Artificial Intelligence is no longer a bolt-on feature. It's embedded deep into the fabric of enterprise software—and CIOs must now rethink how software gets delivered when AI becomes a central building block.
At the Software Excellence Network summit, Dr. Henning Heitkötter, VP and Head of AI Core at SAP, gave a thought-provoking masterclass connecting two fast-moving worlds: AI development and frequent software delivery. His conclusion? The same DevOps principles that fueled cloud innovation must now evolve to handle the AI revolution.
Heitkötter began with a story familiar to many enterprise leaders: the move from months-long release cycles to weekly and bi-weekly delivery cadences.
“But if we want to deliver daily,” he warned, “you can't just split your day into 7 hours coding and 1 hour testing. You need continuous integration and testing woven into the pipeline.”
This challenge becomes even more complex when you add AI artifacts like prompts, models, and benchmark datasets to the equation. Without automated systems to validate their impact, you risk replacing agility with bottlenecks.
The solution? Treat AI artifacts like any other software component:
The complexity of managing multiple microservices is already a known challenge in cloud-native architectures. Add AI into the mix—particularly models or prompt variants tied to specific services—and the traditional branching and backporting strategies quickly collapse under their own weight.
SAP’s approach? Single-trunk development not only for source code but also for AI components. AI prompts and model selections are versioned and deployed together with application code, ensuring they evolve in sync and don't create silos.
This eliminates the nightmare of coordinating hotfixes across disconnected branches—and makes it easier to rollback or test changes without breaking dependencies across the system.
Whether your team stores prompts in code, uses a dedicated prompt registry, or tracks model endpoints externally, the key is consistency. AI features shouldn’t live on a different timeline or process than the rest of the product.
Heitkötter argues that the principles of feature-driven development translate cleanly into the AI world — and may be even more essential there.
Each AI feature, like any other, must begin with clear customer requirements, proceed through design, implementation, testing, and be rolled out incrementally. In fact, he cautions against switching models or prompts for 100% of users without A/B testing and close monitoring in production.
Why? Because no matter how good your benchmark data is, real-world usage reveals gaps. Do users accept the AI’s suggestions? Do they override them? This user behavior becomes the basis for improving both the feature and your prompt dataset.
SAP encourages teams to fold this production feedback back into the benchmark set—a virtuous cycle where every deployment sharpens your system’s intelligence and performance.
None of this works, Heitkötter emphasized, without executive-level buy-in.
When AI projects start as skunkworks or are rushed into production under pressure, teams often cut corners: skipping integration tests, avoiding version control, or forgoing monitoring. The result is brittle, non-repeatable outcomes.
“We must resist the urge to treat AI differently from traditional software,” he said. “It’s tempting to reinvent the wheel for GenAI, but the best organizations apply everything they’ve learned about software engineering to AI from day one.”
CIOs must set the tone—insisting that AI artifacts follow the same discipline as code, and empowering teams to ship continuously, safely, and smartly.
The theory isn’t just academic. Heitkötter shared real-world examples from SAP.
In the pre-GenAI era, some internal tools were already deploying to production on 99% of working days (except Fridays). That culture of daily delivery carried over into their AI efforts.
A recent example? SAP’s internal AI agent runtime platform—launched just over two weeks before the event—had already been deployed to production nine times. By now, that number has almost certainly grown.
These rapid cycles aren’t about shipping unstable products. They’re about building the confidence and tooling to release AI features just like any other business-critical function—safely, incrementally, and always with data to back the decisions.
“GenAI isn't just another technology trend. It's a shift in how we build, test, and deliver software itself. The CIOs who win will be those who apply world-class engineering discipline to every AI artifact their organizations deploy.”
— Dr. Henning Heitkötter, VP, SAP AI CoreNOTE: While I've done my best to represent Dr. Henning Heitkötter's message as faithfully as possible, I still recommend that you watch his entire remarks to fully understand the depth and context of his insights. If you have questions or concerns, please reach out to me at Brandon.Lewis@Seerene.com.