Work Is Becoming Software

A position on where labour, capital, and code are converging

The premise

For two centuries, industrial development adhered to a steady pattern: machines handled the work, and humans made the decisions. The loom wove fabric; the weaver assessed it. The excavator dug; the foreman coordinated. When software emerged, it continued this trend by automating routine calculations, leaving humans responsible for reasoning.

The boundary is now dissolving. This decade's key change isn't just software assisting in work; rather, small parts of a job such as judgments, workflows, negotiations, or craftsmanship are now fully automated by software. Humans are involved only at the beginning and end, to set intentions, review results, and ensure accountability.

We refer to this transformation as shift work into software, believing it to be the most significant structural change impacting all businesses dependent on human labour; in other words, every business.

From tools to workers

Enterprise software has traditionally been presented as a tool: something a worker picks up, uses, and then puts away. The CRM wasn't sold directly to the user; instead, a salesperson utilised it. The IDE didn't directly produce code; an engineer did. In this model, value was linked to the human user, and software pricing reflected its role as a productivity booster for their time. This approach made sense when software couldn't operate independently.

That assumption is now the weakest link in a century of enterprise software economics. When a system can set a goal, plan steps, execute across different platforms, assess its output, and improve without human intervention at every stage, it ceases to be valued as just a tool and becomes akin to labour. This is not figurative; it signifies a literal shift in pricing. A seat license based on a fixed number of work hours is a fundamentally different asset than a system whose output depends on compute, uptime, and orchestration quality. The former has a natural limit, while the latter, in theory, can continue to scale infinitely.

This is why the most significant companies being created now aren't just adding AI features to existing software. Instead, they develop systems that quietly take over roles like qualifying leads as an SDR, drafting initial contracts as a paralegal, triaging filings as a claims adjuster, or resolving tickets as support engineers, often before a human even sees the task. The interface no longer resembles a traditional dashboard operated by a person. Instead, it manifests as a Slack message, an email, or an API call, essentially the same interface a human coworker would use. Software is no longer beside the worker; it now sits in the worker's chair.

Three forces converging

We believe this transition is genuine, not just speculative, because three separate curves are now interacting and influencing each other.

1. Reasoning was initially commoditised and then improved significantly. Over just a few years, the cost of a single unit of machine reasoning has decreased dramatically, while its quality has advanced to include multi-step planning, tool integration, and self-correction features that were once considered academic curiosities as recently as 2022. This development is not just about creating a smarter autocomplete; it has produced a versatile reasoning foundation that's affordable enough to operate beneath business processes continuously, without the economics becoming unsustainable.

2. The enterprise now finally has an interface for such integration. APIs, RPA, and years of SaaS fragmentation inadvertently created the necessary nervous system. Every tool a knowledge worker uses the CRM, ticketing system, spreadsheet, and calendar is now accessible via programmatic connection. This connective tissue mostly remained unused for years, only interacting with humans clicking through multiple tabs. It provides the foundation an autonomous agent needs to act effectively, beyond just offering advice.

3. Labour economics no longer absorbs slack. White-collar headcount growth has diverged from industry output growth. Skilled services wage inflation and tight talent markets in support, ops, junior legal, and finance functions have increased the cost of not automating roles. When automation costs fall below the multi-year fully-loaded cost of a hire, the choice shifts from a technology risk to a straightforward financial decision.

None of these three forces alone would be enough. Together, they explain why work is becoming software feels more like an observation than a thesis.

What actually changes

It's important to be exact about what dissolves and what doesn't, as the simplified idea that AI replaces jobs overlooks where true value is generated.

What dissolves: the unit economics of task execution. The marginal cost of creating a first-draft memo, support ticket, invoice, or deposition summary approaches the cost of computing. Roles that, in practice, consist of many such repetitive tasks at scale are most vulnerable not due to skill deficiency, but because the bundle of tasks aligns with what software can now efficiently handle.

What doesn't dissolve and in fact appreciates: judgement under ambiguity, accountability, taste, and the ability to set the right goal in the first place. A system that can execute a thousand variations of a plan still needs someone to decide which plan, to notice when the output is subtly wrong in a way that matters, and to be the party answerable when it fails. This is not a consolation prize for humans. It is where nearly all of the economic value was always concentrated; we simply didn't have to price it separately from execution before, because execution was scarce enough to dominate the invoice.

The companies that win this transition understand that distinction structurally, not rhetorically. They are not building 'AI that does 80% of the job and leaves the boring 20% to humans.' They are building systems that fully own a well-specified unit of work, and they are building the judgment layer the review, the escalation path, the accountability trail as a first-class product surface, not an afterthought bolted on for enterprise trust and compliance teams.

