A year ago, a CIO meeting about AI often turned into a demo day. Teams showed off the pilots they had built, each one clever, each one a little proud of itself. Varun Thamba remembers those meetings well.
“Until last year when I used to meet CIOs, people were extremely excited to show off the pilots they are doing. Oh, see this, I built this, it works so beautifully. And obviously those are disconnected to the core system.”
Thamba took over SAP’s global AI go-to-market in January, after running AI for Asia-Pacific. He now does the job from Singapore on a schedule that follows the sun westward: “my day starts after lunch and ends at midnight.” From that vantage point, he says, the conversation with customers has changed completely. “The question has now shifted from how cool is your AI story to what value is your AI story bringing in 2026.”
His answer, laid out in the conversation, is more candid than most vendor keynotes. It argues that enterprise AI is a data, context and governance problem before it is a model problem. It also admits that a fully touchless enterprise is not here yet.
An org chart for agents
At Sapphire in May, SAP introduced what it calls the Autonomous Enterprise. The idea combines a unified platform for building, contextualizing and governing agents, a suite that runs core business operations, and a new way for people to work with enterprise software. The platform brings SAP Business Technology Platform, SAP Business Data Cloud and SAP Business AI into one governed environment, anchored by a Knowledge Graph and Joule Studio for building agents. The Autonomous Suite includes more than 50 domain-specific Joule agents.
Thamba describes the architecture as three layers. Joule is the engagement layer for everything SAP. Above the business processes sit autonomous domains such as finance and supply chain. Beneath them is the Business AI Platform, where customers extend or build their own agents.
The domain layer is built much like an org chart. Finance contains processes such as record-to-report. Those processes contain assistants, such as accounts receivable or accounts payable, and the assistants in turn contain agents. “Under each of these assistants, we have six or seven agents,” Thamba explains. An accounts receivable assistant might draw on a matching agent and a posting agent, among others. “And the assistants can then talk to each other across domains. And so then that autonomous enterprise becomes a reality.”
The thesis behind the design is one SAP has repeated all year: enterprise AI only works when it is grounded in business data, governance and operational context. Thamba ties that thesis to a market shift he finds telling.
The SaaSocalypse that wasn’t
Earlier this year, investors worried that AI agents and vibe-coded software would turn SaaS into legacy infrastructure. Thamba argues that fear has faded as buyers learned what agents need to work in practice. “People have realized that agents need to be embedded, need to understand the semantic context, understand the governance. And that’s bringing a lot of the conversation front and center to SAP. And other SaaS.”
It is the strongest argument SAP has. A standalone agent can draft an email. An agent that approves a purchase order, posts a journal entry or releases a payment has to understand the business entities and relationships involved, and the controls around them. That is the job of the Knowledge Graph, which gives agents a structured map of business entities, processes and relationships.
Three shifts reshaping the CIO agenda
Thamba sees three trends driving how enterprises now buy and deploy AI, and each carries a lesson for technology leaders.
The first is that ROI has replaced novelty. He points to rising token costs and the practical limits of agents in real processes as the forces making ROI the “big, big driver.” The useful part for CIOs is the discipline this forces: value has to be tied to a business process, not to an impressive demo.
The second is that AI has to show up where people already work. Thamba frames this as a change-management problem rather than a technology one. “Earlier, people were happy to have AI as an innovation happening in another front end… Now, over a period of time, getting users to bookmark so many URLs, to get them trained on how to use that AI, becomes a big problem.” The answer is to put AI inside the tools people already use. SAP has made Joule available as a mobile app that takes natural-language requests, and customers are asking for it inside Microsoft Teams. “Training is not required. So the change management is extremely minimized… the friction is coming down.”
Thamba says SAP’s CEO frames the idea as “AI is the new UI.” He gives a concrete example of what that means. A finance professional’s priorities shift through the month, from closing and collections at the start to planning at the end. “Can AI adapt in giving you your landing page, which is more adept to the way you do work? So that’s like AI evolving with your process rather than hard-coded to say, these are the five icons in your SAP landing page.”
