From Macros to Agency: A Short History of Automating Knowledge Work

The ambition to automate knowledge work did not begin with artificial intelligence. It began with a spreadsheet, and before that with a magnetic tape, and before that with a clerk who wished the ledger would add itself. The history of office automation is a single long argument with a recurring twist: each generation of tools automates the previous generation's drudgery, declares victory, and then discovers that it has merely relocated the drudgery into maintaining the tools. Agentic AI is the newest entrant in this argument. Whether it breaks the pattern or repeats it is the most interesting open question in enterprise software, and the only way to reason about it is to know the history that produced it.

The first automation was a number that recalculated itself

The personal computer earned its place on office desks not through games or word processing but through a single application. VisiCalc, released on 17 October 1979 for the Apple II, is remembered as the original "killer application," the program that "turned the microcomputer from a hobby for computer enthusiasts into a serious business tool." More than a quarter of Apple IIs sold in 1979 were reportedly bought to run it. The reason was simple and profound: a spreadsheet automates the propagation of consequences. Change one number and every dependent number updates itself. That is automation in its purest form, the elimination of the manual recalculation that had defined accounting for centuries.

Then came the macro: the recording of a sequence of actions for later replay. Lotus 1-2-3, released on 26 January 1983 and the first killer app of the IBM PC, shipped a macro system that by the late 1980s was "the world's most popular application-development language." Microsoft generalised the idea with Visual Basic for Applications, which first appeared in Excel 5.0 in 1993 and let one Office application drive another. The macro encodes a crucial and fragile idea that will recur throughout this story: automation as recorded steps. Record what a human did, replay it forever. It works beautifully until the thing being automated changes shape.

Two other lineages ran in parallel. Electronic Data Interchange, the computer-to-computer exchange of standardised business documents, automated commerce between companies, with the first operational system often cited as London Heathrow's cargo scheme in 1971. And screen scraping, "the programmatic collection of visual data from a source," emerged to wring data out of legacy mainframes that had no API, by emulating the keystrokes of a human at a terminal. Hold that last one in mind. Screen scraping is the direct technical ancestor of the automation wave that would arrive thirty years later.

The discipline of the process

By the early 1990s, organisations had automated individual tasks and wanted to automate the connections between them. This produced Business Process Management, the discipline of discovering, modelling, measuring, and automating the end-to-end processes that thread through a company. BPM needed a shared language, and got one: Business Process Model and Notation, version 1.0, released in May 2004, with the more powerful BPMN 2.0 arriving in January 2011 and eventually standardised internationally as ISO/IEC 19510.

BPM was a genuine advance and a partial one. It could model and orchestrate the steps that already had system support, but the messy human-to-system handoffs, the copying of a value from one screen into another application, remained stubbornly manual. That gap is where the next wave made its fortune.

The robots that watched and repeated

Robotic Process Automation filled the gap with a deceptively simple idea: a software robot that "develop[s] the action list by watching the user perform that task in the application's graphical user interface (GUI) and then perform[s] the automation by repeating those tasks directly in the GUI." The same Wikipedia entry notes the lineage plainly: RPA "has been around for a long time in the form of screen scraping." The 1960s mainframe trick had become a billion-dollar industry.

And it was a real industry. Blue Prism was founded in 2001 and was eventually acquired by SS&C for around 1.6 billion dollars in March 2022. UiPath, founded in Bucharest in 2005, grew into the category leader and went public on the NYSE in April 2021, raising 1.3 billion dollars in one of the largest US software IPOs in history. Gartner pegged worldwide RPA software revenue at 2.9 billion dollars for 2022, up 19.5 percent year on year. The dream of the self-running office had found its commercial form.

But RPA inherited the macro's fragility along with its power, and the inheritance was fatal to many projects. Because the robot mimics clicks on a specific interface, it breaks when the interface changes. The encyclopedic description is unsparing: "Current RPA solutions demand continual technical support to handle system changes, therefore it lacks the ability to autonomously adapt to new conditions. Because of this limitation, the system sometimes needs manual reconfiguration, which in turn has an effect on efficiency." This is the macro problem at enterprise scale. Automation built on recorded steps is automation that shatters when the steps move.

