Keeping a Human in the Loop While an AI Agent Runs Your SOP
Your finance team has a three-page SOP for onboarding new vendors. It tells a human to log in, check the supplier portal, copy the invoice line items into your ERP, and confirm the GL mapping. The process is stable, but the UI of the supplier portal changes every six months. Your RPA bot breaks, a developer has to rebuild it, and your backlog of broken automations grows. Meanwhile, the SOP remains the same. This mismatch between brittle bots and stable procedures is where many organizations get stuck. You want automation, but you also need to keep qualified staff in the loop to handle the exceptions and judgment calls.
Why RPA breaks here
Traditional RPA tools like UiPath, Automation Anywhere, and Blue Prism work by binding actions to selectors, XPath expressions, or object IDs. These identifiers are tightly coupled to the current version of the application. When a finance team updates the supplier portal, the selectors often change. The bot tries to find an element by an old ID or class, fails, and halts. The result is a cascade of manual fixes. A Gartner analysis of RPA deployments found that up to 60 percent of maintenance effort in mature programs is spent on selector changes and process rewrites rather than on new capabilities. You end up with a treadmill: build a bot, wait for the UI to shift, rebuild the bot, repeat. The human who wrote the SOP is sidelined, even though the procedure itself is sound.
What changes with computer use agents
- ●Agents see the screen like a human, so they can act on UI elements without prebuilt selectors.
- ●When the UI changes, they adapt rather than halt, because they read the current layout.
- ●They recover from unexpected states such as pop-ups or errors instead of stopping the workflow.
- ●They can follow SOPs written in plain English, with no need to convert prose into flowcharts.
- ●They work across any application, including legacy systems, Citrix, and virtualized desktops where RPA struggles.
- ●Agents run on cloud VMs, a desktop app, and through an API, making them easy to integrate into existing tools.
A computer use agent can follow the exact wording of your SOP while keeping a human in the loop for checkpoints and approvals.
The difference on the ground
Imagine you have a process that starts with: "Open the procurement portal, navigate to the contract approvals page, and identify all pending items." An RPA bot must know the exact class name and URL path. If the portal rebrands or reorganizes its navigation, the bot fails. A computer use agent looks at the screen, reads the header "Procurement," navigates to that section, and proceeds. When a popup blocks the screen, an RPA bot typically stops and waits for manual intervention. An agent can dismiss the popup or notify a human, then continue. This difference matters at scale. If you run the process 5,000 times a month, a one-second delay per failure adds up to more than an hour of manual work per month. Agents reduce that friction by recovering autonomously and flagging only the truly unique issues.
How to move without the risk
You do not need to rip out all RPA at once. A pragmatic approach starts with a single high‑pain process where the SOP is stable but the UI is volatile. Choose a process that is run frequently, costs a lot of manual time, and involves multiple steps with occasional judgment calls. For example, vendor onboarding or invoice routing. Deploy a computer use agent to run that process, but keep the same checkpoints that a human would use today. Measure the time saved, the change in exception volumes, and the impact on staff. If the agent reliably follows the SOP and flags exceptions, you can expand to other processes. Over time, you can retire brittle bots in favor of agents that adapt to evolving systems. RPA still makes sense for high‑volume, deterministic backend tasks. Focus your agent pilots on the long tail of work where SOPs, UI changes, and exception handling dominate.
The bottom line
Your SOPs are already a blueprint for human work. Computer use agents can read that blueprint directly, act on any screen, and recover when things go wrong. This approach lets you keep qualified staff in the loop for the decisions that truly need human judgment, while the agent handles the routine execution. The result is automation that stays durable even as your systems evolve.
To see how a computer use agent can follow your SOP and keep humans in the loop, book a demo with the Coasty team.