The third trend is the one most likely to interest CIOs who manage a crowded AI vendor list: consolidation, on both the application side and the model side. On applications, if a business can build a driver-facing portal for posting delivery challans on a governed AI platform, it no longer needs a separate PHP or Java stack for the job. That means fewer skill sets, fewer teams and fewer contracts.
On models, Thamba is blunt about the open-ended multi-LLM approach many enterprises began with. “This is a bit like having multiple credit cards. Your loyalty is split. And when you want to use the points, none of them are your platinum or gold.” He adds that the small performance edge one model version holds over another is “not really earth shattering in an enterprise context unless you are doing some cancer research.” The value, he argues, comes from scale, reusability and the ability to switch a capability on or off. That does not mean customers are locked in. The platform is meant to offer a choice of LLMs, and Anthropic’s Claude models are set to become a key reasoning and agentic layer inside SAP’s AI ecosystem.
The data play: from 70–80% to everything
If customers are going to consolidate on one vendor, that vendor has to see beyond its own data. Thamba is direct about SAP’s position. “70–80% of your data is SAP. Bring that 20%-ish over and then you can build something.” Customers, he says, pushed back on a plan that left the rest out: “even if SAP by volume is less, by value it’s a lot for me. So I would rather bring other data to myself.”
That is the logic behind SAP’s recent acquisitions. SAP announced plans to acquire master data specialist Reltio in March, and the Dremio and Prior Labs deals followed in May. The Dremio deal is expected to close in Q3 2026, pending regulatory approval. One analyst reading frames the three as a deliberate pipeline, in which Dremio collects data from various sources, Reltio cleanses and harmonizes it, and Prior Labs supplies models to analyze it.
Thamba’s descriptions fit that framing. Reltio manages master data beyond core SAP objects, such as machinery and digital twins, so customers who kept two master data catalogs could now run one, alongside SAP MDG. Dremio adds a native lakehouse. SAP already partners with Databricks, Snowflake and the hyperscaler data lakes, but customers wanted something that grows organically within the SAP portfolio. “You can now bring petabyte-scale data, you can bring in completely non-SAP data here.”
Prior Labs brings tabular foundation models, which are built for structured data and need little training data. Thamba gives a banking example: “Out of 1,000 calls, only two people take. So then I don’t have enough data set to even build.” For a model that works with scarce positive examples, that is a big advantage. SAP plans to keep Prior Labs independent and invest €1 billion in it over four years.
Together, he says, these moves let SAP serve industries where it has little process history, and domains it never touched before. Some of the largest global banks, he says, are exploring tabular forecasting with SAP even though they would never run core banking on it. “This use case and this acquisition is making them rethink and say, okay, we’re anyway doing our core financials in SAP. Does it also now make value to do the business side there?”
Where growth comes from
Thamba names three places where he sees SAP growing, and each says something about where customers will find value.
The first is faster cloud migration. The long tail of customers on older versions is SAP’s historic weak spot compared with pure-play SaaS vendors. Transformation agents and Joule for consultants are meant to shorten the move, so customers can adopt new capabilities sooner.
The second is process reinvention. For a net-new procure-to-pay design, SAP asks three questions. Can the process be more intelligent, so the buyer knows how much to buy? Can it be more bottleneck-free, with rules and agents trimming approvals? And can it be more autonomous? His example is agentic sourcing, in which an agent tracks a published calendar of government rubber auctions in Malaysia, works out how much is needed and where to store it, and then bids.
The third is industry white space. Standard agents cover common processes, but industries differ. SAP is using a forward-deployed engineering model to build repeatable industry solutions. Thamba is realistic about where that stops: he does not believe every customer’s request can ever be met out of the box, which is why the platform layer matters. Otherwise, he warns, customers end up “taking all the SAP data out, managing two systems of record.”
The three-hour problem that became a three-week problem
Thamba’s best illustration comes from a manufacturer that assembles motorcycles from knocked-down kits. Procurement may be agentic in SAP, but at the receiving bay a person still counts parts by hand, such as spark plugs. When a shipment is short or has the wrong model, the worker returns to a terminal, finds the vendor, and negotiates by email, often sending a photo as proof. The result, Thamba says, is that a cycle that should take hours “is becoming instead of three hours, three weeks.”
