Why most SAP AI agents never leave the sandbox
Every SAP AI agent demo works. That's what a demo is for. Someone types a question, the agent reasons over a tidy slice of data, and the room nods. Then the project stalls — not because the technology failed, but because a demo and a production system are different animals, and almost nobody plans for the gap between them.
The pattern is now common enough that industry analysts expect a large share of agentic-AI projects to be cancelled over the next couple of years, mostly over unclear value, runaway cost and weak governance. In the SAP world specifically, independent user-group surveys suggest that among customers already using AI in production, only a small minority run it on SAP itself. The demos are everywhere. The production systems are rare.
Here are the five checks we run before we'll call an agent production-ready. If your pilot can't pass them, it isn't a go-live candidate yet — it's a science project.
1. Is the data actually ready?
An agent is only as good as what it's grounded in. In a demo, the data is curated. In production, it's the real master data — duplicated vendors, half-filled fields, inconsistent units, the accumulated entropy of a decade of workarounds. Data quality is consistently cited as the number-one reason enterprise agents fail, and it's the least glamorous thing to fix, so it gets skipped.
The check: before building, confirm the agent's grounding sources are complete, current, and clean enough that a wrong answer is the exception, not the norm. If they aren't, the first project isn't the agent — it's the data.
2. Will it survive the next upgrade?
It's tempting to get an agent working fast by reaching into the core — modifying standard objects, wiring logic where it doesn't belong. It demos beautifully and quietly mortgages your next S/4HANA upgrade. Clean-core discipline exists precisely to prevent this: build as side-by-side (BTP) or on-stack (ABAP Cloud) extensions using released APIs, so what you ship stays upgrade-stable.
An agent that breaks at your next upgrade wasn't a shortcut. It was a loan with a balloon payment.
This matters more than usual right now, with ECC maintenance ending in 2027 pushing a wave of migrations. Anything built the wrong way today is technical debt you'll pay off under deadline pressure.
3. Can you see what it's doing — and what it costs?
In a sandbox, nobody watches token spend, latency, or how often the model quietly changes its mind. In production, all three decide whether the agent is an asset or a liability. Observability isn't a nice-to-have you add later; it's part of the definition of done. You need to see every action the agent takes, what it cost, and when its behaviour drifts.
The check: before go-live, confirm you can monitor quality, cost and drift, and that there's an audit trail for everything the agent touches. If you can't answer "what did it do last Tuesday and why," it isn't ready.
4. Is the governance real, or a slide?
Who is allowed to approve what the agent does? Where's the human in the loop? What happens when it's unsure? Governance that lives only in a policy deck evaporates the first time an agent posts something it shouldn't. Real governance is wired in: defined boundaries, escalation paths, and a clear line between what the agent decides and what a person confirms.
The check: the agent should route genuine edge cases to a human by design, not by luck — and the rules for that should be explicit, testable, and owned by someone.
5. Who runs it after go-live?
This is the one almost everyone forgets. An agent isn't a project you finish; it's a system you operate. Models change under you. Grounding data goes stale. Token costs creep. Traditional SAP application management watches your ABAP and interfaces — it does not watch any of this. Without an operating model, agents that launched to applause get quietly switched off six months later, and the organisation concludes "AI didn't work" when what actually failed was the plan to keep it alive.
The check: before you ship, know who owns drift, cost, grounding refresh and audit — and how they're paid to care.
The uncomfortable summary
None of these five is about model cleverness. They're about engineering discipline and an operating plan — the unglamorous last mile that the biggest firms often skip and the smallest teams rarely scope. That mile is the entire difference between an agent that demos and an agent that delivers.
It's also, not coincidentally, exactly the work we do. If you have a pilot stuck in a sandbox — or you want to start one that won't get stuck — a readiness audit is the fastest way to find out which of these five checks stands between you and production.
Book a readiness audit See how we build
Note: figures referenced here come from independent SAP user-group surveys and industry analyst projections; they're directional, not guarantees. Replace bracketed fields before publishing.