Automation Thoughts - A Three-Axis Model

This blog post suggests a three-axis model for categorizing automation initiatives. It is meant to help the reader understand the impact of different types of automation on organizations, the space of possible job-altering automation strategies, and the differences between historical trends and modern automation initiatives. It may also aid speculation about future patterns of automation.

The concepts discussed in this post are discussed in a very linear style: each section builds on the previous one. This may not constitute a good reading experience.

I’ll try to update this post as I continue to think about these concepts, and demonstrate a less linear style in future posts.

What is Automation?

Let’s briefly discuss what distinguishes automation from other forms of technological improvement.

The term automation must refer to a technology that independently handles some of the micro‐steps required to complete a task: rather than having humans perform the full series of mental or physical sub-tasks, these steps are organized into whole actions which can be easily delegated to a machine through its interface. If this weren’t the case, we’d just be talking about tools.

This becomes clearer when we consider the case in which a worker uses a hammer to drive nails. Although the use of a hammer aids the process of completing the task, it does not automate the task. This is explained by the fact that the worker must still swing for each nail, controlling direction and force. Therefore, an automation can not merely make a task easier or faster to complete: we must consider the degree to which the technology abstracts over the necessary steps towards task completion before concluding that it is an automation.

Automation also implies that, whilst human operators no longer needs to manually perform each step (like calculating each digit in arithmetic), humans must shift to a higher-level role: determining how to operate an interface (like deciding which series of buttons to press on a calculator), checking if outputs make sense (like verifying the calculator’s sum), or making judgment calls the system can’t handle (like deciding when calculation is needed). Where an automation appears to operate with complete independence, that higher-level role was occupied by the human operator who set the automation in motion.

I’ll use these observations to suggest an interim definition for automation, which we will refine as we continue: an automation is an interface to a system capable of performing useful actions on behalf of a human, that abstracts much of the mental or physical requirements of task completion away from the human operator.

The Three‐Axis Model, introduced in this post, primarily applies when a system automates the primary tasks that define a paid job role. It only concerns systems of automation which significantly alter the nature of the job, resulting in workforce impacts (discussed later) such as retooling, transforming, or removing entire jobs. It is less applicable to systems that automate the secondary tasks associated with a job, which are only part of the role when they are necessary for the completion of the primary tasks associated with the job.

What’s The Difference Between Secondary And Primary Tasks?

Consider the impact of typewriters on secretarial work. Although writing out text on paper may have consumed a significant portion of a secretary’s time, it would never have been part of their role if it were not necessary for their their primary tasks to be completed: composing, organizing, and managing business communications. Therefore, the task of writing out letters was secondary to the role of a secretary. The role remained (mostly) unchanged: the same personnel continued doing largely the same work with the same workflow, now using the interface of a typewriter.

Why Is “Extent of Automation” Not A Helpful Concept?

If only minor tasks are automated, the result might merely adapt the role: the same employees remain, adjusting their workflows or learning to use the system. Conversely, if primary tasks (the main reason the job exists) are automated, the entire role may be transformed into something quite different or be removed altogether. Yet, if the automation of primary tasks is considered “partial”, it’s unclear how much of the role will be transformed or removed as a result.

Even when automation requires skilled human operators, it can eliminate roles entirely if it covers core responsibilities. For example, if AI automates 80% of an accountant’s primary tasks, the remaining 20% may or may not be absorbed by higher-level management positions - effectively removing the original role. This illustrates why the concept of “extent of automation” is unhelpful: we must discuss how much an automation reduces the time and cost of completing primary tasks separately from its actual workforce impact.

These nuances tie directly into Axis 3 (Workforce Impact): Adapt / Transform / Remove.

How Does Automation Relate To Changes In Working Strategy?

The interface of an automation provides the operator with the ability to select predefined sequences of steps, trigger the automation to execute those sequences, and generate measurable progress on the task. Through this interface, the operator’s role shifts from performing individual steps to choosing which automated sequences to invoke. These options are sometimes not isomorphic to the operator’s manual process, and thus the operator may be forced to choose a different strategy towards completing the task than they would have chosen without the automation.

