Teaching your OutSystems 11 app to understand bank statements, so you don't have to!
- Toni Hatherell
- 3 days ago
- 3 min read
This is the story of how an O11 app teamed up with the Agent Workbench in ODC, proving that you do not need a rewrite to unlock real AI value.
Written by: Natasha De Guzman, Tech Consultant
The Case Of The Mysterious Payment Upload
It always starts the same way.
A notification pops up: "New Proof of Payment Submitted." You roll your chair forward like you're starring in CSI: Accounts Receivable Unit. You click to open the file. And suddenly you're analyzing:
Pixelation
Reflection glare
Camera angle distortion
Whether this was photographed in a moving vehicle
And most importantly: "Does this actually prove someone paid?"
Suppose you're processing dozens of these documents, fine. If you're processing hundreds, it becomes a recurring side quest you never asked for.
So we asked the question: "What if the system handled the obvious cases automatically, and left only the weird ones to humans?"
Turns out: Yes. Very much, yes.
The AI Intern That Never Sleeps
We are not here to replace humans. Humans are great! Humans recognize nuance. Humans can sense "fraud vibes."
AI is here to do the boring, repetitive, pixel-staring part:
AI can detect text patterns
AI can verify numbers and dates
AI can confirm expected formatting
AI does not get tired, bored, or existential at JPG#137
So, instead of forcing analysts to open every document. AI does the first pass, and Humans handle the exceptions.
Operation: Payment Detective
Your existing app lives in OutSystems 11. It works. It's familiar. It's in production. There is zero reason to rebuild it just to add AI. Meanwhile, the Agent Workbench, where the AI reasoning lives, runs in OutSystems Developer Cloud (ODC).
So, instead of rebuilding the system, we integrated:
O11 = File upload screen + back office review UI
ODC = REST API service + storage + AI orchestration
Agent Workbench = Document interpretation + reasoning
Human Review = Final authority
We didn't replace your app. We gave it a sidekick!
The Flow
(aka how and where the magic happens)
Here’s how it all comes together:
User uploads a bank statement
O11 app sends file to ODC via REST API
ODC app stores the file and marks it Pending
ODC app sends the document to AI for validation
Agent Workbench returns {isValid, reason}
ODC app updates the document record
O11 app displays the status
Note - Human Reviewer can override the AI’s decision
Workflow Diagram:
Choosing Your AI Brain
(Don’t worry, you can swap it later)
One of the nice things about using the Agent Workbench in ODC is that you can prototype quickly without needing to set up your own AI infrastructure. ODC provides ready-to-use trial models you can experiment with right away.
Every ODC tenant (including personal edition) has access to:
GPT-4o via Azure OpenAI
Claude 3.7 Sonnet via Amazon Bedrock
Amazon Nova Pro via Amazon Bedrock
These are trial models, meaning they have usage limits rather than time limits. Once the quota is used, the model pauses until it resets, and you can monitor that directly in the Trial Model details page in the ODC portal.
This PoC was developed for a Google-shop organization. Claude would have worked just as well, but since the client’s ecosystem already spoke fluent Google, Gemini was the natural choice.
Because the chosen model resides in configuration rather than deeply embedded code, changing your mind later is simple. If another model ends up providing better clarity, accuracy, or reasoning performance, you can switch it out and redeploy it, without needing rewrites, refactoring, or drama.
My advice is to start with the trial models and let them audition. Keep the one that behaves best.
Human In The Loop
(still the boss)
AI is the intern, fast and tireless, but occasionally confidently wrong. Humans remain the decision-makers.
Back-office reviewers can approve, reject, override, or add notes.
![[Payment Validation Back Office]](https://static.wixstatic.com/media/6c5498_5423ed59703145f89292c179d48392ca~mv2.png/v1/fill/w_900,h_307,al_c,q_85,enc_avif,quality_auto/6c5498_5423ed59703145f89292c179d48392ca~mv2.png)
Future Enhancements
(when this graduates from POC to production)
Replace API Key with OAuth2 / Google Identity.
Convert synchronous AI validation to asynchronous event-driven processing.
Add confidence scoring to auto-approve simple cases.
Add dashboards for throughput and override trends.
Conclusion
We didn’t replace humans; we replaced the part of the job that makes humans question their life choices.
O11 stays familiar. ODC provides intelligence. AI handles routine validation. Humans handle exceptions.
Small architectural shift. Massive reduction in monotony. 10/10 would automate again.
AI + Low-Code + Human Insight = the future of how work gets done.
Want To Build One Yourself?
If this whole “OS11 + ODC + AI intern” thing has you thinking: “Wait… I want to build one of these!”
Good news: OutSystems has already created a learning path for exactly that.
Learn how to build Agent Workbench–powered apps here:
end of article
Comments