Source: Unite.AI
Sathya AG is a technology evangelist, published author of “Enterprise-Grade Hybrid and Multi-Cloud Strategies“, and Senior Principal Architect at Google (Retail Strategic Industries), with over 19 years of experience in enterprise architecture and AI. A global thought leader recognized with the Top AI Influencer Award and Google’s President’s Award, he is the author of Enterprise-Grade Hybrid and Multi-Cloud Strategies and an Advisory Board Member for the CAIO Circle. He is a Fellow of the British Computer Society (FBCS), Senior Member of IEEE, and Stanford LEAD alumnus. A sought-after speaker at AI Forward, Google Next, and NRF, Sathya also volunteers to teach AI/ML to underserved communities.
Disclaimer: The thoughts shared below and opinions expressed are the author’s own and do not necessarily reflect the views, positions, or opinions of his employer or any organization he is affiliated with.
You spent more than a decade at Oracle (ORCL ) before moving into a senior architecture role at Google, where you now work with large retail enterprises. How has your view of enterprise AI changed as companies moved from traditional data modernization projects to generative AI and agentic AI initiatives?
Back in the Oracle days, as an enterprise architect, my core focus was always about building the foundational, highly complex systems that organizations rely on to run their business. We were solving massive integration, governance, and Master Data Management challenges to ensure the enterprise had a single, highly resilient source of truth. Traditional data modernization was fundamentally about passive observation, designing semantic layers and data pipelines that delivered trusted insights to human decision-makers.
The evolution of Generative and Agentic AI has completely changed the architectural paradigm from passive observation to an autonomous system of action. Now, the lens has shifted to whether we can let a system take an action on our behalf (and, correctly, without supervision), at scale. That’s a much higher bar for data quality, context, and control because now a bad data assumption doesn’t just produce a wrong dashboard, it produces a wrong action. An agent that mis-reads inventory data doesn’t just misinform someone, it ends up placing a wrong order.
This is where rigorous enterprise architecture becomes critical and foundational to AI systems. To safely build autonomous systems, we have to seamlessly bridge the gap between deterministic enterprise systems and probabilistic AI. The architecture conversations I lead now are about building a trusted data fabric, ensuring that real-time autonomous reasoning is bound by strict governance, robust observability, and the same transactional integrity we have always demanded of enterprise systems.
You have argued that many AI projects fail not because of the model, but because of data and architecture decisions made much earlier. What are the most common early architecture choices that quietly undermine AI projects later?
The one I see most is companies treating a warehouse or lake as the single source of truth when it’s really just a single destination. They pipe everything into one place and assume that solves the ‘one version of truth’ problem, but it doesn’t, it just centralizes the disagreement. If three source systems each define an active customer differently, funneling them into one warehouse gives you one wrong answer instead of three.
Then there’s the trap of building for exactly what’s in front of you. A pipeline gets tuned perfectly to one dashboard, works great, and then six months later a new model needs the same data faster or fresher and the whole thing has to be torn up rather than extended.
And the one that always comes back to bite people is skipping lineage and metadata because it looks like overhead nobody asked for. It has no payoff; until the day someone asks why the model made a specific call and there’s no way to trace it. That’s when it turns into an expensive fix instead of a cheap design decision.
When working with Fortune 500 enterprises, what signals tell you whether an organization is truly ready to scale AI beyond pilots?
Honestly, I can usually tell within the first few conversations. The difference between an organization experimenting with AI and one ready to operationalize it at enterprise scale comes down to a few telltale signs. I usually look for readiness across three main buckets: Strategic Conviction, Business Integration, and Operational Maturity.
Here is what signals true readiness to scale:
1. Strategic Alignment & Executive Conviction
A Clear AI Strategy, Not Just AI Envy: The organization’s leadership demonstrates genuine conviction regarding their AI strategy. This is easily spotted by asking a few leading questions about how AI maps to their business objectives. If they can articulate the precise business value (e.g., revenue growth, margin improvement, or customer experience) rather than just wanting to “use generative AI,” they are ready.
Committed Enterprise Funding: They have moved beyond isolated R&D budgets. There is a dedicated, cross-functional budget allocated not just for AI projects, but for change management, infrastructure, and ongoing MLOps.
