The bar is no longer “AI”
In 2023 a GPT wrapper with $20k MRR could raise a Seed. In 2025 it can’t. Investors have seen enough wrappers die when OpenAI shipped the same feature for free, or when a competitor built on the same foundation model and undercut on price. The default verdict on a thin AI layer today is pass — not because AI isn’t interesting, but because a feature built on someone else’s API isn’t a company.
What replaces it: proprietary data, proprietary distribution, or proprietary workflow. If you don’t own at least one of those three, you’re a feature. The question every serious investor now asks in the first meeting is: “What happens to your business if OpenAI ships this for free in 12 months?” If your answer is “we’d have to pivot,” the meeting is already over.
What investors actually want to see
For an AI startup raising Seed or Series A right now the standard ask is:
- A wedge use case where AI changes the unit economics by 5–10x, not 20%. A 20% improvement is a feature; a 5x improvement is a product.
- Some form of data moat: proprietary customer data that improves the model with usage, fine-tuned models on corpora a foundation model can’t access, or a workflow that captures structured feedback the underlying model doesn’t see.
- Real usage — not just signups. Weekly active retention above 40% of activated users is the new “traction.” A waitlist is not a business.
- A clear answer to “what happens when GPT-6 ships this for free?” — and the answer cannot be “we’ll move faster.” It needs to be structural: switching costs, network effects, embedded workflow, or proprietary training data.
- At least one enterprise contract above $50k ACV, proving you can sell beyond the prosumer market.
Valuations are split in two
The AI funding market is bimodal. Frontier labs and infrastructure plays — OpenAI, Anthropic, Mistral, Scale AI, Nscale — take the megarounds at valuations that defy conventional analysis. Everyone else competes for a much smaller pool at much more disciplined multiples. The middle is effectively dead: “AI for X” without defensibility now raises at the same revenue multiples as a normal SaaS company — roughly 8–12x ARR at Seed, not 30–50x.
The due diligence has gotten technical
Investors in 2025 are doing real technical DD on AI claims. “Built with proprietary AI” gets unpacked. Investors now routinely ask to see model architecture, fine-tuning methodology, evals against baseline models, and data provenance. Founders who can’t answer these questions in detail are flagged as execution risks even if the product demo looks impressive.
What this means for you
Before you go out to raise, stress-test your pitch against one question: “If a frontier lab released a free feature tomorrow that did 80% of what we do, would customers still pay us?” If the honest answer is no, fix the moat before you fix the deck. The deck is the easy part.
Frequently Asked Questions
Q: What metrics do AI startups need to raise a Series A in 2025? A: AI companies pitching Series A in 2025 typically need $3M–$8M+ ARR, a burn multiple below 1.5, net revenue retention above 110%, weekly active usage above 40% of seats sold, gross margins above 70%, and at least one enterprise contract above $100k ACV. The bar is materially higher than 2022–2023 levels when $2M ARR was sufficient.
Q: What is a “data moat” and why do AI investors require it? A: A data moat is proprietary training or feedback data that improves an AI model’s performance in ways a competitor using the same foundation model cannot replicate — for example, five years of structured clinical notes from hospital partners, or logged human corrections from 10,000 annotators in a specialized domain. Investors require it because AI products without a data moat can be replicated by any team with API access to the same foundation model.
Q: At what valuation multiple are non-infrastructure AI startups raising in 2025? A: Application-layer AI startups without clear defensibility are raising at 8–12x ARR at Seed — similar to traditional SaaS multiples — rather than the 30–50x multiples seen in 2021–2022. Infrastructure and frontier model companies continue to raise at outlier multiples, but they represent a small fraction of deal volume.
Q: How do investors conduct technical due diligence on AI startups? A: AI-focused investors in 2025 routinely request model architecture documentation, fine-tuning methodology and training data provenance, evals against baseline models, and technical interviews with lead engineers. Founders who cannot clearly explain their model’s differentiation from a foundation model are flagged as execution risks regardless of product demo quality.
Q: What does “proprietary workflow” mean as an AI defensibility moat? A: A proprietary workflow moat means the AI is embedded so deeply into a customer’s operating process that switching requires retraining staff, migrating data, and replacing integrations — creating switching costs that persist even if a competitor builds a technically superior model. CrackTheDeck’s pitch deck review workflow is an example: the combination of AI scoring, human expert feedback, IC simulation, and valuation calculator creates a multi-step engagement that a single-feature AI tool cannot replicate.
CTA: Run your AI pitch through CrackTheDeck and see how it benchmarks against real 2025 AI rounds — in metrics, valuation and defensibility.