Introduction

If you’re reading this, you’ve probably had that “wow” moment where within seconds, an LLM assistant completed a task that would have otherwise taken you half an hour. For me, it was when an assistant distilled hours of meeting notes into a concise summary — nearly instantly. As LLM systems continue to improve, it’s clear that the nature of day-to-day knowledge work will change dramatically.

At Theory, we aim to distinguish between sustaining and disruptive innovation. Does a new product or technology support and expand the value of incumbents, or does it change what’s possible so drastically that new entrants displace old?

For knowledge work, we humans are the incumbents. With LLM systems that approximate our work in a fraction of the time and cost, where will humans be supported and where will we be disrupted? Said another way – for businesses of the future, which parts of the organization will remain roughly the same but augmented by AI, and which will be entirely transformed? It depends on if the work is linear or branching.

Linear vs. branching work

LLM-powered automation is different from traditional rules-based systems. LLMs can approximate human reasoning and transform arbitrary data. These capabilities open up new categories of automation that were previously impossible.

But because all they’re doing is sampling from existing data distributions, LLMs are ineffective at open-ended planning. This makes them much more effective at automating linear business workflows as opposed to branching ones.

Linear work is centered around a task queue/backlog; this platform is often the central application used over the course of the day. Some examples are:

• Sales outreach – prospecting and emailing
• Security operations centers (SOC) – alert triage/incident response (see our recent investment in Dropzone AI)
• Accounting – transaction processing and reconciliation
• Supply chain operations – transportation planning, procurement operations

For each task, the sequence of actions required to complete it is generally the same. Outcomes are mostly binary, e.g. open/closed or won/lost. Employees are evaluated primarily on how many tasks they complete successfully and how quickly.

Branching work is composed of activities that vary significantly in both content and sequencing from task to task. Some examples are:

• Complex software engineering
• Data modeling/architecture design
• Strategic planning

Completing each task typically requires planning, external context, and sometimes collaboration. Work may be managed through a queue, but volumes are much lower, and it’s not the primary interface employees look at throughout the day. Outcomes are typically not binary, and employees are evaluated on things like work quality or business outcomes.

Of course, most jobs lie on a spectrum between linear and branching work. A functional group or even an individual often does both. For example:

• SOC and site reliability engineering (SRE) teams are very queue-oriented in terms of operations and evaluation. Individual investigations are often linear and repetitive, but at other times require complex branching analysis to resolve.
• In financial research, some analyses require in-depth, branching work. Others are simple screens applied repeatedly across a portfolio or industry, which are more linear tasks.

LLM systems will disrupt linear jobs

Because linear work is repetitive, you can define logical workflow rails to guide non-deterministic LLM systems. While LLMs can’t reliably plan on their own, they work well when provided with external instructions and validations.

Linear task outcomes are often binary and easily validated. Business value is primarily driven by how many tasks get completed (e.g. number of meetings booked, issues resolved, or transactions reconciled) and how quickly (particularly when tasks are incidents or complaints).

While the amount of work that can be automated depends on the complexity of the task and tolerance for errors, the clear-cut, measurable outcomes of linear work make it practical to tune the system as needed.

These factors make linear work a prime candidate for end-to-end automation. How will this change these jobs? Work will shift from completing tasks yourself to:

• Reviewing LLM outputs, whether in approval workflows (e.g. checking emails before they’re sent) or escalation ones (e.g. investigating a security alert that can’t be resolved autonomously)
• Maintaining LLM systems, including updating workflows, model instructions, and data systems to improve system performance (e.g. updating a sales bot’s prospecting criteria or messaging instructions)
• Doing higher-order work that previously took a backseat to day-to-day operations (e.g. reevaluating sourcing strategy in procurement or proactive security engineering work)

All of the above work is higher-level. Reviewers will need advanced knowledge to resolve issues that were too complex to be automated. System maintainers will similarly need to be experts in their domain (as well as understanding the technology).

Today, most functions doing linear work are pyramid-shaped: the largest group of employees are junior-level individual contributors who complete day-to-day tasks. When LLM systems automate a substantial portion of this work, we expect these organizations to:

• Shrink: Fewer employees will be needed overall because more work will be automated.
• Invert: Positions that do remain will be more managerial or advanced, so organizations will instead look like inverted pyramids or diamonds.

Let’s consider automation in accounting firms. Today, a client team might include two junior accountants, a senior accountant, a manager, and a partner. With a reliable automation platform, they might be able to serve the same client effectively without any fully-dedicated junior accountants. Instead, senior accountants or managers will just review AI-generated actions as if they came from junior staff.

LLM systems will be sustaining for branching jobs (for now)

For jobs on the branching end of the spectrum, organizations will change less dramatically.

The wide variance in work means today’s systems can’t plan and execute tasks reliably. Outcomes are often judged on more abstract quality metrics, so it’s harder to evaluate and tune them.

Instead, LLM systems will provide higher leverage tools, abstractions, or copilots that are operated by human professionals. An engineer might write a natural-language header for a function instead of writing the entire code block. An investment analyst might task an AI agent to pull a particular data point from 100 earning reports.

While linear functions will see roles become almost fully automated, branching functions will see more diffused benefits, where each employee becomes incrementally more productive. Certain activities might go away – e.g. note-taking, summarizing, reporting – but this will not radically change the nature of the job.

Some organizations will use this increased productivity to downsize teams or slow hiring. Others will use it to accomplish more, faster, without changing headcount. Entry-level roles may contract slightly, but overall we don’t expect branching work organizations to change substantially in the near to medium term.

Note: There is a lot of ongoing research into improving LLM system planning, generally by combining LLMs with other models explicitly optimized for planning, or by generating enough planning examples that an LLM approximation works effectively in practice. These are promising but nascent. Given the ambiguous outcomes of branching work and difficulty programmatically evaluating it, we think humans will still lead for the foreseeable future.

Note: Certain types of branching work could see a >10x productivity increase versus an incremental one. One example is creative spaces like digital art, 3D modeling, or copywriting. The magnitude of this impact means there will be more disruptive changes to org structure in those industries. We’ll write more about these soon.

How should startups and enterprises think about disrupted vs sustained knowledge work?

Most B2B software of past decades was designed to augment or streamline existing human workflows. When LLM-powered software might transform them entirely, it’s critical that startups have a hypothesis as to how.

For your use case, do LLM systems disrupt or sustain the human status quo? Should you aim to fully automate a portion of tasks, or partially automate all of them? Should humans be leading, reviewing, or out of the loop? What are the incremental steps to build trust along the way?

They also should anticipate organizational impact and incentives: If your product is successful, how does that change how your customers’ organizations are structured? Should you be designing for front-line workers or managers? Does a top-down or bottoms-up sales motion make sense given individuals’ incentives?

Enterprises will approach organizational change gradually but should anticipate the most substantial impacts and challenges. Will you need to change your operating model or geographic footprint (e.g., what happens to your call centers)? How will you hire and train managers if there are fewer junior roles for them to start with? Is an AI-enabled organization in your industry so different from the status quo that a new entrant will outcompete existing players? Does automation mean that companies will bring in-house (or outsource) functions that were previously too costly or complex to manage?

We are only on the first couple pages of this story but know it will be a big one. If you’re building AI systems to transform traditional business functions, we’d love to hear from you at info@theory.ventures.