Macros to Agents: The Cyclical Evolution of Automation
January 29, 2025
Share
The Step Change in Enterprise Automation & GPT Inflection
TLDR;
1. Cyclic Evolution Over Five Decades: Enterprise automation has consistently mirrored computing advancements, from mainframes to SaaS to AI-native architectures. The introduction of agentic AI tools is the latest in a series of paradigm shifts that have shaped this industry, demonstrating that innovation in automation follows a cyclical pattern rather than being entirely novel.
2. AI-Driven Innovation & New Entrants: Tools like Anthropic’s Computer Use, OpenAI’s Operator, and Google’s Mariner represent advancements, enabling dynamic, natural language-driven task automation and integration across enterprise systems, creating new opportunities.
3. Category Evolution: Automation has progressed through clear phases—API-centric iPaaS platforms, RPA for non-API systems, process discovery for opportunity identification, domain-specific vertical solutions, and the consolidation of these capabilities into unified platforms.
4. AI Tier as Business Logic: The next generation of enterprise software will shift business logic from traditional CRUD databases to AI layers. Co-pilots won’t just interact with existing systems, they’ll eventually replace backends by handling logic and decision-making directly within the AI infrastructure. This marks a fundamental change in how applications are built and operated.
5. Enterprise Determinism & Blended Models: Enterprises are fundamentally deterministic, relying on precision workflows. They will incorporate agentic AI for tasks requiring unstructured input/output, reasoning over user screens, or probabilistic decision-making while preserving deterministic logic for high-stakes processes. This balance ensures innovation without compromising operational reliability.
Latest Releases, History & Cyclicality of Enterprise Automation
The evolution from text-based LLMs → Multi-modal models → Action-oriented agents has ushered a new era of agentic software/co-pilots, giving rise to a cohort of companies and tools that aim to automate manual tasks by emulating human action as it’s done today on the browser. A few key announcements include:
1. Anthropic Computer Use is capable of interacting with tools that can manipulate a computer desktop environment.
3. Google’s Project Mariner that can understand user prompts in Gemini and take actions on the user’s behalf on the Chrome browser.
4. Microsoft Co-pilot as a conversational assistant that seamlessly integrates with Microsoft’s app ecosystem leveraging Power Automate as the underlying automation stack.
While the concept of “emulating user actions” on a computer has seen a notable step function in the capability of an agent to understand a user request in natural language, applying chain of thought to break it down to sub-tasks, identifying underlying applications and taking actions on user behalf while analyzing the user’s screen, let’s take a deeper look into how screen scrapers (rules-based or semantic) have evolved over the past 50 years since the evolution of personal computing to understand the inherently cyclical nature of the core technology & industry.
Evolution of the industry over the past 5 decades
Noteworthy mentions include:
Automation Mirrors Computing Evolution: From mainframes to thick clients to browser based web technologies, and now multi-modal language & vision models, the architecture of software autonomy has consistently followed the underlying computing paradigm, constantly evolving in decade-long cycles.
From Scripting to Natural Language: The recent evolution from low-code workflow automation tools to agentic co-pilots mirrors earlier leaps, such as the democratization of system level scripting via VBA Macros in the 1990s. Traditional RPA and iPaaS tools emulated deterministic tasks, but agents like Anthropic’s Computer Use and OpenAI Operator introduce probabilistic intelligence, bringing the capability to break down a complex task into sub-tasks, identifying underlying application and semantically automating over them. They go a step further by alleviating scripts or workflows with simple natural language inputs. They interpret, reason, and act, lowering the technical bar for autonomy.
Shifting Moats: Earlier automation tools built their moats around proprietary integrations with enterprise systems. Today, the advantage has shifted to leveraging human-in-the-loop data to fine tune models for specific tasks. As users engage with these systems, their interactions can continuously improve model performance through memory and RLHF, creating a feedback loop that enhances accuracy and adaptability. Akin to ChatGPT’s “memory,” but applied to enterprises, models get incrementally personalized for specific tasks and thus cost effective as they learn to manage exceptions with minimal change. The moat now lies in combining automation with user feedback, making these systems more personalized, creating higher switching costs than before.
