Why RPA Programs Stall in Year Two
Robotic process automation sells itself on a clean number. A bot that handles 500 invoices a day replaces roughly four full time positions, costs 30,000 to 60,000 dollars to build, and runs without fatigue. The math works. The pilot works. Then somewhere in year two a lot of these programs quietly stop growing, and the reason is almost never the technology.
The Maintenance Bill Nobody Puts in the Business Case
RPA works at the user interface layer. That is its whole advantage, because you can automate a legacy system with no API, no schema changes, and no vendor cooperation. It is also its whole weakness. The bot is bound to the screen, so a moved button, a renamed field, or a new confirmation dialog stops the run.
SaaS applications make this worse, because they update on their own schedule and do not tell you first. A mature RPA operation budgets 15 to 25 percent of its development capacity purely for keeping existing bots alive. Programs that skipped that line item still discover it, they just discover it as a backlog instead of a plan.
Standardize the Process Before You Automate It
The failure rate for initiatives that skip real planning runs between 30 and 50 percent, and the most common cause is automating a process that was never actually one process.
Discovery usually reveals that what the org chart calls "invoice approval" is a dozen undocumented variations across teams, regions, and time periods. Automate that as it stands and you end up with a bot carrying a hundred exception branches, which costs more to maintain than the work it replaced. Resolve the variations into one standard path first, then build. This step is boring, political, and the highest leverage thing in the entire program.
What AI Actually Changed
Classic RPA needs structured input and deterministic rules, which ruled out most document heavy work. Invoices and contracts vary wildly in layout and terminology between vendors and countries, so they never fit.
Intelligent document processing changed that. Trained models locate and extract the right fields regardless of format, hitting 85 to 95 percent straight through processing on documents that used to require manual review. Process mining changed something else: instead of guessing which processes to automate, you read the system logs and find out how work actually flows versus how people believe it flows.
The newer shift is agentic automation, where the bot plans a multi step workflow and reasons through an unexpected screen rather than falling into a pre-programmed exception path. Every major RPA vendor has shipped or announced this by 2026. At the same time, browser automation frameworks like Playwright and Puppeteer are picking up orchestration features while RPA platforms add real browser engines, so the two categories are converging into one.
Where This Leaves You
Pick processes that are high volume, rule based, stable, and structured. Clean the process before you touch a bot. Budget the maintenance up front and staff it. Treat AI document extraction as the thing that widens what is automatable, not as a way to skip the standardization work.