How to Evaluate an AI Company Before You Invest

The AI investment landscape is crowded with companies that describe themselves as “AI-powered” or “AI-first.” For investors, the challenge is distinguishing between companies with genuine, defensible AI capabilities and those that have added a thin layer of AI marketing to a fundamentally traditional business.

This framework provides a structured approach to evaluation.

Evaluating the Technical Moat

The first question is whether the company has a defensible technical advantage. In AI, moats typically come from one of three sources: proprietary data, proprietary models, or proprietary processes.

Data moats are the most durable. A company that has accumulated a unique dataset through years of operation — customer interactions, sensor data, transaction histories — has an advantage that is difficult and time-consuming for competitors to replicate. Data moats strengthen over time as the dataset grows.

Model moats are increasingly fragile. As foundation models become more capable and open-source alternatives proliferate, the value of custom model architecture has diminished. A model moat only exists if the model’s performance cannot be replicated using publicly available models and the company’s data.

Process moats come from the operational systems that surround AI — the MLOps infrastructure, the feedback loops, the domain expertise embedded in feature engineering. These are often undervalued but can be highly defensible.

Assessing the Team

AI companies are talent-dependent to an unusual degree. The quality of the ML team is a direct predictor of the company’s ability to execute, innovate, and maintain its technology.

Look for depth, not just credentials. A team of three experienced ML engineers who have shipped production systems is more valuable than fifteen data scientists with PhDs who have never deployed a model outside a Jupyter notebook.

Evaluate key-person risk carefully. If the CTO or lead ML engineer left tomorrow, what would happen? Is the knowledge documented? Could the remaining team maintain and improve the systems?

Data Asset Evaluation

Data is the most undervalued and most misunderstood asset in AI companies. Evaluate data along four dimensions.

Provenance: where does the data come from, and does the company have clear legal rights to use it for training ML models?

Quality: is the data clean, labelled, and structured for ML workloads, or does it require significant preprocessing?

Scale: is there enough data to train effective models, and is new data being generated continuously?

Exclusivity: does anyone else have access to this data, or is it genuinely proprietary?

Market Position and Commercialisation

Technical capability without commercial traction is a research project, not a business. Evaluate how the AI capability translates into revenue.

Is the company selling the AI directly (AI-as-a-service), embedding it in a product, or using it to improve internal operations? Each model has different unit economics, scalability characteristics, and competitive dynamics.

Assess the sales cycle and customer concentration. AI products often have long sales cycles in enterprise markets and high switching costs once adopted. This can be an advantage or a risk depending on the stage of the business.

Financial Metrics for AI Companies

Standard financial metrics apply, but with AI-specific adjustments. Pay attention to compute costs as a percentage of revenue — these can be substantial for companies running large models. Evaluate the gross margin trajectory: are compute costs scaling linearly with revenue, or are there efficiency gains as the model serves more customers?

R&D spend should be contextualised against the competitive landscape. Underspending on research in a rapidly evolving field is a risk. Overspending without commercial traction is also a risk.

Risk Factors

The most common risks in AI investments include: commoditisation of the core technology as open-source alternatives improve; regulatory exposure under evolving AI legislation; customer concentration; key-person dependency; and the fundamental challenge that AI capabilities are difficult to value using traditional methods because they are probabilistic, rapidly evolving, and often poorly documented.