The disillusionment, in numbers

The brittleness was not a theoretical concern; it showed up in project outcomes, and the people best placed to know said so out loud. EY, describing itself as one of the largest RPA consultancies in the world, wrote in its report Get ready for robots: "While RPA can transform the economics and service level of current manual operations, we have seen as many as 30 to 50% of initial RPA projects fail." Read that again. A leading vendor of the technology reported that up to half of first attempts failed, and was careful to add that "this isn't a reflection of the technology" but of common mistakes in applying it.

The lesson of the RPA disillusionment is not that automation does not work. It is that automation built on imitating surface behaviour, rather than understanding intent, carries a maintenance tax that compounds until it consumes the savings. The robot that only knows what to click cannot cope when where to click changes.

A partial remedy emerged in the form of intelligent document processing and "intelligent automation," which bolted machine learning, OCR, and natural language processing onto RPA so that bots could, as one vendor puts it, "not only recognize and extract text from documents" but "also understand the context and meaning" (Automation Anywhere). This helped with the input side of brittleness. It did not change the fundamental architecture, which still followed a predefined workflow rather than reasoning about a goal.

The turn from steps to goals

This is the precise hinge on which the current era turns, and it is worth stating in the sharpest possible terms. The Wikipedia entry on RPA draws the contrast almost too neatly: RPA "is based on automation technology following a predefined workflow," whereas AI "is data-driven." Anthropic's engineering essay Building effective agents, published 19 December 2024, sharpens it further into the distinction that defines the field. Workflows are "systems where LLMs and tools are orchestrated through predefined code paths." Agents are "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks," operating as "LLMs using tools based on environmental feedback in a loop," gaining "ground truth from the environment at each step."

There is the whole history in a single contrast. The macro, the screen-scraper, and the RPA bot all follow recorded steps. The agent pursues a goal and adapts the steps to the environment it actually finds. When the interface moves, the brittle automation breaks; the agent, in principle, looks at the new screen and figures out where the button went. And the integration problem that plagued every prior generation, the bespoke connector for every system, is being dissolved by the Model Context Protocol, Anthropic's November 2024 open standard for connecting AI systems to tools and data through one interface rather than many.

The scale of what is at stake, and the warning the history gives

The prize is enormous, and the most-cited estimate is sober rather than breathless. The McKinsey Global Institute's January 2017 report A future that works found that about half of the activities people are paid to do globally could be automated by adapting currently demonstrated technologies, equating to almost 15 trillion dollars in wages, while crucially noting that fewer than 5 percent of occupations could be fully automated and about 60 percent had at least 30 percent of activities that were automatable. Read carefully, that is not a prophecy of mass redundancy. It is a forecast that work gets recomposed, task by task, rather than eliminated whole.

But the history issues a warning that the excitement tends to drown out. Every prior automation wave promised to think and ended up merely repeating, and the gap between those two verbs is where the failures lived. If agents are deployed as RPA was, pointed at brittle surfaces, trusted without supervision, and measured by demos rather than maintenance cost, they will repeat the RPA disillusionment with larger budgets. The escape is the discipline Anthropic itself counsels: "find the simplest solution possible, and only increase complexity when needed." An agent that reasons about a goal is genuinely different from a macro that replays a recording. It is not automatically more reliable. It is more adaptable, and adaptability is only an advantage if the system is built to use it and governed to be trusted.

The line from VisiCalc to the agentic enterprise is forty-seven years long, and it bends in one consistent direction: from automating the calculation, to automating the steps, to automating the pursuit of the goal. Each step up that ladder bought more leverage and demanded more judgement about where to apply it. The agent era will be no exception. The tools that record steps break. The tools that pursue goals are about to meet the same wall the macro met, and the only question that matters is whether we have learned enough from the last forty-seven years to build them so they bend instead.

Sources and further reading

• Early automation: VisiCalc, Lotus 1-2-3, VBA, EDI, Data scraping
• Process automation: Business Process Management, BPMN
• RPA: Robotic Process Automation, Blue Prism, UiPath, Gartner RPA revenue 2022 (ITPro)
• RPA failure rate: EY, Get ready for robots
• Intelligent automation: AWS, Intelligent Document Processing, Automation Anywhere, IDP
• The agent turn: Anthropic, Building effective agents, Model Context Protocol
• Automation potential: McKinsey Global Institute, A future that works (January 2017), executive summary
• Images: VisiCalc (PD), BPMN diagram (PD), IBM System/360 console (CC BY-SA 4.0)