The Due Diligence Gap
Venture capital firms and growth equity investors are deploying billions into AI companies. The due diligence process typically covers: team background, market size, competitive landscape, financial projections, and a technical overview of the AI approach.
What it almost never covers: the quality of the data pipeline, the rigor of the evaluation framework, the depth of the technical debt, and the realistic scalability of the architecture. These are exactly the factors that determine whether an AI company can deliver on its promises.
Having conducted technical due diligence on AI companies for both investors and acquirers, we see the same blind spots repeatedly.
Blind Spot 1: Demo Quality vs. Production Quality
Every AI company has a great demo. The demo uses curated data, controlled scenarios, and often a different model or prompt than the production system. The gap between demo quality and production quality is the most common source of post-investment disappointment.
What to check: Ask to see the production system handling real customer queries in real time. Compare the output quality to the demo. Request access to production monitoring dashboards. If the company resists, that is a signal.
Blind Spot 2: Data Pipeline Fragility
An AI company's model gets all the attention. The data pipeline that feeds it gets almost none. But pipeline fragility is the most common cause of production incidents and quality degradation. Pipelines built for a demo or early customers often break under scale, especially when data sources change format, volume, or availability.
What to check: How many data sources feed the system? What happens when a source is unavailable? How is data quality monitored? When was the last pipeline incident, and how long did it take to resolve?
Blind Spot 3: Evaluation Rigor
"Our model achieves 95% accuracy" is a claim we see in almost every AI pitch deck. The question investors rarely ask: accuracy on what test set, measured how, and validated by whom? Many AI companies evaluate their models on datasets they created, using metrics they defined, with no external validation.
What to check: How large and representative is the evaluation dataset? Was it created by the company or by independent domain experts? How often is it updated? What is the performance on the hardest 10% of cases, not just the average?
Blind Spot 4: Customer Dependency Concentration
AI companies often achieve impressive results for their first few customers because they invest heavily in customization for those accounts. The question is whether those results generalize. A company that achieved great outcomes for 3 customers by spending 6 engineering months per customer does not have a scalable business.
What to check: What is the onboarding effort per customer in engineering hours? Has that effort decreased with each new customer? What percentage of engineering time is spent on customer-specific work vs. platform improvement?
Blind Spot 5: Realistic Cost Structure
AI companies often project unit economics based on future model costs (which they expect to decrease) and future efficiency improvements (which they expect to materialize). Current unit economics are frequently negative, especially for companies that rely on frontier model APIs.
What to check: What is the current cost per customer query, including all compute, API costs, and engineering support? How does this compare to customer revenue? What specific assumptions drive the projection that costs will decrease?
The Better Due Diligence Framework
For investors evaluating AI companies, we recommend adding four assessments to standard due diligence:
- Production quality assessment: Independent evaluation of the live system with realistic data
- Architecture review: Assessment of technical debt, scalability limitations, and pipeline robustness
- Evaluation audit: Independent review of how the company measures model performance
- Unit economics deep dive: Bottom-up cost model based on actual system architecture and usage patterns
These assessments cost a fraction of the investment amount and prevent the most common sources of value destruction in AI investments. The investors who conduct them consistently make better decisions.