At COLLIDE 2026, Actian CTO Emma McGrattan argued that AI stalls on old data problems that need a new fix: treating data as a product and governing it with data contracts. She shared survey results linking company-wide data products to AI in production and a one-project way to start.
Key takeaways
- The problems are old. McGrattan said data access, quality, silos, governance and literacy gaps are decades-old challenges now blocking AI.
- Treat data as a product. Reusable, discoverable data assets with business and data owners and quality guarantees.
- Contracts operationalize governance. A data contract sets what is delivered, how often, for what use and at what quality.
- Fix data at the source. Shift-left quality avoids repeating the same fixes in the warehouse.
- Start with one project. One high-impact AI use case, one contract and one data product build confidence to scale.
The problems keeping enterprise AI stuck in pilots are not new, according to Emma McGrattan, Chief Technology Officer at Actian. Data access, data quality, silos, weak governance and low data literacy have challenged organizations for decades. In a keynote at COLLIDE Data + AI Conference 2026 in Atlanta, McGrattan argued that what is new is the fix: treating data as a product and backing it with data contracts.
McGrattan, who said she has been at Actian for nearly 34 years, opened with figures she said keep chief data officers awake at night. She cited an MIT study finding that 95% of generative AI initiatives deliver zero ROI, and said fewer than 10% of organizations have scaled AI agents, defined as three or more in production.
Why do enterprise AI projects fail to scale?
People still struggle to find the data they need and get it to the right people on time, McGrattan said. Many can't trust the data they have. And fragmentation is severe: she cited a Gartner estimate that the average enterprise has more than 600 data silos, which makes a clear view of a single customer hard to assemble.
Data literacy is a gap too. McGrattan, who grew up in Ireland, said she learned only two years ago that a 51% chance of rain means a 100% chance of rain over 51% of the forecast area. If a data professional can misread something that simple, she asked, how well do business users understand the dashboards they're given?
The deeper problem, she said, is that companies treat data as a byproduct of their applications. "You can't treat data as a byproduct. Treat data as a product," she said.
What is the shift-left approach to data quality?
Many of Actian's customers fix data in their warehouses, McGrattan said, and have to repeat those fixes every time a new source or feed arrives. The shift-left principle means fixing problems at the source, such as cleaning Salesforce data so the warehouse doesn't have to.
She illustrated the cost of waiting with a story from her early years in California. A DMV clerk recorded her height of 64 inches as 6 feet 4 inches. The error would have taken a few keystrokes to fix at the counter, but she didn't correct it until a traffic stop led to a court date and a lawyer. "Fixing the problem with your data at the source is what we really think is important," she said.
What are data products and data contracts?
A data product is a reusable, discoverable data asset, such as a data set or pipeline, that multiple teams can use, McGrattan said. Each one has an owner. Actian recommends two: one from the business and one from the data team. Data products also carry quality guarantees set to the use case. Financial data might need to be 100% accurate, while a marketing campaign might still run if 90% of its email addresses are wrong or duplicated.
A data contract is "a formal agreement between the data team and the business team," McGrattan said, covering what is delivered, how often it is refreshed, its use case and the quality thresholds it must meet. Because the contract sets the terms under which data is delivered, it is how organizations operationalize governance. That matters more when AI systems are making decisions autonomously and at machine speed, she said.
Do data products improve AI success rates?
To test its hypothesis, Actian commissioned a European research firm to survey 300 data leaders worldwide, including CDOs, VPs of data, data architects and ML engineers, McGrattan said. Among the findings she shared:
- 69% of organizations have data products in operational use, up from 48% the year before
- 93% of self-identified data leaders use data products
- 61% of organizations use data contracts, a figure she said is rising quickly
- 85% of companies with AI projects in production use data products company-wide, compared with 25% of those without data product maturity
- For agentic and autonomous AI, 77% of companies with successful deployments use company-wide data products, compared with 23% of those still experimenting
Asked why they adopt data products, respondents pointed to trust. "It's all about trust," McGrattan said, whether the decisions are made by people or by AI.
How should organizations get started with data products?
McGrattan pointed to two Actian customers. Statnett, Norway's national grid operator, found that its manual processes depended on experienced people who handled ambiguity well, something machines do poorly. It put more discipline around its data with data products and added a data literacy program. ACC Automotive Cells, a French maker of EV batteries that grew through acquisitions, wanted a common business glossary across plants that each used their own language, along with formal data contracts, data products and data quality work.
Her advice for getting started: pick one AI project with real business impact, identify the data it needs, define a data contract covering refresh frequency, format and the meaning of its terms, and publish the data product. Narrowing the scope builds the data team's confidence that it can take a project from pilot to production, McGrattan said, and builds the business's confidence in the data team. Each success makes the next data product easier.
See keynotes like this at COLLIDE 2027
COLLIDE, the largest AI conference in the Southern U.S., brings more than 1,000 enterprise data and AI leaders to Atlanta for keynotes, panels and technical sessions. The 2027 conference takes place Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center, with registration coming soon. Data Science Connect's free webinars cover data foundations for AI throughout the year.
Frequently asked questions
What is a data product?
As McGrattan described it, a reusable, discoverable data asset such as a data set or pipeline, with named owners and quality guarantees, that multiple teams can use.
What is a data contract?
A formal agreement between the data team and the business team covering what data is delivered, how often it is refreshed, its use case and the quality thresholds it must meet.
Do data products help AI reach production?
In an Actian-commissioned survey of 300 data leaders, 85% of companies with AI in production used data products company-wide, compared with 25% of those without data product maturity, McGrattan said.
What is shift-left data quality?
Fixing data problems at the source system, such as a CRM, instead of repeatedly correcting them in the data warehouse.
Watch: Keynote: Why AI Fails Without Data Products and Data Contracts on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights
