Application Automation: The Past, Present and Future
Jul 19, 2023
Interested in reading more?
Sign up for our Enterprise Weekly Newsletter.
We'll send you our top, curated content straight to your inbox (along with top industry news, events, and fundings).
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
The world of automation is rapidly evolving. Recent advancements in AI are fundamentally changing the way we work, and have already instilled new behaviors and expectations for enterprise-grade products. Across horizontal and vertical automation, product builders are leveraging Large Language Models (LLMs) to create and automate tasks that were previously perceived as too futuristic.
In this post, I will detail the past, present, and future of automation, along with key takeaways and ideas for how this newfound automaton gold rush could play out.
As always, if you’re working or building in the automation space, we’d love to hear from you.
The Past Generation of Automation
The past generation of automation tools were largely no-code, rules-based engines that aimed to complete some type of business process. These products are primarily horizontal, DIY platforms that are great at solving highly reproducible workflows, interconnected across multiple systems.
To solve these distinct use cases, companies like BluePrism, UiPath, and Celonis improved upon the screen scraping technologies that first came to market in the 1980s and developed Robotic Process Automation (RPA) to help business users automate back office workflows. Similarly, companies like Zapier, Workato, and Tray took a hybrid approach to Workflow Automation by servicing both the business persona and developer to become the connective tissue for data flow within a company.
While these classes of tools service different depths of automation use cases, they provide users with a peace of mind — knowing that the process they recorded will (hopefully) be reproduced.
In RPA, a user could leverage UiPath’s process mining solution to discover and capture their business process, then implement a software bot to mirror the captured process and semi-autonomously arrive at the desired outcome. In practice, RPA targets horizontal, mission critical business processes like claims reconciliations, extracting data from documents, automating order to cash processes, and more. Often, these use cases span multiple tools, data types, and workflows to cut down on hours of manual labor.
As it relates to business outcomes, the most value was derived from automating highly manual, back office workflows. In 2022, UiPath recorded ~$1B ARR by primarily automating these back office processes.
To tap into this, vendors like Uipath and Celonis message their products with relevant, buzzy statements like “improve operational efficiency”, and “allow employees to do higher- level work”.
Conversely, Workflow Automation tools like Zapier help users designate the correct path for data to flow across multiple systems and trigger a response. Often, these use cases could be more lightweight. For example, users could leverage Zapier to send an alert to Slack whenever they get a new user. Or, tools like Workato allow users to connect their entire sales tech-stack like Salesforce, Gong, Clari and more to ensure data is properly routed within an organization.
While we can’t deny the success of these businesses within their respective niches and their place in the automation stack, the potential use cases that were automatable via RPA and workflow differ from the types of actions that are possible by today’s AI native tools.
Today’s AI-Enabled Automation
The recent ZIRP cycle led to a dramatic increase in the number of SaaS tools founded, funded, and scaled over the past 10+ years. This “SaaSification of the Enterprise” has played out across nearly every business function ranging from Sales, Marketing, Customer Success, Finance, and more. Often, these business facing tools digitized workflows, aggregated data from various systems, created user-facing dashboards, enabled actions, and combined it all into one solution fit for a functional use case.
However, with the emergence of LLMs, we’ve unlocked capabilities that previously required too much data or technical skill to make a reality. Now, companies across the maturity stack are all racing to layer on AI-enabled functionality to differentiate their products and add value to users' workflows. We’re seeing this play out as AI-native startups pop up building net-new workflows for users, while incumbents build out their own functionality on top of LLM-powered capabilities.
This rush to LLM adoption is unfolding as both a reactive and proactive strategy — we’re living in a time of constrained buyer budgets and immense competition from a sea of SaaS applications. Especially in tightening capital markets, companies are thinking about how they can capitalize on their market opportunity by differentiating themselves with AI.
As every application software company looks to add AI-enabled capabilities, this generation of automation will look vastly different to the previous generation’s. The past generation required business users to discover automatable use cases and apply lower-level automation capabilities like OCR, and reproducible clicks to back-office business processes. Today, LLMs have enabled vertical automation where every application and workflow across a functional domain is automated. In this generation of automation, we’re bringing automation to the application.
There are two primary differences between the automation tools of the past decade and this generation of automation:
1. Verticalization of Automation: This era will be defined by front office, generative, and action-driven workflows that can be embedded directly into the product. The use cases early adopters first pioneered played squarely into LLMs’ core competencies — LLMs are probabilistic models best used for generating, extracting, summarizing, and entering data. In these early innings, we’ve already seen the emergence of creative automation via Runway, code generation thanks to Github Copilot, and content generation through companies like Stability, Jasper, and more. Beyond that, we should expect to see more niche, but potentially high-value use cases coming to market that can automate the entire workflows within a functional domain (e.g., sales, marketing, finance, etc.).
2. Embedded Automation: Previously, users would need to engage with multiple SaaS tools to complete an action. In this generation of automation, users will be able to automate many use cases & workflows all within a single AI-enabled SaaS application. Products will serve as a command center for use cases and orchestrate data flow across systems of record and other data sources.
While virtually every company races to add AI-enabled functionality, the abundance of AI-enabled tooling will lead to a consolidation of SaaS with the winners being those that can position themselves as a “day in, day out tool”, automate key parts of the user journey and produce a 10x better product experience.
