Every year someone declares that RPA is dead, first because of APIs, then low-code, then generative AI, and now AI agents. Yet companies keep buying RPA platforms, and the biggest vendors are still growing their automation businesses. So what’s actually going on?
The short answer: RPA as a stand-alone strategy is fading, but RPA as a capability is very much alive. It has become one layer in a larger automation stack, the part that reliably executes steps inside systems that can’t be integrated any other way.
Key takeaways
- RPA is not dead. It is being absorbed into broader platforms that combine bots, AI, APIs and people.
- Its weaknesses are real: brittle bots, high maintenance and no ability to handle messy inputs or judgement on its own.
- AI agents are good at deciding what to do; RPA bots are good at doing exactly the same thing every time. The two are complementary.
- The analyst framing has shifted too: Gartner now also evaluates vendors in a wider “business orchestration and automation technologies” category.
- If you run RPA today, the move is to modernise, not rip and replace.
Why people say RPA is dead
The criticism isn’t baseless. Many early RPA programmes underdelivered for the same reasons:
- Brittle bots. Screen-level automation breaks when an application’s layout, a field name or a login flow changes. Teams often spend more effort maintaining bots than building new ones.
- Automating bad processes. Bots were bolted onto messy, undocumented processes, which made the mess faster instead of fixing it.
- Structured data only. Classic RPA can’t read a scanned invoice in a new layout or understand an email, so large parts of real processes stayed manual.
- Scaling pain. Pilots succeeded, but a few hundred bots needed infrastructure, governance and licences that many companies hadn’t budgeted for.
- Better options for some jobs. Where systems have good APIs, integration tools and workflow platforms do the job more robustly than imitating a user.
Why RPA is still here
Despite those problems, there are good reasons the technology persists:
- Legacy systems are everywhere. Banks, insurers, hospitals and governments still run critical software with no usable API. For those systems, imitating a user is often the only practical integration.
- It is deterministic. When a regulator asks exactly what happened to a transaction, a bot with a fixed script and a full log is easier to explain than a probabilistic AI model.
- It is fast to deliver. A well-scoped bot can go live in weeks without waiting for a system replacement project.
- AI needs hands. An AI agent that decides an invoice should be posted still needs something to actually post it in the ERP. Very often that’s an RPA bot or an API call.
What replaced “RPA-first”: the layered stack
Modern automation programmes combine several layers, and RPA is one of them:
| Layer | What it does | Example |
|---|---|---|
| APIs and integration | Connect systems directly and reliably | Sync new orders from a web shop to the ERP |
| RPA bots | Execute precise steps in apps without APIs | Key a claim into a legacy mainframe screen |
| Document and language AI | Turn messy input into structured data | Read an invoice PDF and extract supplier, amount and PO number |
| AI agents | Plan steps toward a goal and choose tools | Investigate a payment mismatch and propose a fix |
| People | Approve, handle exceptions and judgement calls | Sign off a high-value payment the agent prepared |
| Orchestration | Coordinate all of the above end to end | Run the whole invoice-to-pay process with SLAs and audit logs |
Vendors have reorganised around this idea. UiPath, Automation Anywhere and SS&C Blue Prism all now market “agentic automation”, where AI agents orchestrate bots, APIs and people. Gartner continued to publish its RPA Magic Quadrant in 2025, and it now also covers a broader business orchestration and automation category in which the same vendors compete. See RPA vs intelligent automation for how the layers differ.
Rule of thumb: use an API if one exists, RPA if it doesn’t, AI where the input is messy, an agent where the path varies, and a person where the decision is consequential.
What this means if you already run RPA
- Audit your bots. List each bot with its volume, value, failure rate and maintenance hours. Retire the ones that cost more than they save.
- Replace screen steps with APIs where possible. Many systems have added APIs since your bots were built. Swapping a fragile screen step for an API call cuts maintenance.
- Add AI where humans still handle variation. If a bot hands off every non-standard invoice to a person, document AI can probably absorb a large share of those cases.
- Pilot agents on exception handling first. Agents are most useful on the long tail of cases your rules don’t cover, with a person approving the result.
- Keep governance. Agents increase the need for logging, access control and clear ownership, not reduce it. A Center of Excellence is still the right home for this.
What this means if you’re starting now
Don’t start by buying an RPA platform. Start with the process: map it, fix obvious waste, and then pick the lightest tool for each step. For many small businesses that means a workflow tool such as Zapier, Make, n8n or Power Automate cloud flows (see our workflow automation tools comparison), with desktop RPA only for the one or two systems that have no other way in.
Frequently asked questions
Is RPA still worth learning in 2026?
Yes, but learn it as part of a broader skill set. Employers want people who can combine RPA with APIs, document AI and increasingly AI agents, and who understand process design and governance. Pure “bot builder” roles are shrinking.
Will AI agents replace RPA bots?
For some tasks, yes: agents can operate websites and apps directly. But for high-volume, repeatable steps where accuracy and auditability matter, deterministic bots or APIs remain cheaper and more predictable. Expect agents to call bots rather than replace them.
What is the biggest reason RPA projects fail?
Choosing the wrong processes, ones that are unstable, poorly documented or full of exceptions, and then underestimating maintenance. Our do’s and don’ts of RPA covers how to avoid this.
Is RPA the same as AI?
No. RPA follows rules that a person designed. AI makes predictions or generates output based on patterns in data. Modern platforms combine the two, which is what intelligent and agentic automation mean.
Leave a Reply