We have a lot of legacy RPA bots (UiPath) that handle invoice data entry, but they break every time the vendor changes their PDF layout. I’ve heard that "Agentic RPA" is the next big thing. Can AI Agents actually handle the "clicking and typing" on old legacy software, or are they only good for cloud-based API tasks? We are looking for a solution that combines the reliability of RPA with the visual "intelligence" of an LLM to handle non-structured data.
3 answers
Traditional RPA is "brittle" because it relies on fixed selectors—if a button moves 5 pixels, the bot fails. "Agentic RPA" uses Vision-Language Models (VLMs) to "see" the screen like a human. Instead of looking for a specific HTML tag, the agent looks for the concept of a "Submit Button." This makes it much more resilient to UI changes. In 2024, tools like Microsoft’s Copilot Studio and UiPath’s own "Autopilot" are merging these worlds. You still use RPA for the high-speed "boring" tasks, but you use the Agent for the "judgment calls" like reading a handwritten note on an invoice.
Is the latency an issue here? Sending a screenshot to a model like GPT-4o every time you want to move the mouse sounds like it would take a long time compared to a local RPA script.
We switched to a "Vision-first" agent for our SAP entry, and our maintenance tickets dropped by 60%. It’s slower per transaction, but it doesn't break every Monday morning.
That 60% reduction in maintenance is the real ROI. The "hidden cost" of RPA has always been the developers' time spent fixing broken selectors. Agents solve that.
You're right, Gregory, it's definitely slower. You shouldn't use an agent to "click" every single field. The best architecture is "Hybrid": use the agent to extract the data from the messy PDF into a clean JSON format once, and then let the traditional RPA bot "blast" that data into your ERP system at high speed. Use the AI for the intelligence and the RPA for the execution. This saves on API costs and prevents the agent from getting confused by every single pop-up window on a legacy desktop app.