About eight years ago, enterprises were promised a digital workforce. Software robots would click through screens, copy data between systems, and free people for higher value work. Every large company bought some. Many of them are still paying people to keep the robots running.
That was robotic process automation. The pitch for AI agents today is nearly word for word the same, and the industry is pretending this is the first time it has heard it.
The rebrand is already measurable
Gartner put a number on it last year. It estimated that of the thousands of vendors selling agentic AI, only about 130 are real. The rest are what it called agent washing: existing assistants, chatbots, and, notably, RPA products relabeled as agents without the capabilities to match.
Read that list again. RPA is being rebranded as the thing that was supposed to replace RPA.
The same release predicted that over 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Those three reasons are the RPA postmortem, recycled.
How the last robot actually failed
RPA did not fail because the technology did not work. It worked, in the demo, on the day it was built. It failed in month four.
A Forrester study commissioned by Tricentis, itself a vendor in the space, so weigh it accordingly, found that 45 percent of firms dealt with bot breakage weekly or more often. Fewer than one in five were very effective at keeping automation resilient. Most bot logic required custom scripting, and most programs needed advanced programming skills to maintain.
That is the real story of RPA. The cost was never the license. It was the upkeep. Every bot encoded a snapshot of how a process worked on the day it was recorded, and every change to a screen, a field, or a policy silently broke it. Companies discovered they had not removed manual work. They had moved it to a small team whose whole job was fixing robots.
Why agents fail more quietly
The agent industry's answer is that this time is different, because agents reason instead of following a script. When the screen changes, the agent adapts.
That is true, and it is exactly the problem.
A broken RPA bot was loud. It stopped, threw an error, and someone got paged. The failure was visible and the fix was obvious. An agent that meets an unexpected situation does not stop. It improvises. It produces something plausible, writes it somewhere, and moves on. The breakage does not disappear. It changes from an outage into a quality problem, and quality problems are discovered weeks later, downstream, by someone who has no idea an agent was involved.
RPA taught enterprises to budget for bots that break. Agents require budgeting for bots that are confidently, silently wrong some fraction of the time, and almost nobody has a line item for that, because almost nobody is measuring it.
The pattern to watch
The RPA cycle ran like this: a sweeping promise, a burst of pilots, a few showcase wins, then a long plateau where programs quietly stalled because the maintenance burden outgrew the savings. The firms that came out ahead were the ones that treated automation as an operational discipline, with central ownership, monitoring, and someone accountable for every bot in production. The Forrester data pointed the same way: centralized programs scaled, scattered ones did not.
There is no reason to think agents escape that cycle. There is good reason to think the plateau arrives faster, because agents are easier to launch than bots ever were. Anyone can stand one up in an afternoon. That makes the pilot stage cheap and the governance stage expensive, which is the exact shape of the last bust.
So skip the question vendors want you asking, which is whether the agent can do the task. In the demo, it can. Ask the RPA question instead: who notices when it gets this wrong in month four, and how? If nobody can answer that, you are not buying a digital worker. You are buying the same maintenance problem as last time, with the error messages removed.