Striim's Jake Bengtson and Google Cloud's Naveen Punjabi say AI stalls in production when data is stale, pulled straight from operational systems or stripped of business meaning. They outlined four signs your data can support AI agents at scale.
Key takeaways
- Data must reflect current state. Agents acting on a week-old snapshot make business decisions on stale information.
- Derive data, don't hit operational systems. Non-deterministic agents put unknown load on source databases, so replicate data instead.
- Preserve business meaning. Without shared definitions, two agents may answer the same revenue question differently.
- Become event-driven first. Syncing every 15 minutes to an hour won't keep up as agent interactions multiply.
- Share semantic ownership. Lines of business supply the context, and a central team keeps definitions consistent and current.
Many enterprises assume they are ready for AI once models are chosen and pilots are running. In the Data Science Connect webinar "Four Signs Your Data Is Actually Ready for AI," Jake Bengtson, VP of AI Solutions at Striim, and Naveen Punjabi, Director, Analytics and GenAI Partnerships at Google Cloud, argued that the data foundation is usually what holds AI back, and laid out four tests for whether it is ready.
The session was sponsored by Striim. Brian Mink, co-founder and president of Data Science Connect, introduced the speakers.
Why do AI projects stall between pilot and production?
Bengtson said many organizations have run successful proofs of concept, only to see initiatives fail when they move into the workflows employees actually use. He cited a 2025 MIT report on the state of AI in business that found few task-embedded AI deployments returning value, and listed reasons: AI that lacks persistent memory, tools that sit outside the workflows people use, outputs that lack business context, and organizations that deploy AI "like software, not like a capability."
He compared adopting AI to onboarding a new employee. There is a ramp-up period, he said, and the payoff comes after a few months of teaching. Punjabi focused on context. "An AI model, if you think about it, is only as smart as the context it has right now," he said. If a supply chain breaks or customer behavior shifts, a model that doesn't know about it has a gap.
What makes legacy data architecture a barrier to AI?
Punjabi cited a Gartner prediction that more than half of business decisions will be automated or augmented by AI agents by 2027, covering areas such as pricing, inventory, routing and customer service. Agents making split-second decisions on data that is 24 hours old, he said, "invariably will lead to bad decisions."
He named data silos, platforms that lock customers in, and cost among the barriers. Operational databases need to stay in sync with the analytics platform where agents run, and data should sit in open, interoperable formats so business definitions don't change when data moves between systems. He described Google Cloud's approach as open data formats, a universal catalog of technical and business metadata, a choice of query engines, and AI built into the stack.
Bengtson described a maturity path he sees at Striim customers: first move to some form of cloud, then realize that a warehouse built five years ago for a weekly report won't support operational AI, then build real-time pipelines and event-driven processing. A pilot can run on a point-in-time snapshot, he said, but in production "that data needs to be coming together live in real time all the time." He added that traditional machine learning, such as predictive maintenance, needs the same streaming foundation.
What are the four signs your data is ready for AI?
1. Data represents current system state. If the analytical layer is a snapshot from a week ago, Bengtson said, agents will make business decisions on stale information. Punjabi said models need grounding in current data and continuous monitoring for drift after deployment. Every organization has access to the same foundation models, Bengtson added, so the differentiator is the ability to bring in its own data in real time.
2. Data is operationally derived, not operationally accessed. Because agents are non-deterministic and may write many queries to reach an answer, they put an unknown load on source systems. Any database administrator will respond, Bengtson said, "No, you cannot connect directly to my operational database." Punjabi said connecting an agent to a database's MCP server works in some cases, but it doesn't give access to the broader data set or its business context. Replicating changing data from databases and SaaS applications in real time, he said, gives richer context, and each approach has its place.
3. Data preserves business meaning as it changes. The semantic work data teams did for business analysts is "the exact thing the context that AI needs," Bengtson said, and it has to be preserved as data moves through a streaming pipeline. Punjabi gave the example of revenue meaning one thing in SAP and another in Salesforce. Without a shared definition, two agents may answer the same revenue question differently.
4. Data is event-driven by nature. Bengtson described agents as "a natural extension of microservices," and said advanced organizations publish events, such as a sign that a customer may churn, so agents can act without a human kicking off the work. Punjabi said many companies that call their data real time are actually syncing every 15 minutes to an hour. With agents, he said, customer interactions with systems may grow tenfold or a hundredfold, so teams should become event-driven first.
Who should own the semantic layer?
Punjabi said the business context lives within lines of business, so building it takes close collaboration between data infrastructure teams and business users, in a hub-and-spoke model. Infrastructure teams also need to be agile. "If it's going to take 15 days, 20 days to come together on a definition," he said, "it's too slow for business."
Bengtson said leaving it to lines of business alone produces "a hodgepodge" with no accountability, so the best organizations he has seen do both: line-of-business ownership plus a central team that keeps everyone accountable and revisits definitions as the business changes. He warned against shortcuts. Adding semantic meaning to prompts instead of a platform, he said, is "going to be a failing strategy."
Where to keep learning about AI-ready data
Data Science Connect has 24 free live webinars on its 2027 schedule across four series: the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. See the webinar schedule for upcoming sessions. Registration for 2027 events is coming soon.
Frequently asked questions
What are the signs that data is ready for AI?
According to Striim and Google Cloud, AI-ready data represents current system state, is derived from operational systems rather than queried directly, preserves business meaning as it changes, and is event-driven.
Should AI agents query operational databases directly?
Bengtson advised against it, since agents can issue an unknown number of queries and add load; Punjabi said MCP access works in some cases but lacks the broader data set and business context.
Why do AI agents need real-time data?
Agents make split-second decisions on pricing, inventory and customer service, and Punjabi said agents running on day-old data will make bad decisions.
Can you add business context to AI through prompts?
Bengtson warned that putting semantic meaning in prompts instead of a governed platform is a failing strategy.
Watch: Four Signs Your Data Is Actually Ready for AI on YouTube
Event: Webinars · More coverage: Webinars insights