Industry Evolution and Consolidation
Software autonomy has historically centered around enterprises, with tools designed for automating laborious tasks within the enterprise. This industry has been broadly shaped by categories such as iPaaS, enabling programmatic integrations between applications; RPA, which focuses on emulating user actions on screens; process mining, which maps and documents workflows for optimization; and document platforms, designed to extract and process unstructured data like invoices and contracts. While these categories are covered at length in a previous article, Figure 2 below depicts the innovation journey.
The early integration and automation waves.
1. The Rise of iPaaS (1990 – 2005): iPaaS emerged in the late 1990s and early 2000s, driven by the need to programmatically integrate disparate systems, databases, and applications. Tools like TIBCO (earlier generation) and MuleSoft, (newer technology) addressed this challenge by focusing on backend, API-driven integration, enabling programmatic workflows.
2. The Emergence of RPA (2000 – 2010): RPA addressed the gap left by non-integrable legacy systems, automating repetitive UI tasks as a “patch” where APIs weren’t available. Platforms like UiPath and Automation Anywhere advanced the field with GUI-driven tools, empowering non-technical users to build automations without coding expertise. More details here
3. The Rise of Process Discovery and Mining (2005 – 2015): As automation demand grew, companies like Celonis and Kryon simplified the identification of automation opportunities through process discovery tools that mapped workflows, identified tasks for automation, and created process documentation to accelerate enterprise pipelines.
4. Document Automation & Verticalized Solutions (2010 – 2015): Broad platforms like Abbyy, Instabase and Hyperscience expanded document automation, but enterprise commonalities led to specialized vertical SaaS offerings with end-to-end solutions. For example, tools like Rossum, Bill.com, and Klarity automated end-to-end Procure-to-Pay workflows using domain-specific NLP and computer vision models, integrating with specific systems.
5. Category Consolidation (2015 – 2021): In most cases, a single technology wasn’t enough. RPA, API, Intelligent Automation and Process Mining – all were required in tandem to automate complex enterprise workflows. This led to build vs. buy decisions for major providers – a product to platform strategy. While UIPath acquired ProcessGold and Automation Anywhere acquired FortressIQ for process and task discovery, Celonis acquired Integromat to bolster RPA capabilities. Meanwhile, Workato organically blended both RPA & iPaaS capabilities in one platform.
Consolidation, Competition and GPT Inflection
1. Incumbent Automation Bundles & Pricing Pressures (2019 Onwards): As the market exploded, players like Microsoft, Salesforce, and SAP embedded native automation capabilities into their platforms, leveraging existing customer bases and pricing advantages. Microsoft’s Power Platform, for instance, offered low-cost automation as part of enterprise licenses. However, these solutions often lacked integration beyond their ecosystems, keeping broader automation tools relevant.
2. Exits and a few standalone winners: Ultimately, a few players with smaller market share and niche functionality got acquired by incumbents where GUI automation & integrations became accretive to larger providers such as Mulesoft (Salesforce), BluePrism ( SS&C) & Softmotive (Microsoft), while some continue to lead standalone – Celonis, Workato and UiPath. In fact, UiPath went public, surpassing $1 billion in revenue and continues to hold the highest market share in a segment that’s growing at a 20% YoY CAGR.
GPT Inflection: Emergence of Agentic Process Automation
The emergence of GPT 3 in 2022, and multi-modal language models by OpenAI, Mistral and the open source community led to the concept of “AI Agents” that were able to observe and reason a task, generate natural language workflows and code which creates both headwinds and tailwinds for the “workflow automation” industry. While we coverwhat can be done with latest AI capabilities here, here’s how the industry is shaping up:
1. Incumbents bundling Agentic AI: Existing incumbents have been re-positioning their platforms by bundling Agentic AI into their existing platforms, making LLMs available through drag & drop for complex functions such as text parsing, document processing, enhanced screen recording & process discovery.
2. New Entrantsand Verticalization Trends: Adept, Multi-On and others are building browser-native agentic interfaces from scratch, ingesting natural language via chat interfaces, auto generating code to automate workflows and enabling API integrations while “emulating user actions” in remote browser sessions. Kognitos, a leading hyperautomation agentic AI platform has seen tremendous growth as a horizontal leader. Separately, a category of vertical players such as 11x and Harvey.ai have emerged as a key differentiator with domain-specific hyper-trained models.
