SAP Wants to Be Your Agentic Workforce. Your ERP Cannot Be the Whole Story.

SAP now sells a vision it calls the autonomous enterprise, a workforce of AI agents running your business inside S/4HANA. The vision is real and parts of it are already shipping. The catch, and it is a big one, is that the most important work in your company does not happen inside your ERP, so no single system, SAP included, can be the whole story.

I got to pressure test that idea in a long conversation on my podcast, The Integration Layer, with Brent Lewis, who leads the data and AI team at Armstrong, a 166 year old company scaling agents into production right now. He is not a vendor and he is not a keynote. He is the person who has to make this actually work, govern it, and keep it from doing something expensive at three in the morning. What he described lines up almost exactly with what SAP's own customers, and even SAP's own executives, are quietly admitting. This piece is that reality check.

A note before we begin. Brent is sharing his own views and personal experience in this conversation. Nothing here represents the official position of his employer.

When a three year AI plan makes a practitioner chuckle

Ask Brent about the classic multi year technology roadmap and he laughs. "We are asked sometimes to think three years out and it makes me chuckle," he told me. He will commit to about a year with confidence, treat year two as a direction rather than a plan, and refuse to pretend he can see further than that in a field where a model that beats last quarter's best ships on a random Tuesday.

That instinct matters, because SAP's AI story is sold as a multi year transformation, and the honest version of enterprise AI planning is much shorter and much more humble. Hold the destination loosely. Build the foundation that survives whichever model or agent wins.

What SAP is actually shipping

Give SAP its due, because some of this is real. Joule, the copilot layer, is generally available and embedded across S/4HANA Cloud, SuccessFactors, Ariba, and more. Business Data Cloud, launched in early 2025 with a serious Databricks partnership underneath it, is genuine data infrastructure and not a slide. The Knowledge Graph that maps SAP's own data model is a sensible piece of engineering.

Now the part the keynote glosses over. At SAPPHIRE 2026, SAP framed fifty plus Joule assistants orchestrating two hundred plus specialized agents. Those are announcement numbers, not in production numbers. Joule Studio, the tool to actually build these agents, reached general availability only in early 2026, and a long list of the named agents are still beta or roadmap stretching across the year. There is nothing wrong with a roadmap. There is something wrong with reading a roadmap as a description of today.

The tell is in SAP's own customer base. In the DSAG survey of SAP customers, among those who have put AI into production, seventy seven percent are running it on non SAP solutions and just three percent rely on SAP. Only forty three percent have implemented any AI use case at all, and half of those surveyed think AI's potential is currently overrated. This is SAP's own user group, not a competitor's marketing.

Agents are users, and your workflows do not live in your ERP

The single most useful reframe Brent offered has nothing to do with SAP and everything to do with how you should think about all of this. Stop picturing an agent as a tool. Picture it as an employee.

"Agents to me are just like a user," he said. "You give it tools, you give it skills, and it is a digital version of a user." That means the same security, the same system identity, the same audit trail, and the same review you would apply to a human with those permissions. It also means a human is accountable for it. Agents "report to, or they are the responsibility of, a human that uses them in their day to day."

Hold that thought against SAP's pitch, because here is where the ERP story hits its ceiling. "Workflows do not live inside of an ERP," Brent said. A real business process runs from email to a shared drive to the ERP to the CRM and back. An agent that can only see inside S/4HANA is an employee locked in one room of the building. SAP's own analysts have asked the same question out loud, whether SAP's agents will play nicely with others. Most enterprise value lives in the seams between systems, which is precisely where a single vendor's agents are weakest.

Automation, an assistant, and autonomy are three different things

Before you buy anyone's agent, Brent's advice is to figure out whether you even need one. Most teams that say they want an agent actually want automation, a fixed if this then that script with a human reviewing exceptions. Some want an assistant, where a person stays in the loop and supervises. True autonomy, an agent making decisions and acting without waiting for approval, is a much smaller slice than the hype implies.

His method is refreshingly boring. "Start with the workflow. Map out what you do as a human today." Once it is on paper, the automation only steps are obvious, and the genuine decision points, where autonomy might earn its keep, stand out. I have watched this from the ERP side for fifteen years, and I will put it more bluntly. After hours of discovery, most requests for an agent resolve into a request for a well built workflow, and maybe ten to fifteen percent is truly agentic. Buying an autonomous agent for an automation problem is renting a bus to drive yourself downtown.

The layer that breaks first, and it is not the AI

When you scale from a proof of concept to the whole organization, something gives. I asked Brent what breaks first. He named governance, with data as its close cousin, and the reason is quotable. "Good AI needs good data. You are going to get a really strong AI system that is really, really dumb" if the data underneath it is a mess.

