Most venture funds in Central Europe have finance and investment people on the team, not AI engineers. When the deal is an AI or NLP startup, that gap matters more than it used to. The distance between uses AI and has AI has widened sharply in the last two years. A founder can integrate an LLM API in a weekend; building something defensible takes longer and looks different. We help investors tell the two apart.
We work with early-stage funds in Central Europe and the partners who back them.
What we look for
We work through a written checklist that adapts to the target company. The questions we keep returning to, in roughly the order they matter:
- Is the product a real system or a wrapper around someone else's API? What happens to the company if pricing or terms change at the upstream provider?
- Where does the training and evaluation data come from — is it licensed, scraped, user-generated, synthetic? Is the licensing defensible under the EU AI Act?
- How is the model evaluated? Is there a held-out test set, a real evaluation harness, regression tracking? Or is "it looks good in demos" the standard?
- What happens at scale — inference cost trajectory, latency under load, accuracy degradation on out-of-distribution input, cold-start exposure?
- Who owns the IP? Have contractor and freelancer assignments been done correctly? This matters more in Central Europe than in most regions, where distributed teams are the norm.
- How deep is the team? Is there a single person who would take the technology with them if they left? Are the founders complementary, or do they all have the same gap?
- Where does the company sit under the EU AI Act — high-risk, limited-risk, prohibited? Is the documentation in place?
- For LLM-based systems specifically: prompt injection exposure, RAG evaluation quality, hallucination rate, guardrails, observability.
The full checklist is published openly. See below.
How we work
Three formats, depending on stage and ticket size.
Red-flag review. One to three days. A focused look at the most failure-prone parts of the technology. Used for early conviction calls, or when you have a deadline. Delivered as a short memo with red, amber, or green flags by category, and a verbal debrief with the partner on the deal.
Standard AI due diligence. Five to ten days. Includes founder and engineering interviews, architecture review, model evaluation methodology audit, data and IP review, team assessment. Delivered as a structured report with an executive summary and a 100-day remediation plan.
Deep AI due diligence. Two to four weeks. For Series A and later, where code review, MLOps inspection, and a full data licensing audit are warranted. Delivered as a long-form report suitable for investment committee presentation.
What you get
A written report, formatted for an investment committee, with an executive summary on page one — red, amber, or green by category — and the underlying analysis behind it. A verbal debrief with the partner on the deal. Written follow-up answers to any questions the IC raises. A 100-day remediation plan if you choose to invest, structured so the founders can act on it without further interpretation.
Who we are
Crow Intelligence is led by Zoltán Varjú, an NLP engineer and architect with two decades in the field and an exit (Complytron, acquired by SEON), and Orsolya Putz, PhD in cognitive linguistics and adjunct professor at the Technical University of Budapest. We have been working in NLP since before the term became fashionable.
Our open due-diligence checklist
The full checklist we work from is published openly on GitHub. You can use it directly, fork it for your own process, or send it to your portfolio companies as a self-assessment.
github.com/crow-intelligence/ai-dd-checklist