2. Problem-Solution Fit & Business Integration
Solving the Right Problems: A major red flag is confusing standard workflow automation or robotic process automation (RPA) with AI, and force-fitting AI on top just to show AI adoption to the board. Ready organizations understand the distinct value proposition of AI. They focus on use cases where AI offers a paradigm shift in capability, rather than just checking a technology box on a legacy process.
Business Unit Ownership: This is a quieter signal, but often the greatest impediment to scaling. If AI excitement lives entirely within an isolated “innovation lab” with no skin in the game from the actual business units, pilots will fail to take off. True readiness is signaled when business stakeholders actively co-develop the solution and own the final business outcome, rather than throwing a proof-of-concept over the wall to operations.
3. Operational Maturity & Governance
Day-One Security & Compliance Engagement: A lack of involvement from security and governance teams, or trying to cordon off a critical AI use case from them to “move faster,” is a recipe for deployment failure. Mature organizations bring InfoSec, legal, and data governance to the table on Day 1. They view these teams as critical enablers who build the guardrails necessary to scale safely, not as roadblocks.
Ultimately, organizations that successfully scale AI treat it not as an IT science experiment, but as a transformational business capability with the right sponsorship, the right guardrails, and the right business alignment.
What does an AI-ready data foundation actually look like beneath the surface, especially for enterprises with fragmented data across ERP, CRM, commerce, cloud, and legacy systems?
Most enterprises don’t have a data volume problem rather they have a data trust problem. This is what gets exacerbated with AI layered on top. It’s less about having a shiny new platform and more about a handful of unglamorous properties holding true across systems.
For example, the same customer, product, or store must be recognized as the same entity, regardless of whether it’s sitting in a legacy ERP, a modern CRM, or a commerce platform. If your systems can’t agree on who a customer is, your AI personalization and forecasting models are working off fractured data resulting in a hallucinated picture. You can’t predict customer lifetime value if your AI treats one customer as three different people.
Second, data freshness that’s matched to the decision, not defaulted to whatever the source system happens to do. Nightly inventory updates are fine for long-range planning and useless for real-time fulfillment; a real foundation supports both speeds without flattening everything to the slowest one.
Third, a lineage layer that lets you answer ‘where did this number come from’ in minutes, not a multi-day investigation. Most fragmented enterprises aren’t short on data, rather they’re short on a layer that makes the data they already have trustworthy and traceable. That’s the actual foundation, not whatever tool gets bolted on top of it.
You often discuss the 5 Vs of data: Volume, Velocity, Variety, Veracity, and Value. Which of these do enterprise leaders tend to underestimate the most when preparing data for AI?
Data Veracity, without question, closely followed by Data Value.
Volume is a given in today’s big data ecosystem. Everyone’s drowning in data. Variety and Velocity get a budget because they’re visible on a roadmap – new sources, pipelines and real-time feeds.
Veracity is invisible until it costs you something – usually a model produces an obviously wrong output, and someone traces it back to duplicate customer records or a field three teams populate three different ways.
Lastly, value gets underestimated differently – leaders assume that if data exists, it has value, without asking whether it actually moves a decision. Most enterprises have a small fraction of their data doing any real work and nobody’s mapped out which fraction that is.
In retail specifically, AI use cases can span personalization, forecasting, inventory, supply chain, customer service, and store operations. Where do you see the biggest gap between AI ambition and data readiness?
While AI ambition is spread evenly across retail, data readiness is highly asymmetrical.
Personalization and forecasting have the most reasonable readiness – retailers have been collecting transaction and browsing data for years.
The bigger gap is supply chain and omni-channel, especially the connection between online and physical stores. Omnichannel is incredibly complex because it forces a retailer to solve two of the hardest data problems in real-time: unified customer identity and fluid inventory. Most retailers have real-time e-commerce inventory but only periodic, sometimes end-of-day, visibility into store-level stock.
Customer service and store operations are still in early stages. Ambition is high, but the underlying operational data, i.e. staffing, task completion, real-time store conditions, is often the least digitized part of the business. Ambition is spread evenly across these use cases; readiness isn’t.
How should companies modernize legacy data infrastructure without disrupting mission-critical systems that the business still depends on every day?
