Blog
Building apps with AI assistance in OutSystems: what stays in your hands
· 4 min read
Your client wants a new app: an onboarding journey for employees, a request process for dealers, an approval flow for invoices. In the past, that started with a long list of screens and weeks of build time. Increasingly, you now hear: just let the AI build a first version, it saves weeks. OutSystems is responding to this publicly and in a big way, with AI that helps build the app itself, not just AI that runs inside the app. For an SME considering such a platform, the question is not whether it can be done, but what is still left for a human to do.
What OutSystems publicly offers in AI assistance for building
OutSystems calls its main AI feature for building itself Mentor. You type a request in plain language, such as building an onboarding app, and Mentor first delivers a blueprint: an overview of what it is going to create before anything is actually built. Only then does it generate the basic structure, screens and data models that a team continues working with. That is how it is described on the Mentor product page: you can "request a complex change, review Mentor's implementation plan, and only then give final approval before the update is applied."
Besides building screens, the platform also offers AI agents as a building block within an app: functionality that carries out tasks on its own, such as retrieving data, summarising it or preparing a decision, and that you deploy as part of a larger business process. These agents are combined with workflows that keep track of the structure, status and overview of the entire process, according to the documentation on agentic workflows in OutSystems Developer Cloud. In short: AI helps not only with typing code, but also with designing the app and with carrying out tasks once the app is running.
Where people keep the final step
What is interesting is that OutSystems, in its own documentation, explicitly builds in moments where the AI stops and waits for a person. In the Mentor Studio development environment, according to the documentation, it works like this: for a complex change, Mentor Studio first proposes the changes it is going to make, and waits until you accept them before anything is applied. Afterwards, a comparison view lets you see exactly what has changed before you continue. So the AI does not quietly carry on; it presents the plan first.
For agents that carry out tasks on their own while an app is in use, a similar mechanism exists, only during execution itself. The documentation on agentic workflows describes the so-called human activity block: a component that pauses the workflow until a specific user or group has completed a task. The explanation is clear about when that is appropriate: agents may need human oversight for high-impact decisions, or when the agent itself has little confidence in its answer. In practice, this means you decide at which step in a process a person must review things before an agent continues, for example before a request is finally approved or an email goes out.
That is precisely the distinction the platform itself makes between what AI proposes and what a human decides: AI generates a plan or a result, but applying it is a separate, visible step that you take.
What we recognise from our own build work
Those two habits, showing a plan first and only applying it afterwards, and having a person review at the right moment, match how we work ourselves when building applications, OutSystems or not. For the MEA Certification Platform, installation management used to be handled manually by email, resulting in chaos and delays for dozens of dealers and hundreds of vehicles a year. The solution now in place, a fully automated platform where dealers manage their own installations and certificates are generated automatically, only works reliably because the automation builds in a checkpoint at the right places instead of charging ahead on its own everywhere.
One lesson that sticks: always ask an AI tool for a measured outcome, never just for confirmation that something works. AI tools readily confirm that a feature is correct without really having tested it, and that applies just as much to an AI building an app as to an AI agent carrying out a task within that app. As long as there is a concrete plan to review and a clear moment at which someone checks or approves the result, AI assistance remains an acceleration of the work rather than a risk.
What does this mean for your business?
If you are considering AI assistance for building or running a business application, it helps to have a few things clear in advance:
- Never let a plan or result go through automatically without someone looking at it first, especially not for steps involving money, customer data or irreversible actions.
- Decide per process where an agent may act on its own and where a person must review before the next step takes place, rather than keeping the same pace everywhere.
- Always ask for a verifiable result rather than reassurance: you test a generated screen yourself, you verify an agent's answer yourself, even if the AI says it is correct.
- Start small: a first app or a first agent with a limited goal, so you can see how the review moments work in practice before you extend it to a larger process.
AI that helps build software mainly changes the pace, not the question of who is responsible for what ultimately goes live. Would you like to discuss what that looks like for your process? Feel free to get in touch.