Here’s an example: a brute-force search tool that presents false positives replacing a human operator’s manual screening process. The machine performing the screening may present a list of potential matches, but the operator must choose which ones to investigate further, which involves manually checking each potential match. We know that this is different from the operator’s manual process, since if the actions taken by the machine were isomorphic to the operator’s manual process, the operator would not need to choose which matches to investigate further.

It’s not merely that introducing an automation reduces the number of steps required by humans for the task. When a new strategy for task completion is required, it transforms the practical aspects of completing the task from the operator’s perspective, which further complicates our previous discussion about determining the extent of an automation.

These nuances tie directly into Axis 2 (Oversight): Manual / Supervised / Autonomous.

---

The Three‐Axis Model and Its Rationale

This model is organized around three questions. By concatenating the answers, we can create a “label” for any automation initiative:

1. Ownership: Who builds and controls the system?
   • In‐House: Internal teams develop it and run it.
   • Contracted: An external vendor develops it, then hands it off for the organization to run.
   • Outsourced: A third‐party handles both creation and ongoing operations.
2. Oversight: How much human oversight is involved in day‐to‐day operation?
   • Manual (Human in the Loop): A human must explicitly guide or approve major actions.
   • Supervised (Human on the Loop): The system runs on its own most of the time; humans monitor and may intervene.
   • Autonomous (Human out of the Loop): The system requires minimal or no routine human involvement.
3. Workforce Impact: What is the impact on employees who used to do these tasks?
   • Adapt: The same employees remain but adapt (learn new skills, adjust workflows).
   • Transform: The role changes drastically; many current staff are not a good fit, so partial displacement or redeployment occurs.
   • Remove: The job is eliminated altogether (the system’s automation of core tasks makes the role obsolete).

Example: “In-House, Supervised, Adapt” refers to a system that is built and operated by an internal team, with human oversight required for major actions. The workforce remains, but employees learn new skills and adjust their workflows.

Independence of Axes

For this model to be useful, it is important for the axes to be (mostly) distinct.

As we saw earlier, an automation’s interface design directly influences task completion strategies, which in turn determines both the Workforce Impact (by dictating skill requirements) and the Oversight needed (through the need for human judgment at decision points and the complexity of interventions).

However, these axes remain worth discussing separately. I’ll demonstrate that this is the case:

1. Ownership doesn’t tightly correlate with Workforce changes. One might imagine an “Outsourced” system to which employees “Adapt”, if the external SaaS product only automates minor tasks. Conversely, an “In‐House” system could completely Remove roles if it fully automates the core function.
2. Oversight doesn’t tightly correlate with Workforce changes: a non-autonomous automation might be used to adapt, transform, or remove roles, as we illustrated with the example of accountant roles possibly being absorbed into higher-level management positions.
3. Any degree of Oversight can co-occur with any Ownership model.

Therefore, there are twenty-seven possible automation profiles in this model.

Why This Model Is Useful.

Firstly, there is a clear relationship between the essential aspects of how the automation works - specifically how actions are presented to the operator via the abstraction mechanism - and the two main axes of the system: Oversight and Workforce Impact. The abstraction style directly determines oversight requirements, with more complex abstractions requiring frequent operator decisions leading to Manual or Supervised models, while simplified abstractions that eliminate decision points enable Autonomous operation. This same abstraction mechanism, combined with job market and organizational context, shapes workforce impact by influencing whether roles can be retained with retraining, must transform significantly, or will be eliminated entirely, as well as determining how remaining tasks might be redistributed.

Secondly, the Oversight and Workforce Impact axes represent two logically distinct aspects that people often conflate when trying to quantify the “extent” of an automation. When someone claims an automation is “more automated” than another, they might be referring to either its degree of autonomy (Oversight) or its displacement of human roles (Workforce Impact), which are separate considerations. By establishing this framework, we can trace a direct line of reasoning from an automation’s interface design through to these two components of its “extent”: the interface’s abstraction mechanism determines both oversight requirements and, in conjunction with organizational context, workforce changes. This makes tractable what was previously an ambiguous discussion about how “automated” a system is.