While this might sound like a tall feat, that’s because it is. To put this in perspective, today’s SaaS incumbents and up-and-comers are not the sleepy laggards of old. Many of these players are experimenting with AI and have teams of highly skilled technologists working to implement this technology – Github unveiled Copilot to suggest code in real-time which led to 1M users in 5 days, HubSpot created a Content Assistant to help customers leverage CRM data for content marketing, Zoom launched Zoom IQ for meeting transcriptions and insights, Ramp shipped Ramp Intelligence, an AI-enabled accounting automation product to complement its spend management platform, and the list goes on.
Where startups used to have a competitive advantage in speed to market, many no longer do. With LLMs, anyone can build these workflows. What previously required teams of highly sophisticated ML engineers months if not years to build, is now straightforward enough to only staff teams with two or three savvy hackers. As we iterate on models trained on hyper-specific, high quality data, we will only see the performance of these models improve.
For founders looking to automate key processes in B2B SaaS with net-new products, it’s imperative that they identify greenfield opportunities or build 10x better product experiences for existing markets.
The Future: Autonomous Agents?
The future of automation may look more like the past than one might think. Early adopters are currently exploring the viability of commanding software bots via natural language to complete a number of tasks and automate a variety of use cases.
Dubbed “Autonomous Agents”, companies like Adept, Cognosys, Agemo, Superagent, AgentGPT, Layer, and AIAgent are AI-powered programs able to create, finish, and reprioritize tasks until the agent has achieved a goal, operating with little to no human intervention.
While the idea of Autonomous Agents has been around since the 1990s, the emergence of LLMs have spurred a renewed interest in the idea to automate both vertical and horizontal use cases. Projects like Auto-GPT and BabyAGI have inspired a newfound gold rush for what could be the future of automation. Users can provision agents with access to either a specific suite of tools, or the entire internet, and can decide which tool to use based on user input.
Agents work by breaking down larger scale tasks into smaller parts and leverage memory to control their actions:
On the flipside, here at Work-Bench we’re looking at recent advances through an enterprise lens to better understand how companies can leverage agents to be more productive in their work-life. Many of the enterprise use cases we’ve encountered mirror that of the underlying LLMs — a generative, front-office feel to them where an agent could automate data gathering, summarization, or generation. We’ve seen companies aiming to build horizontally as “a productivity agent to fully augment, then automate your workforce”, to agents building vertically within a specific use case like internal tools or code generation.Still, while startups are split on going horizontal vs. vertical, there are still limitations across both approaches.
Limitations to Productionizing Autonomous Agents
Memory: The context window per agent is limited to one question. The agent will not be able to retrieve answers from previous questions. This restricts the agent's ability to recall past answers and could cause issues with future dialogues for ongoing problem sets.
Generalized Models: While LLMs are powerful, they lack context around specific questions / prompts and can be outperformed by smaller models trained on more relevant data.
Security and Authorization: Models can be misused and harmful, so depending on the severity of the use case, extra guardrails should be put in place to ensure data integrity and permissions management.
Compute Costs: Running third-party models and calling an API will trigger inference cost. Run at scale, these costs could be expensive to a vendor.
Opportunities to Further Agents Potential
Marketplace of APIs and Integrations: When ChatGPT Plugins were released, OpenAI assumed that companies wanted to embed their applications into ChatGPT. The reality is that companies want to embed natural language interfaces that allow users to leverage the power of LLMs with their company’s own data. Agents take this a step further and seek to unify public and private data to complete a task. In order to make this a reality, builders will require a suite of accessible and trusted APIs that agents can call to complete a task. This could look similar to Rapid API or completely new.
Logs, Metrics, Tracing, and Debugging: To begin tackling mission critical enterprise use cases, users need the ability to monitor agent behavior to ensure that when something breaks, the user can quickly view the action and “debug” the task. This could come in the form of third-party monitoring solutions like GenTrace, or be built directly into the agents interface.
Human in the Loop: While some are excited about the availability of fully autonomous agents, we believe that incorporating a human in the loop is the best foot forward for workflow automation type use cases. Allowing agents to work fully autonomously could prove to be risky in certain use cases and lead to serious problems when the agent fails.
Summary + Takeaways
Automation’s natural evolution from horizontal RPA and workflow automation tools to AI-enabled vertical automation to autonomous agents and beyond will permeate into every industry and use case. The power and flexibility of today’s foundation models are enabling each of these automation sectors to improve their technology and usability, allowing users to automate a variety of tasks and functions.
While it’s still early, and many of the low-hanging fruit use cases have been built, we’re excited about founders building AI-enabled application software in niche areas that serve as a wedge into larger use cases and markets. Given Work-Bench’s obsession with enterprise go-to-market, we know there’s still greenfield opportunities when you compare true customer pain points and unique product positioning.
Automation is a Large but Fragmented Market: There will continue to be a place in the market for each automation technology - RPA & Workflow, AI-enabled vertical automation, and Agents - given each has a unique underlying use case. There’s opportunity to build large companies by verticalizing once horizontal automaton use cases.
Automation is Commoditized and How to Win: As existing vendors and new entrants leverage LLMs to underpin their technologies, the biggest opportunity for startups is to improve upon usability and workflow creation of net new use cases. With the widespread adoption of LLMs, the power of automation has been commoditized and product builders will need to differentiate by stitching together specific use cases and improving upon manual and the previous generation of automation.
Autonomous Agents Aren’t There Yet: Autonomous Agents are exciting, but lack guardrails for mission-critical use cases. In their current state, model hallucinations, retrieval, and a lack of monitoring capabilities make it challenging for enterprise customers to adopt agents for high-value use cases.
As we continue digging into this topic, we’re eager to speak with more practitioners and builders in this category. If you’re an early stage startup building in the future of automation, we’d love to hear from you!