Let’s play Do it or Ditch it with a16z’s Enterprise AI Report
"Do it or ditch it" breakdown for AI strategy to help you stay nimble while building stronger SR&ED claims.
A16z's latest enterprise AI report reveals 5 key habits of large companies adopting AI. Here we analyze which of these enterprise strategies actually make sense for startups—and which ones could hurt your innovation claims.
A16z published a report a month ago about how CIOs and R&D execs at enterprise companies are approaching their AI inclusion. You can read it here. Here's our very quick summary—with the SR&ED lens.
Habit: Use Multiple AI Models for Different Tasks
Do it or ditch it: Do it—at your scale.
Why: Enterprises are moving away from a “one model fits all” approach, instead deploying different models for coding, writing, and business strategy (e.g., Anthropic for code, Gemini for system design, OpenAI for Q&A). Founders - experiment with several models to find the best fit for each use case, but avoid overcomplicating—pick 2-3 that serve your immediate needs, not 5+.
SR&ED angle: Multi-model experimentation creates some systematic investigation activities. If they go beyond vanilla integration, take note since those are highly claimable. Document your comparative analysis, performance benchmarks, and integration challenges.
Habit: Fine-Tune Every Model for Custom Needs
Do it or ditch it: Ditch it (for now).
Why: Off-the-shelf models are getting smarter and context windows continue to grow. My guess is that for 90%+ of your use cases, prompt engineering on off-the-shelf models will deliver results close to fine-tuning, with less cost and complexity. Fine-tune only if you hit a wall with prompts.
SR&ED angle: Here's the counterintuitive part—fine-tuning can actually be better for SR&ED claims than prompt engineering. Fine-tuning can often involve systematic experimentation with training data, hyperparameters, and model architectures that create clear technical uncertainties. Prompt engineering, while valuable, is often too routine to be strongly claimable. If you do fine-tune, document everything: your hypothesis, training iterations, performance metrics, and technical challenges overcome.
Habit: Aggressively Expand AI Budgets
Do it or ditch it: Ditch it.
Why: Enterprises are ramping up their AI spend (75% YoY growth), but this is driven by their scale and use cases. Founders - focus on high-ROI experiments and keep budgets tight. Since AI pricing goes up with usage, the more you use, the more you dip into your limited kitty.
Habit: Shift to Off-the-Shelf and Third-Party AI Tools
Do it or ditch it: Do it.
Why: Enterprises are buying more third-party AI apps instead of building everything in-house, especially for non-core needs. Founders - leverage proven tools to save time and resources, reserving custom builds for true differentiators.
SR&ED critical point: This is where most founders hurt their claims without realizing it. Using established AI tools for routine tasks (like basic chatbots or standard image recognition) isn't claimable—there's no technical uncertainty. However, integrating these tools into novel workflows, customizing them for unique use cases, or combining multiple tools to solve unprecedented problems can be highly claimable. The key is proving you're advancing the state of the art, not just implementing existing solutions.
Habit: Build Complex Agentic Workflows Early
Do it or ditch it: Ditch it (for now).
Why: Enterprises are redesigning around AI agents, but deep integration makes switching hard and increases complexity. Startups should keep workflows simple and modular until there’s a proven need for advanced automation.
SR&ED angle: Here's where the tax implications get interesting. Simple, modular approaches may be better business strategy, but complex agentic workflows often create more claimable technical uncertainties. If you do build agentic systems, document the systematic investigation around agent coordination, decision-making logic, and error handling. The technical challenges around "how do we make multiple AI agents work together effectively" are rich SR&ED territory.
Are you in the AI consulting space?
If so, CIOs remain wary of outcome-based pricing for AI projects due to unclear metrics and unpredictable costs, preferring usage-based and seat-based models. If you’re a service provider, consider incorporating a hybrid pricing structure that includes the latter.
Copy the habits that keep you nimble, cost-efficient, and focused on real customer value. Skip the ones that only make sense at enterprise scale or add unnecessary complexity. But always remember—the way you implement these strategies can make or break your R&D tax credit claims.