The Next Generation of Productivity: Generative Process Automation
Key Takeaways From this Piece (TLDR):
- Despite numerous evolutions, RPA systems still face several limitations to implementation and reaching end-to-end process automation
- Pre-trained LLM models dramatically lower the barrier to entry for process automation
- Generative AI will solve issues around discovering and optimizing the most disruptive workflows by partnering process mining capabilities with autonomous code generation
- Orchestrating multiple models and prompt engineering will empower end-to-end fully-autonomous agents
- Model reinforcement will allow bots to become smarter over time and even mimic individual-level problem solving
- Could your position be entirely eliminated by artificial intelligence? It could be one of the 300 million jobs at risk of being completely replaced by Generative AI, a shift predicted in a Goldman Sachs Research report released last month.
Could your position be entirely eliminated by artificial intelligence? It could be one of the 300 million jobs at risk of being completely replaced by Generative AI, a shift predicted in a Goldman Sachs Research report released last month.
This isn’t the first time we’ve been promised a world where machines do all the work for us. McKinsey’s 2017 report foretold an astounding 800 million jobs would be replaced by 2030, driven predominately by swift advances in Robotic Process Automation (RPA). RPA has seen massive enterprise adoption, but has decelerated with major players in recent years. Those speculations about enterprise-wide automation by McKinsey and others have fallen flat. Even still, with the near vertical pace of innovation now pushed forward by Large Language Models (LLMs), it is hard not to get swept back into the hype.
Although enterprises have begun to test the waters of standalone Generative AI products, we’ve yet to see mass adoption due to concerns around security, explainability, and hallucination (generating confidently incorrect answers when it doesn’t understand the question). Given enterprises’ long history of engagement with RPA and its adjacent solutions, adding GenAI to battle-tested software might offer the right wedge to gain trust. It also has the potential to address the shortcomings of previous automation efforts to exponentially accelerate our pace to a Fully Automated Enterprise. Let’s explore how enterprises have evolved from RPA to Intelligent Process Automation, where these evolutions still fall short, and how GenAI will potentially fill these gaps.
The Evolution of Enterprise Automation
RPA emerged in the late 1990s / early 2000s to prompt computers to replicate, coordinate, and replace highly repetitive tasks that were low-skill but high-cost. Synchronizing information into database systems and generating reports were ideal tasks for bots. In the decades since, step-changes in augmentation have made RPA easier to adopt for more complex tasks.
1. Cognitive RPA: Early 2000s
In the beginning of the 2000s, linking data from an employee’s expense report to the HR department’s accounting system looked like progress. However, data captured in the system wasn’t recorded, and it couldn’t be moved from one system to another. This led to the development of screen scraping abilities that were powered by an early AI model called Optical Character Recognition (OCR). Data was scraped, stored, and then transferred to the next critical application, and each step was wholly coordinated.
RPA’s Biggest Flaw:
- Implementation: Enterprises interested in implementing present-day RPA still often rely on consultants to take that leap. Systems integration consultants require months to map out coordinated processes, build connections between those systems, and train business leaders to maintain the systems with as little human intervention as possible. Consultants are integral in determining which processes and providers are best aligned with the scope of the project, and what technical resources are needed that the enterprise may or may not have.
How Will GenAI Solve This? Generative Process Mining:
Process Mining is already used by enterprises today looking to accelerate their time to value with enterprise automation. In its current state, it offers a view into the world of machine-generated data, allowing enterprises to identify where the biggest pains and potential gains in productivity live. Rather than rely on an outside-in perspective, process mining gives data-driven evidence of where enterprises are losing money or leaving it on the table because of broken processes. However, once these inefficient processes are identified, they still require implementation consultants or RPA solutions to write the code to automate them further adding to the time to value.
The next critical step is to combine Process Mining with Large Language Models. This will generate code snippets ad hoc, or design new processes from scratch so machine-driven insights come to life. Generative Process Mining would first identify inefficient patterns in user behavior at scale, then map optimizations, outline potential edge cases, preempt process breaks, and perform significant predictive maintenance.
Companies like Orby.ai are already combining process mining, GenAI, and RPA to bypass the need for implementation consultants. Orby observes worker behaviors to mine for the most repetitive tasks. Without the need to hire implementation consultants or allocate internal developer resources, Orby designs and orchestrates automation scripts to immediately augment or replace mundane processes.
2. APIs and Integration Platforms Go Mainstream: 2010s
In the 2010s, metadata management and logs became standard to software infrastructure. This combined with universal APIs enabled applications to read and write to each other. Integration platforms (IPaaS) like Workato, Zapier, and Tray.io popped up, and RPA’s capabilities cultivated even more connections and multi-step automations.
Integration Platforms’ Biggest Flaw:
- Intelligent Routing: Automations from integration platform players are dependent on the triggers (when this happens, do that), primitives (action types, i.e. copy data, create new account, send notification), and connections (which programs can communicate natively) within scope. While this can handle highly repeatable tasks with predictable behavior, mapping complex workflows with shifting contexts becomes time consuming and computationally expensive. Rather than spending time to map out every possible permutation of expected behavior, a better solution would be intelligent systems capable of adapting on the fly.
How Will GenAI Solve This? Self-driving Software
If you’ve had the unfortunate experience of building macros in Excel, you’re familiar with the record button. Once clicked, Excel will record each action taken by the user and systematically convert each step into code. While you might not know how to program in Visual Basic, messing around with the record button and attempting different tasks will quickly give you an idea of the syntax to be replicated or modified for different outcomes. Sadly, the record button only exists in Excel, but foundational models like Adept AI’s ACT-1 and fully autonomous open-source agents like Auto-GPT are teaching machines to replicate human behavior like this in software interfaces.
