Enterprise AI Budgets Soar as GenAI Becomes Essential and Off-the-Shelf Apps Eclipse Custom Builds

Budgets: AI Spend Surpasses Expectations and Is Here to Stay

Just over a year ago, enterprises began shifting how they build and buy generative AI. Since then, the landscape has changed dramatically. Recent conversations with over two dozen enterprise buyers and surveys of 100 CIOs across 15 industries reveal how AI spending and strategies are evolving for 2025 and beyond.

AI budgets have grown faster than anticipated. Enterprises have moved from pilots and innovation funds into making AI a recurring, permanent line item in IT and business budgets. Companies are now mixing multiple AI models to balance performance with cost. Market leaders include OpenAI, Google, and Anthropic, while Meta and Mistral remain popular open source options. Procurement now resembles traditional software buying, with more rigorous evaluations and hosting decisions. Complex AI workflows are increasing switching costs. Meanwhile, off-the-shelf AI applications are overtaking custom builds, rewarding AI-native third-party solutions.

1. Budgets Are Bigger Than Expected and Still Growing

AI budgets have grown beyond already high forecasts, with enterprises expecting around 75% growth in the next year. One CIO summed it up: “What I spent in 2023 I now spend in a week.” Growth is driven by discovering more internal use cases and rising employee adoption. Tech-forward companies are starting to focus on customer-facing AI, which could push spending even higher.

2. Generative AI Moves to Permanent Budget Lines

Innovation budgets once made up 25% of AI spending; now they’re just 7%. More AI expenses are covered by centralized IT and business unit budgets, signaling that AI is no longer experimental but essential. One CTO noted, “More of our products are adding AI enablement, so our spending growth will rise across all of these products.”

Models: Multi-Model Usage and Market Leadership

3. Multi-Model Strategy Is the Norm

Using multiple AI models in production is standard practice. Beyond avoiding vendor lock-in, enterprises buy different models because each excels in specific use cases. The share of companies using five or more models rose from 29% to 37% in a year. For example, Anthropic’s Claude is strong in coding and content generation, while OpenAI models shine in complex question-answering. This diversity encourages a best-of-breed approach and helps avoid lock-in.

4. Clear Leaders Are Emerging in a Crowded Market

• OpenAI leads overall market share.
• Google and Anthropic have gained significant ground.
• Open source models like Llama and Mistral see higher adoption in large enterprises focused on on-prem deployments.

OpenAI’s GPT-4o is the most deployed production model. Google’s Gemini 2.5 and Anthropic’s Claude Sonnet 3.5 models have helped these companies enter the market. Google’s cost-performance ratio stands out, with Gemini 2.5 flash costing 26 cents per million tokens compared to GPT-4.1 mini’s 70 cents. Anthropic is especially popular with tech-forward companies due to its coding strengths.

5. Closed Source Models Offer Better Price-to-Performance for Smaller Models

Model costs are dropping about tenfold every year. Closed source models like xAI’s Grok 3 mini and Google’s Gemini 2.5 Flash offer attractive price-to-performance ratios for small and medium workloads. This shift encourages more enterprises to pick closed source models, especially when integrated with their existing ecosystems.

6. Fine-Tuning Is Less Critical as Models Improve

With growing model intelligence and longer context windows, enterprises rely more on prompt engineering than fine-tuning. This approach reduces costs and avoids vendor lock-in. Fine-tuning remains relevant for very specific use cases, such as domain adaptation in video search. But overall, prompt engineering delivers similar results at a fraction of the effort.

7. Reasoning Models Are Early but Promising

Reasoning models expand the complexity of tasks LLMs can handle. While still early in testing, enterprises are optimistic about their potential. OpenAI’s reasoning models lead adoption, with 23% of surveyed enterprises using them in production. Other players like DeepSeek have lower adoption, mostly among startups.

Procurement: AI Buying Adopts Traditional Software Rigor

8. Buying Processes Become More Rigorous and Price-Sensitive

Model selection now follows structured evaluation frameworks. Security and cost have gained importance alongside accuracy and reliability. Enterprises carefully match models to use cases: top-tier models for critical applications and cost-effective options for simpler tasks. One leader said, “For most tasks, all models perform well enough now, so pricing has become a much more important factor.”

9. Hosting Preferences Vary but Trust in Providers Grows

Some enterprises prefer hosting models directly with providers or via platforms like Databricks, especially when their cloud provider doesn’t host the chosen model. Early access to the latest models is a key driver. Trust in direct hosting with providers like OpenAI and Anthropic has increased significantly compared to last year.

10. Switching Costs Increase with Complex AI Workflows

Initially, enterprises designed applications to minimize switching costs and treat models as interchangeable. Now, agentic workflows—multi-step processes requiring tailored prompts and guardrails—make switching more difficult and expensive. One leader noted the extensive engineering needed to maintain quality assurance when changing models.

11. External Benchmarks Help Narrow Model Choices

With more models available, enterprises use external benchmarks like LM Arena as an initial filter. Internal testing and employee feedback remain essential, but benchmarks provide a useful starting point similar to Gartner’s Magic Quadrant for software evaluation.

Rise of the App: Buying AI Applications Outpaces Building

12. Shift from Building to Buying AI Applications

Enterprises are increasingly buying third-party AI applications rather than building their own. The mature AI app ecosystem offers better ROI through continuous optimization and dedicated AI teams. Internally built tools often lack competitive advantages and require heavy maintenance. Over 90% of companies testing customer support AI apps are evaluating third-party options. Regulated industries like healthcare remain an exception due to data privacy concerns.

13. Outcome-Based Pricing for AI Apps Remains Challenging

CIOs are cautious about outcome-based pricing due to unclear metrics, unpredictable costs, and difficulty in attribution. Most still prefer usage-based pricing until clearer standards for measuring AI app value emerge.

14. Software Development Stands Out as a Killer Use Case

AI adoption in software development has surged, thanks to high-quality off-the-shelf tools and clear ROI. One CTO reported nearly 90% of their code is AI-generated with tools like Cursor and Claude Code, up from 10–15% a year ago. This trend signals strong growth potential across industries.

15. Prosumer Market Drives Early App Growth and Enterprise Demand

Consumer familiarity with AI brands like ChatGPT fuels enterprise adoption. Employees drive demand through personal use, influencing CIO decisions to purchase enterprise versions. This consumer-to-enterprise pull accelerates AI app growth.

16. AI-Native Vendors Outpace Incumbents in Quality and Speed

AI-native companies innovate faster and deliver better AI products than incumbents retrofitting older solutions. Buyers prefer AI-native vendors for their superior product quality and rapid feature delivery, especially in coding tools. User satisfaction data confirms this shift in preference.

Looking Ahead

Enterprise AI is no longer experimental. It’s now a strategic investment with clear budget commitments and structured procurement. Model diversity by use case is embraced, with leaders emerging in the market. Enterprises lean on external benchmarks and prefer buying AI applications over building them. The market is settling into a pattern similar to traditional software but with the fast pace and complexity unique to AI.