AI TRANSFORMATION
AI product development costs: evidence and planning considerations
Key takeaways
- The widely cited 55.8% result came from one controlled JavaScript HTTP-server task. It is not a full-lifecycle productivity measure.
- Gartner's 25% to 30% figure is a 2028 forecast. McKinsey's figures estimate potential R&D value rather than realised savings.
- Public adoption and spending signals show momentum, but they do not provide a reliable price for an enterprise AI product.
AI product development cost depends on product scope, data readiness, integration, assurance and the operating model after launch. Published research can inform assumptions, but task experiments, forecasts, economic models and adoption surveys must remain separate.
AI development productivity evidence
| Evidence | Finding | What it measures |
| Microsoft Research and GitHub, 2023 | Developers with GitHub Copilot completed the task 55.8% faster than the control group. | A controlled experiment in which recruited software developers implemented an HTTP server in JavaScript. It did not measure discovery, architecture, testing, release or production operation. |
| Gartner, published 2025 | AI could drive 25% to 30% productivity gains across the software development lifecycle by 2028. | A strategic forecast. It is not an observed cross-industry result, and Gartner advises leaders to look beyond task-level time savings. |
| McKinsey, 2023 | GenAI could deliver productivity value equivalent to 10% to 15% of overall R&D costs. | An estimate of economic potential across product R&D activities, not a promised budget reduction. |
A delivery plan should therefore use task evidence to test individual activities and an internal baseline to estimate programme-level cost.
AI R&D investment and economic potential
| Evidence | Finding | Scope |
| McKinsey, 2025 | $360 billion to $560 billion in potential annual economic value from using AI to accelerate R&D. | Modelled potential across industries representing about 80% of large corporate R&D expenditure. It is not realised revenue or savings. |
| WIPO, 2025 | Almost $1.3 trillion in 2024 corporate R&D expenditure. | Data available for 1,510 of the world's top 2,000 corporate R&D spenders. WIPO notes that this does not represent smaller firms. |
| OECD, 2026 | Inflation-adjusted R&D spending in its panel rose 6% in 2025, up from 3% in 2024. | A panel of 60 large companies representing close to 46% of R&D reported by the top 2,000 global investors. Growth was led by digital industries. |
These findings describe the scale and direction of R&D investment. They do not reveal what an individual software product should cost.
Enterprise AI adoption and spending signals
| Evidence | Signal | Boundary |
| Menlo Ventures, 2025 | Estimated enterprise GenAI spending of $37 billion in 2025, up from $11.5 billion in 2024. | Market-sizing model informed by a survey of 495 US enterprise AI decision-makers. It excludes chips, inference and model serving, and AI features embedded in existing software. |
| Microsoft, FY2025 Q4 earnings call, July 2025 | Microsoft reported 20 million all-time GitHub Copilot users. | A cumulative vendor-reported total from an investor disclosure, not a monthly or daily active-user measure. It does not establish paid usage, delivery depth or productivity. |
| Deloitte, 2026 | Surveyed organisations expanded sanctioned AI access from fewer than 40% to around 60% of workers during 2025. | Survey of 3,235 business and IT leaders across 24 countries and six industries. Access does not establish regular use or business return. |
| Forrester, May 2024 | 67% of surveyed AI decision-makers planned to increase GenAI investment within the following year. | Investment intent from a survey whose sample size is not disclosed in Forrester's public summary. It is not observed spending or adoption. |
Building an AI product cost baseline
Generic solution-price matrices hide the variables that usually determine enterprise cost. A credible estimate starts with the delivery evidence below.
| Cost driver | Evidence to establish before estimating |
| Product and workflow scope | Users, decisions, channels, exception paths and measurable acceptance criteria. |
| Data and integration | Source quality, access constraints, system interfaces, migration and retrieval requirements. |
| Assurance and governance | Risk tier, evaluation method, security controls, human approvals and audit requirements. |
| Production operation | Expected volume, model choice, latency, observability, support ownership and change frequency. |
Estimate the first release and the operating run rate separately. Then test the highest-risk assumptions before funding the wider roadmap.
Sources
- Microsoft Research (2023). The Impact of AI on Developer Productivity: Evidence from GitHub Copilot.
- Gartner (2025). How to Capture AI-Driven Productivity Gains Across the SDLC.
- McKinsey & Company (2023). The Economic Potential of Generative AI.
- McKinsey & Company (2025). The Next Innovation Revolution Powered by AI.
- WIPO (2025). Global Innovation Index 2025, Global Innovation Tracker.
- OECD (2026). Tracking Business R&D in Real Time.
- Menlo Ventures (2025). The State of Generative AI in the Enterprise.
- Deloitte (2026). The State of AI in the Enterprise.
- Forrester (2024). Generative AI investment-intention summary.
- Microsoft Investor Relations (2025). Fiscal Year 2025 Fourth Quarter Earnings Conference Call.
