Margin of Safety #44: From Process Automation to Process Reinvention

Reinvention redesigns the system for machines and not humans

Back to process automation. We’ve previously discussed the difference between process automation and process reinvention (link to post). This week, we’re going to dive in further. To that end, we suspect that the history of physical process automation can give us some previews of what reinvention will look like. Notably, while current enthusiasm for artificial intelligence emphasizes near-infinite flexibility, we think digital systems will eventually encounter the structural and economic shift similar to those encountered by their physical counterparts.

In the industrial sector, the deployment of automation is dictated by a fundamental trade-off between volume and variety. This spectrum is typically anchored by two distinct operational models: the job shop and the flow shop. A job shop is designed for high-variety, low-volume production, relying on general-purpose tools and human expertise to manage unique specifications. In a digital context, this is mirrored by using large language models as assistants that help humans highly navigate bespoke tasks (eg, OpenClaw or many AI desktop tools). Conversely, a flow shop—or assembly line—is optimized for standardized, high-volume repetition (eg, automating SOC tier1 triage or other highly repeatable security tasks). While exceptionally efficient, it is difficult to reconfigure versus the job shop. But that same specialization allows organizations to drive orders-of-magnitude gains in throughput, cost efficiency, reliability, and control—at the expense of flexibility—forcing a strategic choice between adaptability and scale.

(Source: Gemini Nano Banana Pro)

Engineering pragmatism suggests there is no universal ideal for flexibility. Increasing the degrees of freedom in a system, whether through robotic sensors or complex model parameters, necessarily increases the cost of development, power consumption, and likely maintenance. In the case of LLMs, the power consumption is quite literal – more flexible use cases often require larger models, which in turn need beefier hardware and more electricity. This leads the principle of simplicity to prevail; operators favor the most rudimentary automation that can successfully complete a task and all its variants, pushing high volume automation to specialized per-process solutions; for all that Toyota assembly lines are phenomenally expensive, it’s still cheaper to build out specialized lines than more general purpose ones. [1] Job shops, on the other hand, benefit from the lower ratio of fixed:variable cost ratio associated with minimally specialized processes or giant, all purpose AI models.

We think the logistics infrastructure developed by Amazon over the past fifteen years provides a fascinating case study in this vein, while also demonstrating how a process can be automated with just-right levels of flexibility. In 2010, fulfillment centers (FCs) were designed around human physical limitations, featuring wide aisles and static shelving at reachable heights – see this video we dug up (especially around the 1-1:30 minute mark). Amazon slowly added more and more automation though, and each generation of fulfillment center was slightly more automation-friendly than the alst. By 2025, the model has shifted to an automation-forward design, visible in this video. Modern FCs utilize goods-to-person systems where shelving units move to stationary robots or workers, storage density ignores human navigational requirements, bins appear redesigned to minimize errors, and actual fencing protect robots and humans from accidentally interfering with each others’ workflows. It also reframes the human role to one more frequently focused on supervision and repair.

This is what genuine process reinvention looks like. A modern FC is not merely a manual warehouse assisted by machines; it is a system that a purely human workforce would find physically impossible to operate. The arrangement, density, and sheer expanse of goods exceed human cognitive and physical limits (and the density precludes humans armed with forklifts). Digital process reinvention follows a similar path. Initial efforts typically involve automating existing human steps, such as using software to fill out a digital form. True reinvention occurs when a process is redesigned to rely on capabilities that humans do not possess.

We believe you can detect a fully reinvented process by looking for specific indicators.

1. Inhuman scale, where the data volume or transaction frequency handles a load that would be mathematically impossible for a human team to manage. For example, many fraud detection systems can handle so many queries that there aren’t enough humans in the world to successfully replace the automated system with manual investigation. We think modern GRC is an interesting area for this sort of reinvention: much of it happens on a quarterly or annual basis due to constraints in human throughput. With automation-forward reinvention, ongoing compliance could be on the table

2. Process appropriate flexibility. The level of flexibility should be fully sufficient but not excessive, with higher volume (and cost) processes typically having more specialization in the name of efficiency. It should be able to deal with reasonable levels of variability, but your customer support agent probably shouldn’t be able to traffic in bitcoin. On the other hand, your general-purpose desktop agent should be able to do far more things, but likely at far higher per-operation cost than something more specialized.

3. Structural dependency. A reinvented process cannot revert to manual operation; if the automation layer fails, the business logic ceases to function because the workflow is not built to be a human-operable sequence. This is the shelving of the modern amazon warehouse; if the robots fail, there’s no practical way for humans to replace them. Another example might be highly asynchronous security investigations. Humans are relatively bad at context switching, so if a task involves a multitude of irregular waits, a human operator will likely struggle to stay on top of it.

4. Machine-centric data schemas or structures. Today, many key workflows revolve around human-friendly data, with processes going so far as to convert structured data into human language and back again. A hallmark of non-reinvented processes is when AI is used to operate on natural language representations of structured data, rather than skipping several steps and operating more directly on the underlying data itself.

Conversely, if a process remains structurally unchanged and features human-friendly formats, scale, and structure, we question whether it’s experienced a full re-envisioning.

The transition toward an automation-forward enterprise will likely result in an environment that appears unrecognizable to those accustomed to current manual workflows. Just as an operations leader from 2010 would find the specialized density of a 2025 fulfillment center alien and impractical, the future digital enterprise will operate on a logic optimized for machine efficiency rather than human comfort. Success in this phase of development will depend on moving beyond the replication of human effort and towards creative thinking about what actually needs to get done, and what the most effective, maintainable way to do it is.

If you’re building something in this space, feel free to reach out to jpark@forgepointcap.com and kshih@forgepointcap.com.

This blog is also published on Margin of Safety, Jimmy and Kathryn’s Substack, as they research the practical sides of security + AI so you don’t have to.

[1] AI can likely change this on the margin by reducing some classes of development cost. But we expect the core tension will remain; more flexibility means more complexity, and complexity is rarely free-as-in-beer.