Here SAP agrees with him, and that is the most important agreement in this piece. SAP's own CTO, Philipp Herzig, has admitted a significant gap between AI innovation and actual outcomes, and put it plainly: "no AI agent can compensate for a broken data model." An SAP customer, an ExxonMobil executive, said the quiet part just as clearly. "Data is the asset which has been trapped. If we cannot get this foundation right, we will pay the price for that evermore."

This is why my podcast is called The Integration Layer. The data and integration foundation is not a detail beneath the AI story. It is the story. From the ERP seat, I will say it more strongly than Brent did. The first thing that fails is almost always the data layer. If you make it all the way to your third layer, humans struggling to supervise the agents, then congratulations, your data and governance were good enough to get you there.

SAP knows this, which is why its whole clean core message ties AI value to standardizing your processes and cleaning your data first. That is genuine engineering hygiene. It is also a convenient commercial gate, because the road to clean core runs through a RISE with SAP migration. Notice the move. The reason the AI has not delivered value yet becomes your customization debt, and the cure is the cloud contract SAP wanted to sell you anyway. The advice is correct. Just be clear eyed about who benefits from the framing.

Do not marry one vendor, including SAP

Armstrong made a decision early that is worth copying. They refused to lock into a single AI vendor. Instead of buying one model for everyone, they built what Brent calls a model garden, a secure environment offering up to eighty models, with logic that routes a given prompt to the right one. "To pick a horse in that race is very dangerous," he said, because the pace of model innovation is almost mind blowing and whatever you marry today is outpaced within months.

This is the sharpest edge of the SAP reality check. When you lean your whole agent strategy on one large platform's built in agents, you inherit that platform's pace of innovation and its roadmap. Even SAP does not build its own frontier model. At SAPPHIRE 2026, SAP named Anthropic's Claude as the primary reasoning model behind Joule. The largest enterprise software company on earth chose flexibility over building its own. Take the hint.

Brent's mental model for which model goes where is the best I have heard. Think of your agents as an org chart. At the top, the orchestrator making complex decisions, you want the smartest frontier model you can afford, an Opus class model. As you move down toward the repetitive worker level tasks, you drop to smaller, cheaper, often open source models you can even run on your own hardware to control token cost. You do not put your most expensive model on mundane work, and you do not put a lightweight model in charge of the hard calls. Portability across that whole stack, from cloud to on premise, is what keeps you free.

The point is scaling, not shrinking

One more place where the loudest AI narrative is, in Brent's view and mine, simply wrong. Some vendors and executives sell agents as a way to cut headcount and shrink SG&A, and the share price twitches upward on the announcement. Brent called that a dangerous move, and he is right on the math. "It is hard to even theorize how you reduce staff by that much without cutting your productivity."

The better use, echoed by two CEOs he cited, is to use AI to scale rather than to cut. Take a hundred people and let an agent workforce make them as productive as a thousand. Chase the adjacent market you never had the capacity to enter. That is also the honest competitive story, because the moat that scale, process, and headcount used to buy a large company is thinner than it was a year ago, and a small, focused firm with good data and good agents can now punch far above its size.

How a practitioner actually plans for this

I asked Brent what guidance he would give a leader trying to build a six to twelve month AI strategy. His answer was not a framework with a clever acronym. It was discipline. Stay informed relentlessly, through podcasts, partners, and the analysts. Pull the threads together yourself, then hand your own strategy to your favorite model and tell it to attack the plan, find the gaps, poke holes. He and his CIO do exactly this to each other.

I would add two things I tell clients. Chase the low hanging fruit first, the unglamorous use cases that return real ROI, so you are funding the ambitious work with wins rather than promises. And pace yourself honestly. Brent's analogy is a marathon. If you have not trained, no amount of ambition gets you to the finish line safely, and going out too fast just gets you hurt. The companies that spent the last two years cautious cannot sprint their way to an autonomous enterprise this quarter, no matter what the keynote implied.

SAP's agent story is worth taking seriously, and I am not telling you to avoid it. I am telling you to right size it. Let SAP's agents do the work that genuinely lives inside SAP. Build or buy an orchestration layer above your systems for everything that does not. Fix your data foundation first, because it gates all of it. Keep your models swappable. And measure the roadmap against what is actually generally available today, not against the stage.

The autonomous enterprise is not a product you purchase. It is a foundation you build, one honest layer at a time.

If you are working through where AI genuinely fits in your SAP or ERP landscape, that is what I spend my days on. Come say hello.

Shubhendu Tripathi is an AI and ERP strategy consultant based in Toronto, and the host of The Integration Layer, a podcast on AI, enterprise systems, and the work of making them fit together. Connect on LinkedIn or reach out at tripathis@qubittron.com.