Short- & Long-Term Implications
There are 3 key transformations that GPT brings to the ecosystem – A shift in the underlying stack of SaaS and legacy applications, a shift in interoperability and introduction of enterprise models.
“The notion that business applications exist is probably where they’ll all collapse in the agent era. They’re essentially CRUD databases with a bunch of business logic. The business logic is going to these agents and these agents are going to be multi-repo CRUD. They’re not going to discriminate between what the backend is, and are going to update multiple databases and all the logic will be in the AI tiers. Once the AI tier becomes the place where all the logic is, then people will start replacing the backend. As we speak, we’re seeing pretty high rates of wins on Dynamics backend and agent use and we’re going to go aggressively and collapse it all whether it is in customer service, or finance & operations. People want more AI native biz apps. Co-pilot to agent to my business application should be very seamless”
Secondly, Given that the business logic now resides within the AI layer built with foundational model layers, we may also see a new paradigm of REST and SOAP APIs (Or let’s call it AI-PIfor a fancy name)that leverage new protocols exchanging payloads in different formats based on the underlying language model infrastructure. This fundamentally changes how the iPaaS industry integrates across multiple business apps.
And finally, while most open source models are trained on the open internet, most enterprise data and applications are proprietary. Training on enterprise data is key for multi-modal models as seen with Adept Fuyu 8B and Cohere Command R and we may see numerous on-premise deployments, contextualizing existing models with enterprise data, or smaller domain specific models trained on vertical datasets such as PaLM 2.
The last thing to keep in mind is that regardless of technology and product cyclicality, Enterprises will always run on deterministic outcomes. They will also continue to have organizational silos and organic growth. How does this work with GPT?In the words of Daniel Dines:
“LLMs are not good at following repetitive steps but LLMs work really well when information is unstructured, and cannot be expressed in rules, with the intent that human input on that part of the process will be removed. Every process has a non-deterministic part and a deterministic part and it makes sense to culminate them into one platform…. Both automation and agents imitate people. Agents will make recommendations. Agents are not going to take action directly. There will be a progression from agent to a human user calling for validation and then calling an action. Most enterprises create a lot of rules-based workflows and precision workflows”
So, what’s next?
As enterprise automation transitions from deterministic workflows to AI-driven systems, incumbents are leveraging their scale, distribution, and established ecosystems to embed AI into their platforms, solidifying their hold on broad, horizontal use cases. However, this shift opens the door for new entrants to innovate, particularly in areas incumbents struggle to serve—niche verticals, domain-specific solutions, and cutting-edge integrations tailored to x unstructured data and complex workflows. Let’s discuss further…
About the Authors
Aditi Charnoubi is a recognized AI and business transformation leader specializing in automation, operational excellence, and strategic growth. As Managing Partner of Tysons Advisory, she helps government and enterprise leaders adopt AI-first strategies to modernize operations, enhance brand equity, and drive increased stakeholder value.
With 20 years of leadership experience, Aditi has scaled AI platforms, optimized multi-million-dollar P&Ls, and led enterprise transformations generating millions in cost savings. She played a key role in a Top 10 US Bank merger and architected AI solutions at scale, supporting millions of users.
A thought leader and board advisor, Aditi partners with industry-leading analysts and expert networks, providing market intelligence and technical due diligence to assess business competitiveness and innovation potential. She actively engages with organizations such as Chief, NVTC, and Public Sector Network, driving advancements in AI-powered leadership, brand growth, and enterprise transformation. Learn how Aditi and her team can drive innovation and transform your organization at TysonsAdvisory.com.
Sid Sahu is a Product and GTM leader with deep experience in enterprise software, specializing in business productivity, automation, and knowledge management. Currently working in Product Management for Search and Personalization at Amazon, Sid has built and scaled enterprise search and workflow automation tools across venture-backed software companies, including Strivr and Automation Anywhere. Passionate about the future of enterprise software, Sid frequently writes and advises startups in these domains.