Thirdly, the Ownership axis provides a practical framework for analyzing concrete choices organizations face when pursuing automation. Since the necessary extent of oversight is determined by the limitations of the automation technology, and the workforce impact is mostly outside of the organization’s control, their possible choices are determined mostly around the Ownership axis.

Consider a future organization that has the means to automate the work of some junior developers using an “AI Agent” product. Let’s assume that they’re certain that the “supervised” model of oversight is the only way to ensure tasks get done correctly. They must now determine whether they they should transform the role of junior developers, employing them to assist the agent, or scrap the role entirely, delegating the oversight workload to senior developers. If they scrap the junior developer role in favour of a commercial (outsourced) agent product, they run the risk of becoming more dependent on the external vendor than is ideal. If they own and maintain their own “AI Agent”, their risk of regret is lower, but they must also consider the cost of building the product. This leaves them with six possible choices, determined by prepending “In-House”, “Contracted”, or “Outsourced” to “Supervised-Transform” and “Supervised-Remove”. Their choice of automation strategy is now a matter of weighing the tradeoffs between these six scenarios, some of which are easily eliminated as undesirable.

---

A Historical Overview of Automation (1985–2025)

NOTE: Much Of The Following Was Created Using OpenAI’s Deep Research In Combination With Manual Intervention And Strategic Prompting.

This appears to be a fair summary of the modern history of digital automation through the Three-Axis Model.

Between 1985 and 1995, robotics matured in heavy manufacturing. Automotive plants, especially in Japan and the United States, led the adoption of industrial robots, aiming to automate repetitive or hazardous tasks. Many large firms began by developing robotics in‐house, particularly those that sought proprietary control over advanced equipment. Yet, as robotics technology standardized, contracted solutions became common - manufacturers purchased preconfigured robot cells from specialized vendors and operated them on their own shop floors. During this time, fully outsourced approaches (where a third party would own and manage the automation entirely) remained rare, mainly because integrating robotics with daily operations demanded close coordination with the firm’s production workflow. Regarding Oversight, few processes attained a truly “autonomous” mode. Although discrete tasks such as welding or painting were heavily automated, human supervisors stayed on the line to correct malfunctions and approve changes in production. The “human out of the loop” approach, sometimes called “lights‐out manufacturing,” was attempted sporadically but often ran into reliability problems. In terms of Workforce Impact, Japanese companies, guided by cultural traditions that prioritized preserving employment, tended to move workers from manual assembly tasks into robot maintenance or supervisory roles. This exemplified an Adapt dynamic. In contrast, some American manufacturers, under severe competitive pressures, used these new robots to remove a portion of assembly labor. One prominent case was General Motors’ large‐scale automation efforts, which involved substantial up‐front capital investment and workforce reductions, though they did not always yield the targeted productivity gains (as documented by Keller (1989) and others).

During the same 1985–1995 interval, offices began introducing personal computers and database systems to automate clerical tasks. At this time, Ownership was commonly contracted in the sense that organizations purchased commercial software from major vendors. Larger corporations sometimes deployed in‐house solutions to integrate data‐entry or payroll systems with existing processes, but the emergence of packaged software (e.g., from IBM or Microsoft) usually made outright development less appealing. Because software tasks such as data verification or file updates still required active human input, Oversight was predominantly “manual” or at most “supervised,” with nightly batch processes that needed IT professionals on standby. The Workforce Impact in clerical settings often fell under Adapt, since employees remained but learned new workflows - typewriting yielded to word processing, while spreadsheets replaced hand calculations. Although entire clerical departments sometimes shrank, large‐scale removal of white‐collar roles was less common than in high‐volume factory settings. Even in banking, the rise of Automated Teller Machines (ATMs) did not fully remove the teller role. Instead, a partial transformation took place - tellers began focusing on customer relations rather than repetitive cash‐handling (Bessen, 2015).