The new shape of a company

If work transforms into software, the organisational chart turns into a source code repository, and the balance sheet begins to appear unusual: the cost of goods sold now includes compute expenses for functions that formerly relied solely on headcount.

This has three downstream effects worth underwriting for.

Gross margins are now sorted by task rather than by industry. A services company that previously had 30-point margins due to its labour-intensive nature can, in theory, shift towards software-like margins, though only for the portions of its work that are truly automatable. This leads to a new type of business: the margin-arbitrage incumbent. Such companies take a low-margin, labour-heavy sector (like bookkeeping, claims processing, initial legal review, or logistics coordination), re-implement the execution layer through re-platforming, while maintaining the trust and judgment provided to customers. These firms appear as software companies on their income statements but operate as professional services in customer relationships. This hybrid combining software margins with service trust is currently one of the market’s most undervalued patterns.

Headcount now becomes a variable cost that can be adjusted quickly, both up and down. In the past, the main challenge during a downturn was that labour was sticky, hiring took quarters to reverse, and layoffs imposed high organisational costs. With a workforce largely composed of software, it can be provisioned like infrastructure, scaled in near real time based on demand. This marks a real improvement in the resilience of a company's cost structure. Additionally, this shift is likely to be disruptive for career paths, compensation, and organisational loyalty. Companies that manage this transition with more empathy than what the balance sheet requires will attract talent that others cannot compete for.

The lowest level in every profession is often redesigned or eliminated. Traditionally, junior roles like associates, analysts, and support representatives have served as the training ground for future senior decision-makers, with repetitive work acting as a form of apprenticeship. If software takes over these tasks, the training pathway that sustains the development of judgment disappears. This second-order effect, which is rarely discussed but structurally crucial, poses a significant risk: by automating the training process, we might inadvertently reduce the supply of the very judgment skills that are expected to remain rare and valuable. Anyone building in this field must address where the next generation of senior judgment will come from, or the market for such judgment will eventually stagnate.

What we look for

When we evaluate a company operating on this thesis, three questions do most of the work.

Does the system own an outcome or simply assist with a task? Assists with indicates a feature. Owning an outcome end-to-end, with clear service levels and accountability, characterises a business. A key indicator is the pricing model: outcome-based or usage-based pricing suggests the builder believes in their core thesis; seat-based pricing on an AI tool typically indicates the opposite.

Is the judgment layer a genuine product, or just an apology? Top companies in this space consider human review, escalation, and audit trails as vital intellectual property, key elements that enable risk-averse customers to trust autonomous systems with critical tasks. Less effective firms see human oversight as a temporary fallback, expecting it to disappear as the model improves, but this misunderstands the original value of the oversight.

Does the company truly understand its core offering? Successful firms recognise they are not just selling software to individual workers. Instead, they are providing labour services to a business—priced, insured, and accountable like traditional labour, but delivered through software. This results in a different type of company with unique cost structures, competitive advantages such as data flywheels and workflow lock-in, and a different customer relationship than those from the SaaS era. Founders still using SaaS-era terms like productivity tool, copilot, or assistant may be underestimating the true nature of their product or have yet to see the full picture of what they are building.

The honest counterweight

We need to clearly identify where this thesis lacks evidence. Autonomous systems fail in ways that differ fundamentally from human failure, often silently, confidently, and sometimes on a large scale before detection. Trusting a software system with regulated, high-stakes roles like medical triage, credit scoring, or legal filings requires institutions to proceed cautiously. This caution isn’t mere friction to be bypassed; it’s a necessary constraint that successful companies must consider when designing their systems. The social contract issue—what happens to those whose tasks are absorbed, and how quickly is not simply a PR challenge. It’s a genuine cost borne by individuals, and companies or investors ignoring this will ultimately be rightly criticised.

We believe the thesis remains valid regardless. This isn't because the transition will be seamless, but because the economic incentive - the disparity between the cost of software execution and human labour for well-defined tasks is too significant and persistent to go unexploited. Markets tend to bridge such gaps. The only uncertainty is who will close it, and whether they'll develop the judgment layer and accountability mechanisms that ensure a trustworthy transition, or simply release the execution layer and leave others to handle potential failures.

The bet

Work has traditionally been a bundle: execution, judgment, and accountability, sold together because they were inseparable; one couldn't buy judgment without also purchasing time. Now, software can unbundle this package for the first time in history. Execution is represented by code, while judgment and accountability, if implemented properly, become more valuable and explicitly rewarded than ever before, as they are no longer concealed within an hourly rate.

That is the company-building opportunity in front of us: not 'automate work,' but rebuild the architecture of work around the fact that execution is now nearly free, and judgment is the only scarce input left.

Everything else is implementation detail.