Rethinking the Org Chart: How AI Changes the Way Work is Structured

The question firms are actually facing

Walk into any large organisation in 2026 and you will find people running AI pilots. Customer service teams testing chatbots. Legal departments experimenting with contract review. Finance functions exploring automated reporting. In most cases, the question being asked is: "which of our tasks can AI do?"

It is a natural question, and the task-level exposure literature discussed in the first article in this series provides a useful starting point for answering it. But it is the wrong place to stop. The exposure framing treats each task in isolation — as if automating 60% of a job's tasks straightforwardly removes 60% of the need for that job. In practice, the impact depends entirely on how those tasks relate to each other and to the tasks performed by other people in the organisation.

The deeper question is not "which tasks can AI do?" but "given that AI can do some of our tasks, how should we redesign the way work is organised?" That is a question about job design — and it turns out to have a surprisingly precise structure.

Why jobs exist: three forces in tension

Jobs are not natural units. They are bundles of tasks, and the way tasks are bundled is a design choice — one that firms have been making, mostly implicitly, for as long as organisations have existed. Why are certain tasks grouped together in one role and not others? Three forces shape this decision, and they pull in different directions.

Coordination costs favour bundling

When two tasks are interdependent — when performing one well requires knowing about the other — there is a cost to splitting them across different people. That cost shows up in meetings, handover documents, waiting for responses, misunderstandings, and the quality loss that comes from imperfect communication. When one person does both tasks, most of this cost disappears. You do not need to schedule a meeting with yourself.

This force favours broad jobs with many tasks — ideally, all heavily interdependent tasks bundled into the same role.

Specialisation returns favour splitting

A worker who concentrates on fewer tasks performs each one better. This is the logic of the division of labour, as old as Adam Smith. A generalist who writes code, manages projects, and handles client relationships will typically be less proficient at each than three specialists. The more diverse the tasks in a job, the greater the penalty to performance.

This force favours narrow, focused jobs — which directly conflicts with the coordination argument.

Wage structure favours grouping by skill level

When a job includes tasks requiring very different skill levels, the firm must pay the worker at a rate that reflects the most demanding task. A solicitor who spends half her time on high-value legal analysis and half on routine document formatting is being paid solicitor rates for the formatting. A junior administrator could do the formatting at a fraction of the cost — but splitting the work would introduce coordination costs.

This force favours grouping tasks with similar skill requirements, which may cut across the coordination structure entirely. The most tightly interdependent tasks in a process may require very different skill levels.

A job is a compromise between coordination, specialisation and cost. AI changes the terms of the compromise.

Every existing job in an organisation represents a locally optimal compromise between these three forces, constrained by the reality that a worker has a finite number of hours in a day. The question is what happens to this compromise when AI enters the picture.

A concrete example

Consider a legal team reviewing contracts. A solicitor ($200/hour) reviews the contract and a paralegal ($40/hour) summarises the key clauses. The tasks are highly interdependent: the solicitor needs the summary to focus the review, and the paralegal needs to understand the solicitor's priorities to produce a useful summary.

Should these tasks be bundled into one job or split across two people? If the solicitor does both, the firm pays $200/hour for three hours of paralegal-grade work — a $480 premium. If the work is split, the firm avoids the wage premium but incurs coordination costs: briefing meetings, back-and-forth on priorities, delays while the paralegal waits for clarification.

The right answer depends on the relative magnitudes. And AI can change every one of them.

Four channels, not one

Most discussion of AI and work focuses on a single channel: can AI do this task? But AI affects the job design problem through four distinct mechanisms, each operating on different aspects of the organisation. Getting the full picture requires thinking about all four.

1. Task automation

The most discussed channel: AI performs a task entirely, removing it from the human workload. In our legal example, an AI tool might draft the clause summary automatically, eliminating the paralegal's task.

But the impact depends on the role that task played in the broader structure. If the automated task was a hub — connecting many other tasks through coordination requirements — its removal can fragment the way work is organised across an entire team, opening up redesign possibilities that go far beyond the single task. If it was a peripheral task, the impact is minimal.

This is a point the standard exposure indices miss entirely. They treat each task as independent; in practice, the structural position of a task in the coordination network matters as much as whether AI can perform it.

2. Coordination compression

AI tools can reduce the cost of coordinating between tasks performed by different people, even when neither task is automated. Shared dashboards, AI-generated handover summaries, real-time translation between technical and non-technical language — these reduce the meetings, misunderstandings and delays that make it expensive to split interdependent tasks across workers.