Legacy data platforms often suffer from operational bottlenecks, severe data quality issues, and organizational scaling limits. The most effective way to protect daily operations during a modernization journey is to strictly avoid a single, high-risk big-bang migration. The primary objective of the data platform modernization should be making data demonstrably more useful for organizational decision-making, rather than treating the effort as a pure IT platform swap.
To cater to the AI era, organizations should adopt a data medallion and data mesh architecture. This strategy creates a structured pipeline where data is refined progressively from raw to business-ready, effectively preventing poor-quality data from decision making. By separating data into distinct logical layers, organizations establish clear lineage, making it simple to trace how information transforms from source to destination.
By packaging verified data into accessible products, organizations empower their teams to make strategic, evidence-based decisions rather than relying on intuition. Crucially, this approach establishes the foundation of data trust required for enterprise AI. Building AI agents on top of curated data products ensures the models learn from secure, accurate information rather than messy, unvetted datasets.
Governance is often added after an AI project is already underway. What changes when governance, privacy, security, and data quality are treated as architecture requirements from day one?
Everything moves faster, breaks less, and scales better. With the probabilistic nature of generative AI, governance is extremely critical but yet the most overlooked aspect in many organizations.
When you treat these elements as day-one architecture requirements rather than a last-minute audit checklist, you shift from damage control to velocity. Here is what changes:
First, you avoid the painful “rip-and-replace” phase where fully built AI projects are scrapped because they violate compliance or privacy standards.
Next, security and data quality aren’t patched on top; they are embedded into the pipeline. Your models run on reliable data, yielding outputs leadership and users actually trust.
Lastly, instead of hitting a regulatory wall when moving from proof-of-concept to production, the path to deployment is already cleared and automated.
In short, governance isn’t a brake on innovation rather it’s the steering system that allows you to drive fast safely.
As an advisor to the CAIO Circle, how do you think the Chief AI Officer role should evolve, and where should its responsibilities begin and end relative to the CIO, CDO, and business-unit leaders?
Being part of the CAIO Circle has been a genuinely useful window into this, because right now the role is all over the map because many enterprises created it reactively, in response to generative AI hype, rather than filling a clear operational gap. I believe this is going to settle quickly and it needs to. To work effectively, the executive breakdown should be crisp:
The CIO owns the infrastructure and system reliability. The CDO owns data as an enterprise asset (its quality, governance, and accessibility). The CAIO owns the translation layer between raw AI capability and real business outcomes.
That translation layer means taking explicit ownership of the AI portfolio, determining what to build, identifying the areas of business that can be improved and served better through AI technology and serving as the ultimate authority on where AI should and shouldn’t be deployed.
The biggest gap in most enterprises today is evaluating risk and value tradeoffs across competing business units. Blocking a high-risk AI initiative that a single business unit desperately wants requires someone whose incentives aren’t tied to that unit’s short-term targets and who is technical enough to evaluate the risk directly. That independent, technically grounded gatekeeper is precisely what the CAIO role needs to become.
Looking ahead, what will separate companies that achieve durable AI advantage from those that simply experiment with the latest AI tools?
It won’t be ones who have access to the best models. Model capability is converging rapidly, and raw technology will not remain a lasting differentiator.
Instead, the core test of long-term AI success – whether internal or customer-facing – comes down to one thing: Trust.
Trust is directly proportional to an organization’s ability to build an AI-ready foundation and an environment where new capabilities can be tested, safely deployed, and scaled at speed. Concretely, that requires three things:
A Reliable Data Foundation: Clean, trustworthy data so that every new use case doesn’t demand a massive, one-off cleanup project.
Agile Governance: Risk and compliance processes fast enough to keep pace with how quickly underlying models evolve.
Outcome-Driven Metrics: An organizational culture that measures AI by business value, not tech sophistication.
The companies that will still be stuck in “experimentation mode” three years from now will be the ones that never built that foundation – they will keep re-learning the same data and governance lessons on every single AI project.
The winners with a durable advantage will be the ones where the tenth AI use case becomes dramatically cheaper and faster to deploy than the first, because the foundation was built right the first time.
Thank you for the great interview, readers may also want to read his book Enterprise-Grade Hybrid and Multi-Cloud Strategies.