SAP’s answer combined technologies that exist today. A thermal-sensing camera counts the parts and matches them to the master data for that model. If the match is correct, the system pays the supplier automatically. “It needs to validate into SAP, it needs to check master data within SAP, it needs to trigger payments within SAP. So, that is where is the focus for us.”
The lesson for CIOs is that the biggest process gains often sit at the edge of the system of record. Vision, IoT and external data feed the process, and an agent grounded in master data closes it.
SAP has gone further in a demo. One scenario watches for Strait of Hormuz disruptions by scraping the web, checks the customer’s catalogue to see where a delayed shipment is, and then helps the planner decide whether to wait for the ship or airlift the goods. It has not been released to customers. “We have not yet GA’d it because of the implications of it.”
Openness, with guardrails
On how outside agents may interact with SAP, Thamba is firm. SAP is a founding member of the agent-to-agent (A2A) framework alongside Google, and Joule already exchanges requests with Microsoft’s Copilot, with credentials managed through Active Directory. But the core application layer is closed to anything ungoverned. “Only trusted, governed APIs are allowed to touch an SAP system because we are responsible for the SLAs.”
The technical rule follows from that. The integration suite offers an MCP server at the open platform layer, but traffic to core applications goes through A2A, “because MCP as a technology is hard to govern.”
Beyond that, SAP is reworking its partner marketplace for what it calls the agentic era and launching a Business AI Platform Validated Partner program, with a new agent store for certified partners. Thamba admits the marketplace “is not yet there.” SAP has also announced a €100 million fund to accelerate agent deployments with partners.
The reality check
The most useful part of the interview comes when Thamba is asked whether an autonomous enterprise exists today. His answer is measured. Many of the agents go live in October, and more by December. Customers with early versions are seeing results: he cites ABB and JK Cement, where order and purchase-requisition creation is running “50%, 60%, 70% faster” in those steps. There are also pointed examples in recruitment and claims processing. But he is clear about the ceiling. “Nobody is today saying, okay, let’s take an entire record to report process and completely make it touchless.”
Even when the agents are generally available, he expects customers to choose different levels of control. One may want the agent to run a three-way match and then park the result for a human to approve. Another may let the agent post anything it is more than 90% confident in and send the rest to a person. The technology will be ready before some organizations are. “Then it is a little bit of a psyche thing,” he says, and the pace of adoption will depend on how quickly leaders are willing to hand over decisions.
What CIOs should take from this
Start with value. Every agent deserves a business case that names the process metric it moves, by how much, and what the agent costs to run. Thamba’s emphasis on ROI over novelty is a useful filter for any pilot that still lives in its own front end.
Treat context and governance as buying criteria rather than afterthoughts. Agents are only as safe as the semantic layer and the controls beneath them, so ask any vendor how its agents see business entities and how their actions are audited.
Count your AI contracts, too. Multiple platforms and multiple model relationships split your leverage. Consolidation can improve cost and reuse, but it has to be weighed against lock-in.
Set the autonomy dial deliberately. Confidence thresholds and human-in-the-loop points should be defined process by process, as Thamba’s three-way-match example shows, and the gains are often hiding in physical or semi-manual steps outside the ERP, which means planning the connection back to the system of record. All of it rests on the foundation. Agents are available on the cloud releases, and analysts reading Sapphire noted that AI deployments depend on a clean core and cloud migration. Legacy estates will be a drag on adoption.
The bottom line
SAP’s pitch is that enterprise AI is a data, context and governance problem before it is a model problem. As its CTO put it, enterprise AI stalls because the data isn’t ready for AI agents, not because the models aren’t good enough. The acquisitions, the Knowledge Graph and the A2A-only rule for core systems all follow from that view.
What remains to be proved is delivery. Several agents are still weeks from general availability, the end-to-end touchless process Thamba says nobody has yet is the real test, and customers will need to see how the acquired technologies are integrated and priced. Thamba’s own candor helps SAP’s case, though. A vendor that says plainly where autonomy ends is easier to plan with than one that promises everything. By December, when more agents are live, CIOs will have real numbers to judge whether this model delivers.