Artificial intelligence holds immense potential, but many projects fail before they even take off. Across our industrial projects built with knowledge graphs, three obstacles come up systematically and block the path to scale.
The curse of incomplete data
The most frequent problem in industrializing AI is the lack of complete, reliable data. On the development of an algorithm predicting the cost and lead time of aircraft modifications, feasibility was demonstrated on a limited dataset, but the lack of exhaustive data prevented scaling. The time data scientists spend chasing data blocks industrialization and slows internal validation.
Structure data from the start
Building a business ontology, a map of dependencies and relationships between data, before even launching the proof of concept delivers three things.
- +Standardize data collection.
- +Speed up proof-of-concept velocity with test datasets.
- +Ease access to real data at industrialization time.
Uncoordinated data exploration
Without centralization, the search for data becomes chaotic. After a fire at a factory supplying titanium screws for the A320, six different teams explored separate datasets and produced inconsistent results, each with its own sources, without coordination.
A shared knowledge foundation
A knowledge-driven system, whether a marketplace of graphs or of agents, directs teams to reliable sources, lets them build on work already done in other departments, and reduces errors while speeding up industrialization.
The myth of unrealistic expectations
Finally, many projects fail on a lack of realism. Enthusiasm for AI leads to outsized expectations: we've had a client ask that an AI automatically generate a Figma mockup from a UX study. Theoretically possible, but at a cost rarely justifiable.
The challenge is explaining what AI can do efficiently and economically, by clearly delimiting the boundaries of what's possible.
Master the foundations before taking off
The success of an AI initiative doesn't depend solely on the algorithm or the data scientist's talent, but on the maturity of the project's foundations. The failures we've seen aren't technical: they're organizational and methodological, a mix of unreliable data, unclear governance, and a gap between ambitions and economic reality.
For AI to deliver on its promises, you first have to solve the data and governance problem, and move from a craft-based stage to an industrial, realistic approach.