A second wave of digital transformation arose between 1995 and 2005, driven by the rapid spread of the internet, enterprise software, and global competitive pressures. Large manufacturers and retailers widely implemented Enterprise Resource Planning (ERP) systems that integrated multiple operations - procurement, production scheduling, inventory tracking - into a unified framework. These projects were frequently contracted: well‐known vendors like SAP or Oracle provided the core software, which internal teams then customized. A few industry leaders, including major retailers, built in‐house equivalents that served as proprietary logistical backbones. On the Oversight axis, ERP automated many day‐to‐day transactions (for example, matching purchase orders to invoices) in a mostly “supervised” fashion: once configured, the system executed standard routines independently, leaving humans to supervise data integrity and handle exceptions. Because these solutions replaced entire administrative workflows, Workforce Impact often involved role transformation, if not outright removal. Business Process Reengineering - popular at the time - encouraged reorganizing staff around new digital processes rather than retaining prior structures, which sometimes triggered large‐scale layoffs, especially in the U.S. By contrast, countries with stronger labor institutions, such as Germany, negotiated strategies to keep workers on staff, thereby converting potential removals into transformations or adaptations.

In service industries during this 1995–2005 window, call centers and e‐commerce underwent significant automation. Some corporations built in‐house online platforms for competitive advantage, whereas others contracted prebuilt solutions for customer management systems. Furthermore, the outsourced model emerged when organizations delegated entire call center operations to specialized service providers who managed the infrastructure and labor force. Interactive Voice Response (IVR) systems and basic self‐service websites required limited daily human input for routine tasks, so the Oversight dimension was largely supervised: automated scripts handled the initial call flow, while human agents were alerted for complex issues. The Workforce Impact varied widely. When automation was introduced but retained in‐house, many employees simply adapted, learning new tools for customer interaction. In an outsourced scenario, however, the original in‐house support teams were often removed and replaced by vendor personnel, reflecting a more complete displacement outcome.

By 2005–2015, robotics and software automation had achieved higher sophistication. In manufacturing, sensors, machine vision, and data‐driven controls enabled “smart factories.” Germany’s “Industry 4.0” concept illustrated how equipment, products, and IT systems could integrate seamlessly. The largest firms frequently combined in‐house development of custom software with contracted robotic modules, aiming for synergy between specialized hardware and proprietary data analytics. A few smaller entities started experimenting with “robotics‐as‐a‐service,” effectively an outsourced arrangement, though this was still not dominant. The Oversight axis in smart factories was sometimes described as “human on the loop,” where robots performed the bulk of routine tasks, but engineers or technicians continuously monitored them, intervening whenever anomalies arose. In regulated or safety‐critical plants, manual oversight still prevailed for final approvals or exception handling. Regarding Workforce Impact, co‐determination laws in Germany and cultural emphasis on lifetime employment in Japan often led to an Adapt approach, or at most a structured Transform that preserved internal labor. In industries facing tough cost competition, particularly consumer electronics assembly, more dramatic job removals occurred when advanced machines took over assembly tasks once considered too delicate for standard industrial robots.

Simultaneously, service‐sector automation progressed to include basic forms of machine learning. Data‐driven processes started replacing routine accounting, document review, and customer‐service tasks. While many large companies built initial AI or analytics modules in‐house, they often contracted with specialized software vendors for certain vertical solutions (such as churn‐prediction models). Fully outsourced AI solutions were still relatively new but found favor in use cases like recommendation engines or chatbots, where a third‐party service could plug into the organization’s infrastructure. In all these cases, oversight usually remained “supervised.” Early chatbots or algorithmic screening tools could handle daily operations, but humans were needed to correct errors and retrain models. The workforce response ranged from Adapt - employees learned to interpret AI outputs - through Transform, as entire job roles moved from “manually doing the work” to “overseeing the automation.” In some cost‐driven call centers, enough routine work was offloaded to chatbots or automated phone menus that entire in‐house teams were removed.

