Accenture chief executive Julie Sweet says only 5% of organisations have the enterprise AI foundations needed to make AI repeatable at scale. The company’s view is that the hard work is less about finding a bigger model and more about building the data, infrastructure and controls around it.
Accenture AI Watch analysis
What happened
In an interview with Diginomica, Sweet said Accenture has worked on more than 18,000 advanced AI projects since late 2022, and grown its AI and data workforce from 40,000 to nearly 110,000. Yet, she said, just 5% of organisations have the enterprise AI stack needed for repeatable use and the prospect of return on investment.
Accenture chief AI and data officer Lan Guan described that stack as including data and knowledge, models, an agentic core, governance, security and safety, scaled operations and a user interface, all built on robust infrastructure. For an airline AI companion, Guan said, the company had to connect customer data and clean up lengthy policy and procedure documents because an off-the-shelf model did not know the airline’s specific information.
Accenture also says only 10% to 20% of enterprise workflows justify frontier-model capabilities. Its advice is to match models to the work and consider open-weight models where organisations want more control over their data. Read the Diginomica interview.
Why it matters
The figures put a useful gap between AI pilots and dependable, organisation-wide systems. If data is messy, access rules unclear or infrastructure unfinished, a powerful model alone will not make an operation repeatable. And if only a fraction of workflows warrant frontier models, paying for the most capable option everywhere may be a costly way to learn that lesson.
Our read
Accenture has a commercial interest in selling the consulting, platforms and infrastructure that make this “enterprise stack” sound essential. Even so, its practical point is worth taking seriously: the model is one ingredient, not the whole kitchen. The 5% figure is the company’s assessment, not an independent census, but it gives organisations a sharper question than “Which model should we buy?” Start with one workflow, identify the data and permissions it needs, and work out whether a cheaper model will do.
What to watch
- Whether Accenture publishes the method behind its 5% readiness estimate.
- Whether its reported AI projects translate into measured customer returns.
- Whether organisations increasingly reserve frontier models for the workflows that need them.
Discussion spark: Should organisations invest first in stronger data, governance and infrastructure, or keep experimenting with models until a clear use case earns that investment?
Sources and evidence
not affiliated with or endorsed by Accenture AI