While the “Turing Test” has previously been limited to judging an AI’s ability to pass for a human through conversational intelligence, Generative AI is quickly gaining the ability to navigate software like any digitally-native knowledge worker. Responding to highly contextual tasks with human-level accuracy has been out of reach for RPA. Currently RPA uses computer vision and machine learning to handle repetitive tasks like filling out forms or generating reports. But Universal AI collaborators open up higher order intelligence tasks that require dynamic, sequential processes such as flagging, escalating, and remediating fraudulent claims.
3. No-Code Orchestration: 2015–2020
More recently, the need to handle even more complex processes without the additional developer resources prompted the rise of low-code and no-code orchestration platforms like Instabase who developed drag-and-drop visual interfaces that mapped applications end-to-end.
No-Code’s Biggest Flaw:
- Low-code Requiring Code: Similar to IPaaS, No-Code players have defined actions and triggers. As soon as a new connection or task is introduced, that either requires the vendor to build that functionality natively into the solution, or rely on internal resources to build glue code (code that connects systems together) to fill gaps.
How Will GenAI Solve This? No-Code + Copilot = “Co-Code”
Companies like Hyperscience and Instabase offer process automation solutions through composable blocks of models, actions, triggers and user interfaces. No-code workflow orchestration was designed to address the hurdles for implementing and updating process automation without the need for systems integrators or significant developer upkeep. Although these preconfigured components do lower the barrier to entry for automation and increase time to value, they still suffer from issues with extensibility (i.e. how can you expand a process if there is no pre-configured block to handle it). Building out new connectors for new data sources or linking multiple processes together for additional scope requires a custom buildout from the vendor’s own business services, and prevents them from truly becoming self-serve solutions.
An already feasible Generative AI approach to workflow mapping could be done with simple language interfaces. While they still require some tweaking, Copilots (AI assistants that sit adjacent to existing workflows) fluently generate code snippets. Constructing a recipe for a novel process would be limited to knowing how to properly engineer the prompt. Moving forward, GPT-4’s ability to understand images makes it easy to see a world in which an internal process designer lays out an event-model system in a charting tool like Miro, then converts the connections to code that executes tasks either via plugins trained on LLMs or existing APIs.
To improve the scope of their automations, existing no-code solutions should introduce Copilots alongside their orchestrators. No-code still provides a level of visual specificity that language interfaces cannot. A real-life example of this is something we’ve all undoubtedly faced: teaching our parents new technology that they cannot understand so we have to either draw it out for them or just do it ourselves. In the future, designing and orchestrating application workflows will resemble today’s conversations between product managers and developers, but with far faster iteration cycles and likely far less frustration between both parties.
4. Intelligent Process Automation: 2020-Present
While “generalist” RPA solutions had strong performance on prescriptive tasks that did not require human reasoning, its performance was limited on more specialized and domain specific processes. For example, while a horizontal RPA solution could transcribe a bank customer’s loan repayment history it could not compute a credit score or predict a probability of default. Prediction required specially trained models that inherently became vertically focused. Companies like Inscribe, which addresses document fraud, and Ikarus, which manages accounts payable, emerged to offer better performance within niche markets.
Intelligent Process Automation’s Biggest Flaw:
- General vs. Specific Intelligence: Tasks break when intelligent systems lack sufficient context to handle task complexity. Humans are brought in to share their training and experience to manage workflows that are out of model.
How Will GenAI Solve This? The Second Brain and the Personal Digital Twin
We’ve been experimenting with creating digital versions of ourselves for quite some time. In 2013, A Verizon employee outsourced his entire job to a Chinese consulting firm before ultimately being discovered and fired. Generative AI is already edging closer to bringing your brain to bytes. Companies like Character.ai have already created convincing clones of both fictional and actual people to answer questions and generate novel perspectives.
AI hackers and avid note takers have already begun using GenAI-powered knowledge bases like Mem.ai to act as their second brain. While computers are much better at certain tasks out of the box, like reading the entire library of Congress, without training it would not be able to answer highly subjective, nuanced questions like, “What are your personal thoughts on the politics and law of the Second Amendment?”
VC and AI hacker Yohei Nakajima created a personalized chatbot “Mini Yohei’’ to answer a vast number of potential questions from companies that might pitch him, portfolio companies looking for support, or LPs asking for clarity in a volatile market. Companies like Personal.ai are already building individualized baby GPTs for each user. Imagine a world where instead of pointing an “Out of Office” auto-response to your personal cell number, you let your digital avatar triage any after hours emails directly. Go ahead and have that extra margarita on your vacation. Your digital twin will agree you deserve it.
The Fully Automated Enterprise
Generative AI presents the most compelling case for systemic automation and solves many of the biggest current bottlenecks by giving bots the ability to optimize, replicate, and most importantly, reason. Systems that previously took significant time and expense to implement will use pre-trained, API-accessible models to become increasingly self-service. Tasks that were limited in scope or scalability will see exponential productivity gains as software learns how to orchestrate itself, handle out of model exceptions, rapidly switch contexts, and improve via feedback-driven reinforcement. To summarize:
- RPA companies should combine Generative AI and process mining to not only identify where the biggest productivity gains live, but also proactively create code to replace tasks without human intervention
- Workflow automation and integration platforms dependent on APIs should leverage task-driven agent approaches to minimize human-dependent triage and maintenance
- No-Code orchestrators should allow text-based Copilots to extend capabilities
- Leveraging knowledge bases will enrich intelligent process automation with deeper contextual problem solving
Whether you’re an upstart or an incumbent, if you’re looking to augment or replace human processes, designing your future product roadmap squarely around GenAI is imperative.
I’m Nick Giometti, a VP on Geodesic Capital’s investment team. Reach out to discuss the future of data and all things infrastructure. If you are a founder looking to take your business international, I’d love to hear from you.
