurTech: Pricing Models, Fraud Detection, and Risk Systems
InsurTech qualifying work tends to focus on ML model development and data engineering at the frontier of actuarial and statistical practice. Qualifying examples include: developing novel causal inference models for risk pricing that go beyond standard GLM and gradient boosting approaches where the causal structure of the risk required experimental model development; building novel real-time fraud detection systems using graph neural networks or sequence models on claims or transaction data where published architectures underperformed on the specific fraud patterns observed; engineering novel telematics data processing pipelines for UBI pricing where the signal processing, trip segmentation, and feature extraction required experimental determination; developing novel survival models or credibility estimation approaches for thin-data segments where standard actuarial methods were insufficiently precise; and building novel document fraud detection systems for claims processing with technically uncertain image or text classification accuracy.
AccountingTech: Document AI and Reconciliation
Accounting automation is a large market with well-established off-the-shelf solutions, which means HMRC's competent-professional test is particularly important here. Qualifying work includes: developing novel document understanding models for invoice, receipt, or purchase order parsing that achieve technically uncertain accuracy on the diversity of document formats encountered in practice; building novel entity resolution and matching systems for accounts payable or accounts receivable where standard fuzzy-matching and deduplication approaches were insufficient; engineering novel reconciliation algorithms for complex group accounting, multi-currency, or intercompany-elimination scenarios where standard approaches required novel computational work; and developing novel anomaly-detection systems for audit or fraud purposes trained on proprietary accounting transaction datasets.
RegTech and Compliance Automation
Regulatory technology (RegTech) is a cross-sector category spanning all three verticals. Qualifying work includes: developing novel natural language processing systems to detect regulatory changes in legislation, FCA publications, or FRC guidance that affect a company's obligations; engineering novel obligation-to-control mapping systems using knowledge graphs or other structured reasoning that required experimental development; and building novel change-impact classification systems that predict which internal controls a regulatory change affects, where the accuracy and completeness requirements exceeded what rule-based and standard ML approaches could deliver.
What Does NOT Qualify: Professional Services Tech Anti-Patterns
These are the most common false positives in LegalTech, AccountingTech, and InsurTech claims:
- Applying known LLMs or ML frameworks via API without novel development. Calling an OpenAI, Google, or AWS API to process legal, insurance, or accounting documents is integration work, not R&D, even if the prompt engineering is sophisticated.
- Professional services delivery using the technology. Lawyers reviewing contracts, accountants producing reports, and underwriters pricing risks using the software are delivering professional services. That work does not qualify as R&D.
- Standard data warehousing and BI. Loading insurance or accounting data into a data warehouse and building dashboards using standard BI tools is not qualifying R&D.
- Regulatory compliance and legal work. Obtaining FCA authorisation, drafting terms of business, and managing PII under UK GDPR are compliance activities, not R&D.
- Training data annotation without technical development. Having lawyers or accountants annotate documents to build training datasets is data labelling. It is a necessary precursor to qualifying ML development but is not itself qualifying R&D.
- Standard CRM and case management system configuration. Configuring Salesforce, ServiceNow, or a legal case management system using documented configuration options is not qualifying R&D.
Qualifying Costs for Professional Services Tech Companies
Staffing costs. Software engineers, ML engineers, NLP researchers, and data scientists working directly on qualifying technical projects, apportioned by time. Legal, actuarial, and accounting professionals are claimable only for the time they spend directly supporting qualifying technical R&D, such as providing expert annotation in a structured R&D context, not for their professional delivery work.
Subcontractors. UK-based NLP researchers, actuarial data scientists, and academic NLP labs engaged on qualifying work are claimable at 65% for unconnected parties. Overseas parties are excluded under the merged scheme's UK-only rule.
Cloud and data costs. Cloud compute for ML model training and data processing directly in qualifying R&D, and data licences for training corpora (legal databases, insurance claims data, accounting transaction datasets) licensed specifically for R&D use, are claimable from April 2023. Production serving costs are excluded. See our qualifying expenditure guide for the full rules.
Software licences. Development tools, ML frameworks, and annotation platforms used directly in qualifying R&D, apportioned to R&D use. See the glossary for definitions of qualifying expenditure and the PAYE cap.
ERIS for LegalTech, AccountingTech, and InsurTech
Early-stage LegalTech and InsurTech companies building technically complex products before they have scaled commercial revenue frequently meet the ERIS intensity threshold. Where qualifying R&D expenditure exceeds 30% of total expenditure and the company is loss-making, ERIS at 27% applies. For a pre-revenue LegalTech company with £500,000 of qualifying R&D out of £1.4m total (intensity 36%), ERIS returns £135,000 versus £100,000 under the standard merged scheme.
Worked Example: A LegalTech NLP Company
A UK LegalTech company (SIC 62010) builds a contract review platform using proprietary NLP models. Total expenditure for the year to 31 March 2026 is £1.6m. The company is loss-making. It employs 8 software and ML engineers, 2 NLP researchers, 4 legal and commercial staff (excluded). A specialist review identifies:
- £390,000: qualifying engineer and researcher salaries at 75-100% R&D apportionment.
- £70,000: cloud compute for model training and experimentation on qualifying NLP projects.
- £35,000: legal data licence for proprietary contract corpus used directly in qualifying R&D model training.
- £28,000: UK university NLP subcontract for named entity recognition architecture work (65% of £43k).
Total qualifying spend: £523,000. R&D intensity: £523k / £1.6m = 32.7%. Company is loss-making and above the 30% ERIS threshold.
Credit under ERIS: £523,000 x 27% = £141,210 payable as cash. Under the standard merged scheme at 20%: £104,600. ERIS adds £36,610 in cash recovery.
HMRC Enquiry Risks for Professional Services Tech
HMRC's main concerns in LegalTech and InsurTech claims are: legal or actuarial staff included at high R&D time apportionments without evidence that their time was spent on technical development rather than professional delivery; data annotation work presented as qualifying R&D rather than as a precursor to it; cloud cost claims that include production model serving alongside qualifying development and training; and project narratives that describe the business problem ("we needed to review contracts faster") rather than the technical uncertainty ("our clause extraction model achieved below 80% recall on sub-100-word clauses and we did not know whether a transformer-based approach would resolve this"). An experienced adviser will build the narrative around the technical uncertainty, not the business outcome.
What to Do Next
If your LegalTech, InsurTech, or AccountingTech company has engineers and researchers working on genuinely novel NLP, ML, or data systems, the claim is worth assessing. The eligibility checker takes five minutes. For a full review, request a free assessment. Related guides: SaaS R&D tax credits, SaaS platform engineering R&D, merged scheme overview, and qualifying expenditure categories.