From 2015 to 2025, the proliferation of deep learning and broader AI capabilities has brought new decision points for organizations. Many have embraced outsourced AI providers, typically large cloud services that deliver “AI‐as‐a‐service.” Others prefer contracted approaches, purchasing custom AI systems and hosting them internally to maintain closer control. Select corporations in data‐heavy sectors continue to develop advanced AI solutions in‐house, seeking strategic differentiation. As AI accuracy in language processing, image recognition, and anomaly detection improved, the potential for “autonomous” oversight also increased. Yet serious obstacles remain, including legal liability for mistakes, algorithmic bias, and new regulations that mandate human review for high‐risk applications (as seen in proposed EU legislation). Consequently, many businesses favor a “human on the loop” stance even when technology permits greater autonomy.

In terms of Workforce Impact, this latest period has triggered a widespread transformation of knowledge work. Some organizations aim for an Adapt path, providing employees with continuous training so they can supervise or collaborate with AI tools (for example, using AI suggestions to draft documents or interpret large data sets). Others find it more efficient to remove certain roles entirely, especially junior‐level positions that handle repetitive digital tasks, causing concern about the long‐term development of expertise. There have also been strategic debates about whether removing so many entry‐level roles can compromise future talent pipelines. In countries committed to long‐term employment, such as Japan, businesses often focus on adaptation, allowing AI to supplement an aging workforce rather than displace it entirely. Meanwhile, firms in regions with flexible labor markets may treat automation primarily as a mechanism to reduce headcount, especially when facing competitive pressures.

Across the four decades, recurring strategic themes emerge. Ownership decisions often hinge on balancing upfront development costs against a desire for control over proprietary processes: large manufacturers and high‐tech leaders typically invest in in‐house or heavily contracted solutions, whereas smaller organizations often find outsourced models more practical. Oversight remains vital because automated systems can malfunction or produce erroneous outputs, especially in evolving contexts. Even advanced AI rarely achieves a fully “human out of the loop” condition outside highly structured settings (such as warehouse robots performing uniform tasks). Workforce Impact is strongly shaped by the degree to which automated systems target core tasks that define a job. Where primary functions are automated, the role is likely either transformed beyond recognition or removed. Where automation focuses on peripheral tasks, employees adapt and remain integral to the process. Cultural and regulatory environments mediate these outcomes. In Germany, co‐determination practices frequently slow the push toward role elimination, incentivizing upskilling. In the United States and other laissez‐faire labor markets, organizations that see financial value in rapid redeployment can remove entire job categories more easily. In Japan, retaining staff even when new technology arrives aligns with an employment culture in which permanent job reduction is frowned upon, leading to adaptation or gradual transformation rather than abrupt layoffs.

In summary, the evolution of work automation from 1985 to 2025 can be understood by examining how organizations chose to own their automation initiatives, how they balanced the need for human oversight, and which workforce impacts they deemed acceptable or strategically desirable. Initially, the scarcity of off‐the‐shelf solutions encouraged large firms to develop robotics and IT systems in‐house, while smaller firms purchased simpler contracted tools. Over time, a thriving ecosystem of automation vendors enabled more outsourced arrangements, particularly for standardized services and AI capabilities. Despite major technological advances, complete autonomy has remained limited to narrowly defined tasks with predictable parameters, and the most common oversight mode is still “supervised.” Finally, workforce impacts have never been uniform. Workers in jobs that overlap heavily with primary tasks targeted by automation often face displacement, but jobs in settings where firms - by law, culture, or strategy - emphasize upskilling see more adaptation or role transformation. Observing these patterns reveals that automation decisions hinge not simply on the power of new technology, but on broader organizational, cultural, and regulatory frameworks that continue to shape how work evolves in the face of ever‐advancing automation.

References

• Bessen, J. (2015). Toil and Technology. Finance & Development (IMF).
• International Federation of Robotics (various reports).
• McKinsey Global Institute (various analyses on automation trends).
• Keller, M. (1989). Rude Awakening.
• EU Draft AI Act (2021).