RPA follows coordinates. Agents should follow intent.
The last automation wave left enterprises with brittle bots and a maintenance bill nobody budgeted for. Here is what is fundamentally different about computer agents — and what is not.
| Author | Retypd Engineering |
|---|---|
| Published | |
| Reading | 6 min read |
Most enterprises have a drawer full of bots. They were built with good intentions during the last automation wave, and many of them still run. But ask an operations leader how they feel about them and you will hear a familiar story: months to build, constant breakage, and a growing team whose only job is keeping the bots alive.
That story is worth understanding, because the same failure modes threaten the current wave of AI agents if we are careless.
Why bots break
Robotic process automation records where things are: the coordinates of a button, the CSS selector of a field, the position of a column in a table. When a vendor ships a redesign, a field moves, or a pop-up appears that never appeared in testing, the bot does exactly what it was told — in the wrong place.
The result is a maintenance burden that industry surveys routinely put at 30 to 40 percent of total programme cost. Worse, the breakage is often silent. A bot that keys a value into the wrong field doesn’t crash; it just quietly creates a mess for someone downstream.
What “following intent” means
A computer agent like Autofil doesn’t remember that the invoice number goes in the third box from the top. It understands that this step is enter the supplier’s invoice reference, and it finds the right field the way a person would — by reading the label, the layout and the context.
That change has three practical consequences:
- Resilience. When a screen changes, the agent re-locates the target by meaning. Autofil then flags the change for a human to confirm, rather than silently adapting forever.
- Unstructured input. The platform that operates the screens also reads the documents, emails and faxes, through Retypd Extract — so you don’t need a separate document-processing project in front of every workflow.
- Authoring by demonstration. Because the agent learns intent, the people who know the process can teach it by doing it. You don’t need specialist developers translating their knowledge into scripts.
What must not change
Flexibility is a double-edged sword. An agent that can adapt to a moved button can also, in principle, adapt its way into doing something nobody intended. That is why we believe intent-following agents need more governance than bots, not less:
- Bounded planning. Autofil plans only within an approved Playbook. It cannot invent new steps in production.
- Verification. Every write is read back and checked before the next step begins.
- Policy outside the model. Limits and approvals are enforced by Retypd Governance, which the model can neither see nor modify.
- Human sign-off on adaptation. Self-healing proposes; people approve.
The honest comparison
| Legacy RPA | Computer agents (governed) | |
|---|---|---|
| Finds elements by | Coordinates & selectors | Meaning & context |
| When the UI changes | Breaks | Adapts, then asks |
| Unstructured inputs | Separate tool | Native |
| Built by | Developers | Process experts |
| Governance | Per bot, bolted on | Platform-level, built in |
Agents are not magic, and they are not a reason to abandon the discipline enterprises learned the hard way. They are a reason to finally get the return that discipline was supposed to deliver.