Does agentic AI replace RPA, or do they serve complementary, fundamentally different automation needs?

Aaron Levie · May 22, 2024

RPA is a frail early preview of agents rather than the same thing, because it performs rote screen-level actions, handles variability poorly, and lacks the intelligence of modern AI models

RPA looks at the screen and performs routine actions; general intelligence applied to business tasks unlocks far more

Scope: agents don't negate the enterprise need for RPA; doesn't require full AGI

13:27 20VC: Box's Aaron Levie on Predictions for the Next Wave of AI: Will Foundation Models Be Commoditised | How the Business Model of SaaS Changes Forever | Startups vs Incumbents: Who Wins | App vs Infrastructure Layer: Where is the Value?

George Sivulka · Jan 22, 2025

RPA is not an AI application in the modern sense — it is essentially simple computation, AI in the ten-years-ago sense

What customers now ask for over 800-page credit agreements or 230-page CIMs is open-ended work like finding inconsistencies or events of default, not copying numbers between fields

25:29 20VC: Why All AI Companies Are Under-Valued | The Future of Foundation Models: Scaling Laws, Generalised vs Specialised, Commoditised? | From Unable to Afford Rent to Raising $130M From Index and Peter Thiel with George Sivulka @ Hebbia

Harry Stebbings · Jan 22, 2025

RPA handles low-skilled, low-level cognitive processes while agents handle high-skilled, ambiguous decisions

Scope: relayed from a prior podcast conversation

26:08 20VC: Why All AI Companies Are Under-Valued | The Future of Foundation Models: Scaling Laws, Generalised vs Specialised, Commoditised? | From Unable to Afford Rent to Raising $130M From Index and Peter Thiel with George Sivulka @ Hebbia

David Luan · Jun 24, 2024

RPA and AI agents are fundamentally different technologies: RPA suits high-volume tasks that always look the same, while agents continuously think, reevaluate and plan at each step toward a goal

RPA is like factory robots following a painted yellow line from station to station, whereas agents are more like full self-driving — replanning at every step to solve the goal

Scope: there are many areas where you don't want variability and should use RPA instead

34:23 20VC: Why Foundation Model Performance is Not Diminishing But Models Are Commoditising, Why Nvidia Will Enter the Model Space and Models Will Enter the Chip Space & The Right Business Model for AI Software with David Luan, Co-Founder @ Adept

Daniel Dines · Dec 18, 2024

Agentic AI does not replace RPA because RPA's sweet spot — rule-based, structured-input tasks spanning multiple business systems and up to hundreds of steps — is precisely where LLMs perform badly

LLMs are not good at following repetitive rule-based steps, just as you would not use an LLM to multiply two numbers; rule-based automations reliably encode company knowledge and work until the underlying system changes

11:54 20VC: UiPath's Daniel Dines on Why Agents Do Not Mean RPA is F***** | Why We Have Reached the Upper End of Scaling Laws | The Future of Work in an Agent World and What Everyone Misunderstands About Enterprise AI

Daniel Dines · Dec 18, 2024

The future shape of work is semi-autonomous agents doing most of the job with humans validating from their inboxes, orchestrated by a rule-based enterprise workflow

Orchestration is how work is done today — rule-based workflows connecting different people to do their jobs

Scope: orchestration layer remains rule-based

19:08 20VC: UiPath's Daniel Dines on Why Agents Do Not Mean RPA is F***** | Why We Have Reached the Upper End of Scaling Laws | The Future of Work in an Agent World and What Everyone Misunderstands About Enterprise AI

Daniel Dines · Dec 18, 2024

GenAI has so far failed in the enterprise because of unpredictability, but will succeed next year once embedded in agentic workflows surrounded by rules and human validation

Putting LLMs inside rule-based enterprise workflows with humans in the loop limits the unpredictable aspect, unlike a chatbot

Scope: success expected within the next year

25:03 20VC: UiPath's Daniel Dines on Why Agents Do Not Mean RPA is F***** | Why We Have Reached the Upper End of Scaling Laws | The Future of Work in an Agent World and What Everyone Misunderstands About Enterprise AI

Daniel Dines · Dec 18, 2024

Enterprises will not buy rule-based and non-rule-based automation from two different vendors, because deterministic and nondeterministic steps sit inside the same end-to-end business process and should live in one framework

Long processes like order-to-cash or procure-to-pay contain both kinds of steps, so one technology must connect and automate all parts — just as you don't run two different Workdays

14:22 Enterprises want one unified framework for deterministic and nondeterministic automation not separate vendors

Daniel Dines · Dec 18, 2024

RPA and agentic AI are not fundamentally different universes because both imitate people doing a process

The common denominator is imitation of human process execution; the difference is agents are more fragile and need more exception handling and retries

16:17 Rpa and agents are not different universes both imitate human process execution

Daniel Dines · Dec 18, 2024

Enterprises will keep building rule-based precision workflows rather than letting judgment be made case by case, just as they don't let people freely decide routing today

Humans also make errors of judgment, which is exactly why enterprises encode decisions in rules

18:21 Enterprises will keep encoding rule based workflows rather than case by case agent judgment

Daniel Dines · Dec 18, 2024

The right way to deploy agents is to start from processes and decompose them into rule-based and non-deterministic parts, not to start from job roles like BDR

Focusing on smaller tasks within a process leads to immediate successes, the same approach that worked with RPA

22:53 Decompose processes into rule based and nondeterministic parts rather than start from job roles

Daniel Dines · Dec 18, 2024

The biggest misconception about agents is that they will be good at rule-based tasks — they are not

Per-step error rates compound: 0.99 success over 100 steps yields a tiny overall success rate, and LLMs give different answers to the same question each time

23:58 Agents are not actually good at rule based tasks due to compounding per step error rates

Daniel Dines · Dec 18, 2024

Agentic AI requires thinking about end-to-end processes rather than piecemeal task-by-task automation, and isolated chat agents are the wrong model

Customers taught him they need the end-to-end picture of the process; value comes from enterprise workflows that connect agents

27:16 End to end process thinking not isolated chat agents is the right agentic model

David Luan · Jun 24, 2024

Incumbent RPA vendors are poorly placed to deliver agent solutions because agents are fundamentally disruptive to their business model

The RPA model involves consultants mapping processes and RPA engineers building workflows over six to nine months, whereas an agent simply observes an end user doing the job and can then be invoked in natural language

35:30 Incumbent rpa vendors are poorly positioned to deliver agents since it disrupts their consulting model

David Luan · Jun 24, 2024 · hedged

The agent market is a fundamentally different and far larger market than RPA — the share of today's work addressable by agents is on the order of 1,000x to 10,000x what RPA can address.

Very little work is addressable by RPA; judging agents by UiPath's outcome is like judging self-driving by the market for autonomous warehouse rovers before self-driving existed.

52:00 Agent addressable market dwarfs rpas by orders of magnitude