The short answer: an AI use case on SAP reaches its first production release when four decisions have an owner. Someone defines what a correct result is. Someone chooses the identity the agent uses in SAP. Someone decides when human approval can relax. And someone can switch the agent off. This post explains how we settle each decision with you, and what we need from your side to do it. For the offer itself, the team shape and the release gates in summary, see our AI forward-deployed engineers page.
What we need from you
Our team is two or three engineers who work inside yours. However, the decisions that get a use case into production sit with your people, so the work needs three things from you:
- A process owner who can say what a correct result looks like, and who will later decide how much the agent may do on its own.
- Access to the SAP data, other systems and documents the use case depends on, under your normal security approvals.
- The people who will run it after go-live, so they learn alongside our engineers from the start.
For that reason, we recommend starting with one use case and one owner, not a programme. A single process with a clear owner lets you prove the controls before you scale them. If you have not chosen the use case yet, our getting started with AI agents white paper helps you pick one.
Decision 1: what “correct” means
First, the engineers map the SAP data, other systems and documents the use case reads, plus where the output has to land. Then they sit with your process owner and agree what a correct result looks like for this one process. For example, which cases must the agent handle, and which must it hand to a person?
That agreement becomes the evaluation test set: labelled real cases from your business. As a result, “good enough” stops being an opinion. Your process owner can see how the agent performs on your own cases, including the awkward ones, before anyone relies on it. Later, when a model, a prompt or an SAP process changes, the same set can be run again to check whether quality held.
Decision 2: which identity the agent uses in SAP
The agent acts either with the user’s own SAP authorisations or with a scoped technical user. These two choices behave differently, so we make the choice explicit for each use case.
- The user’s own authorisations. The agent can see and do only what the person using it can. In other words, your existing SAP roles keep doing their job, and an assistant never shows someone data they could not open themselves.
- A scoped technical user. This fits background work that no single person triggers, such as working through a queue. We limit its scope to what the use case needs, and we review it like any other privileged account.
We start that review with your security team early, not at the end. Therefore, the security review is part of the plan from the start, rather than left to the end. For the full set of controls, see AI governance, evaluation and security.
Decision 3: when human approval can relax
At first, a person approves every write or posting to SAP. For example, the agent drafts the action, and someone in your team confirms it. That is the safe default, but it is not meant to last for ever.
Your process owner decides, with us, whether and where approval can change, and we give them the evidence to decide. That evidence comes from the evaluation set and from how the agent performs in production. So approval might relax for one well-understood case type while staying in place for the rest. In short, the agent earns trust case by case, rather than receiving it at launch.
Decision 4: who can switch it off
From the first release, we monitor usage, cost and errors. In addition, you can switch the agent off at any time, and the process falls back to how it ran before. Before go-live, your team agrees who holds the switch and when to use it, rather than working that out after the first incident.
What you keep
At hand-over, you keep the code, the runbook, the evaluation suite and a team we have trained to run it. Because they live with you, your own team or another provider can take over the run, or we can keep it running as managed AI operations under an AMS contract. Our AI FDE page sets out the full hand-over.
What this looks like in practice
For a Saudi petrochemicals manufacturer, Altivate delivered AI built with Joule Studio and SAP BTP, using the generative AI hub, across finance, procurement and supply chain, and sales and customer service. It is now running in production across finance, procurement and customer service.
Whether a use case suits Joule Studio or a custom agent is a separate choice. Our guide to Joule Studio or a custom agent on SAP explains how we make it. Also, if you are still deciding who should do this work at all, compare AI FDE teams, consultancies, staff augmentation and in-house hires.
Frequently asked questions
What happens at the start?
The engineers map the data and systems the use case depends on, and they agree with your process owner what a correct result looks like. They also start the authorisation review with your security team.
Who decides when the agent can act without approval?
Your process owner decides, with us. We supply the evidence from the evaluation set and from production. Then approval can relax for some case types while staying in place for others.
Can we run it ourselves afterwards?
Yes. You keep the code, the runbook, the evaluation suite and a trained team. Managed AI operations is an option, not a condition.
Take your first use case to production
Bring one process and the SAP data it touches. We will work through the four decisions with you and scope a first production release.

