Interest and evidence are not the same
Artificial intelligence attracts more investor attention than any other sector we review. That attention is real, and it has translated into completed transactions. It has also produced a large number of companies whose public-market narrative rests on sector sentiment rather than on anything specific to the business.
Public investors distinguish between the two faster than private ones, and they do it by asking about ownership, dependency and margin.
What you own
The first question is what the company actually owns. Models trained in-house on proprietary data, with the weights and the data rights held by the entity, is one position. A product built on a third-party foundation model with a well-designed application layer is a different one — not a worse business, but a different investment case that must be described accurately.
Data rights receive particular attention. If training data was scraped, licensed under terms prohibiting commercial use, or obtained from customers without contractual permission, that is a disclosure issue and potentially a legal one.
Dependency and concentration
Where a company depends on a single model provider, a single cloud vendor or a single large customer, expect that dependency to be examined closely. The relevant question is what remains if terms change: what would happen to margin, to capability and to the product if pricing doubled or access were restricted?
A defensible answer usually involves either a genuine ability to substitute providers or a value layer that survives the substitution.
Margin after compute
Gross margin after inference cost is the metric that most often separates AI businesses that read well from those that do not. Companies reporting rapid revenue growth alongside compute costs growing faster are describing a problem, not a trajectory.
Public investors will calculate this whether or not it is presented. Presenting it, with the plan for how it improves at scale, is materially stronger than allowing it to be inferred.
Deployment, not pilots
Pilots, proofs of concept and letters of intent are not revenue, and they should never be presented in a way that suggests they are. Production deployments with named customers, defined contract terms and renewal history are.
Conversion rate from pilot to production is one of the most informative metrics an AI company can report — and one of the few that a sceptical investor will accept as evidence of product-market fit.
Where AI companies are consistently weak
In our experience, AI companies are strongest on technical capability and weakest on corporate foundations. The recurring issues:
- IP created by contractors without assignment clauses
- Training data with no documented provenance
- Revenue recognition on usage-based contracts that will not survive audit
- Capitalised development costs that an auditor will expense
- Capability concentrated in two or three researchers with no retention arrangements
All are addressable. None is addressed quickly once a process is under way.
Regulatory direction
AI regulation is developing in every major jurisdiction. Companies operating in health, credit, employment, biometrics or public-sector decision-making should expect questions about how they will meet obligations that are being defined now.
A company that can describe its position credibly is in a stronger place than one that has not considered the question.