ChatGPT
Perplexity
Most mid-market teams that adopted RPA years ago are now fielding the same board question. The agents have arrived, the bots keep breaking, so why not point one at the work the bots could never handle.
It reads as the efficient move, and it is the most expensive one on the table. Point a smarter tool at a workflow that already breaks, and it runs that same failing process faster, which fixes nothing, because the workflow was the constraint all along.
The maintenance bill that sent you looking at agents is still exactly where it was. That cost, and where it comes from, is the decision sitting underneath the choice of model.
What does running an RPA estate really cost beyond the license?
The licensing line on an automation budget describes a small slice of what the estate consumes. A bot built against a screen breaks the moment that screen moves, and someone has to rebuild it, so the standing cost of RPA lives in repair work; the contract was never written to show.
HfS Research interviewed 359 RPA power users and found that licensing accounted for only 25% to 30% of the total cost of running RPA, with implementation, governance, and ongoing upkeep covering the rest.
A separate Forrester study commissioned by Tricentis found that 45% of firms experienced bot breakage weekly or more often, and that firms weak in resilient automation were four times as likely to say they could not control their RPA costs.
Most could not even state their program’s full bill. These are enterprise figures, and the same math reaches mid-market estates built the same way, on bots that break the moment an underlying screen moves.
Before you price a single agent, pull two numbers from your own estate. Your weekly breakage rate and your services-to-licensing split. For most teams, that second number is much larger than the AI budget next to it.
Pointing an agent at a broken workflow inherits the break
Two instincts compete when the bots start to strain, and both cost more than they seem to.
One wraps an agent around the existing bot and trusts the intelligence to cover the cracks. The other tears the whole estate down and rebuilds every process from the ground up.
The wrap-it approach fails first.
An agent placed inside a workflow built around a screen-scraping bot becomes a faster traveler down that same failing path, and the gain remains small because the workflow was the real constraint the whole time.
A more capable runner does not widen a road that was paved for a cart.
McKinsey’s research on the agentic advantage puts a ceiling on it. Agents embedded inside legacy processes deliver roughly 5% to 10% productivity gains, the same narrow band a straight tool swap leaves you in.
The rebuild-everything approach fails for the opposite reason: it spends hard on processes that were never the problem. The real work is to find the line between the two, and your estate already shows you where it runs.
Stay updated with Simform’s weekly insights.
Redesigning the workflow is where the real return lives
The payoff appears only after the workflow is rebuilt around the agent, which is why the tool, on its own, moves the number so little. Capability sitting atop an unchanged process inherits every limit the process already had.
In McKinsey’s customer-service example, reimagining the workflow around the agent rather than threading the agent through it resolved up to 80% of common incidents autonomously. It cut resolution time by 60% to 90%.
That figure pertains to the customer-service context in which it was measured and shows the size of the gap between a swap and a redesign.
A companion McKinsey study explains why so few teams reach numbers like that. Fewer than one in ten enterprises have scaled agents to deliver real value, and eight in ten cite data limitations as the cause.
Gartner arrives at the same conclusion from the integration side. It predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, and advises rebuilding workflows with agentic AI from the ground up, which usually beats threading agents through legacy systems.
Redesign your highest-incident workflows one at a time, because the distance between a 5% swap and a workflow rebuilt for agents is the distance between motion and return.
When should a process remain on deterministic RPA?
Not every automated process should move, and the estate tells you which ones to leave by showing you how they break.
When a bot fails because the interface beneath it shifted while its logic remained stable and rule-bound, the work still belongs in deterministic automation. The fix is a sturdier connection to the underlying system.
When it fails because inputs vary or arrive as unstructured documents, no amount of selector repair holds, and that is where a reasoning agent earns its higher running cost.
AT&T drew exactly this line inside Microsoft Power Automate. Its team put AI Builder to work telling invoices apart from permits, a classification no rule-based selector could do reliably, and the flow now runs at roughly 99% accuracy with a projected 3,500 hours saved each year.
Microsoft frames the same division in its platform guidance. It keeps deterministic RPA for rule-based high-volume work and for bridging legacy apps that lack APIs.
Unstructured inputs go through AI Builder, and goal-based agents are reserved for objectives that scripted steps cannot follow.
Sort each process by how it fails before you sort it by which department owns it, and expect a good share of them to land as a hybrid of both.
A done-right rebuild retires the bill it replaces
Re-architecting the right processes shrinks the estate, and that is the part most business cases leave out.
When the work worth moving consolidates onto a single governed platform, the redundant bots and the service contract keeping them alive switch off along with it. The savings come from two places at once: the new capability you gain and the old cost you no longer carry.
A Forrester Consulting model commissioned by Microsoft, built on a composite 30,000-employee enterprise, credited $9.5 million in three-year savings to retiring legacy automation tooling as the organization migrated processes off its incumbent tools, for a 248% return.
Legacy-tool retirement accounted for close to a sixth of the modeled benefit, which means a real share of the payoff comes from switching the old estate off before the new capability adds anything.
That is an enterprise-scale modeled composite, so treat the numbers as a directional ratio for a mid-market estate rather than a forecast.
Scope the move to the processes whose running cost justifies it, retire the rest on a deliberate schedule, and a large part of the transition funds itself out of the spend you already committed.
Ungoverned agent sprawl is the RPA story replayed at higher speed
The discipline matters because the estate you are about to build grows faster than the one you are replacing. The same instinct that let RPA bots multiply without an inventory will let agents multiply the same way, and agents act on their own authority once they are running.
Deloitte surveyed more than 3,000 leaders and found that only 21% have a mature approach to governing autonomous agents, even as roughly three-quarters expect to be running them within two years. Deloitte describes the result as agents scaling faster than their guardrails.
That gap is the original RPA problem returning at speed, the same missing inventory, and the same unknown cost, now spreading through systems that take action without waiting to be asked.
The way to keep the new estate from becoming a second maintenance tax is to set up the agent inventory and the governance around it before the estate forms, not once it has already sprawled.
Would you rebuild the process if the agent cost as much as the bot does?
The honest answer to that question is the whole decision. If the agent in front of you arrived carrying the same maintenance bill your bots did, the processes you would still choose to rebuild are the ones whose running cost earns the move, and the others can stay where they are.
That test is answerable today with two numbers you already hold, your breakage rate and your services-to-licensing split, and it is what separates an estate that compounds value from one that compounds cleanup.
Simform helps mid-market teams re-architect RPA estates into agentic automation through its Agentic RPA services, so the work worth rebuilding moves to agents and the brittle bots stop billing you for upkeep.