When coordination costs fall, the case for broad, bundled jobs weakens. The firm can afford to split tasks across more specialised workers, because the penalty for doing so has decreased. The prediction is counterintuitive: AI that improves coordination between people can lead to more specialisation, not less.

In the legal example, an AI system that automatically keeps the solicitor and paralegal synchronised — surfacing the right clauses, flagging priority changes in real time — reduces the coordination cost of splitting the work. The two-person structure becomes more efficient, not less.

3. Task acceleration

AI makes a task faster without removing it. An AI coding assistant does not eliminate the "write code" task — it reduces it from six hours to two. The worker still does the task, but has four hours freed up.

This pushes in the opposite direction from coordination compression. When workers can do their existing tasks in less time, they have capacity for additional tasks. The firm can consolidate roles — fewer, broader jobs — reducing coordination overhead and the fixed costs of additional headcount. This is how AI reduces employment even when no task is fully automated: each worker absorbs more of the workload, and the firm needs fewer workers overall.

This channel is largely absent from the public debate about AI and jobs, which focuses almost entirely on whether AI can replace tasks rather than whether it can accelerate them.

4. Skill-requirement reduction

AI can allow a less-skilled worker to perform a task that previously required specialist expertise. A junior analyst using a well-designed AI tool can produce work that previously required a senior consultant. The task still exists and still takes time, but it no longer commands a specialist wage.

This changes the wage calculus. If AI reduces the skill requirement for the most demanding tasks in a job, the cost penalty for bundling high- and low-skill tasks together shrinks — the existing job structure becomes cheaper to maintain. Alternatively, the now-cheaper task might be profitably unbundled and assigned to a lower-cost worker, creating a new, narrower role.

In the legal example, if AI tools allow a paralegal to perform some of the solicitor's analytical tasks, the gap between the two roles narrows. The firm might merge them into a single "AI-augmented legal analyst" role at a blended rate — or it might push further in the other direction, creating a three-tier structure with an AI operator handling the routine analytical work at an even lower cost.

The critical insight is that these four channels can push in opposite directions simultaneously. Task acceleration favours consolidation (fewer, broader jobs); coordination compression favours specialisation (more, narrower jobs). A firm that only thinks about channel 1 — which tasks can AI automate? — will miss the reorganisation opportunities from the other three, and may make design choices that are locally rational but globally suboptimal.

Thinking in graphs

To make this precise enough to act on, it helps to represent a firm's production process as a network. Tasks are nodes. Interdependencies between tasks are edges, weighted by the cost of coordinating across a boundary. A job is a cluster of nodes — a partition of the graph.

The optimal partition balances coordination costs (minimise edges cut across jobs), specialisation returns (keep clusters focused), wage efficiency (keep skill levels within each cluster similar), and a hard constraint on how many hours a worker has in a day. This is a well-studied class of problem in computer science — capacitated graph partitioning — and while it is computationally hard in full generality, good approximate solutions are tractable for the scale of real organisations.1

The value of this representation is that it lets you simulate the impact of AI before committing to a reorganisation. Shock the graph: remove an automated task and see how the remaining structure re-optimises. Reduce coordination weights and see whether roles should split. Compress task times and see whether roles should merge. The output is not a prediction — it is a structured way to explore redesign options and understand their trade-offs.

Beyond headcount: software and systems

The same logic applies to the firm's technology stack. A software system is, in this framing, another way to perform tasks — one that carries its own coordination costs (integration, data handoff, API maintenance), its own specialisation structure (each platform does some things well and others poorly), and its own cost structure (licence fees, development costs, maintenance).

AI shifts the calculus in two ways. First, AI capabilities are being integrated into existing platforms, changing what each system can do and potentially making some standalone tools redundant. Second, and less obviously, AI dramatically reduces the cost of custom software development. Tasks that previously required buying a commercial platform — because building the equivalent in-house was too expensive — may now be cheaper to build. The buy-versus-build boundary moves.

A firm that re-evaluates its workforce and its technology stack simultaneously, using the same framework, is likely to find opportunities that are invisible when the two are considered in isolation.

What this means in practice

This framework is not a theoretical exercise. We are developing it into a practical analytical tool that can be applied to a real organisation's task structure. The process involves mapping the firm's tasks and their interdependencies, calibrating the coordination costs and skill requirements from operational data, modelling the impact of AI across all four channels, and exploring the redesigned partition to identify specific roles that should merge, split, or be restructured.

The output is not a recommendation to "automate 30% of your workforce." It is a structured analysis of how your work could be reorganised — which roles become broader, which become more specialised, where new roles emerge, and what the implications are for hiring, training and technology investment.

If this is relevant to your organisation, we would welcome a conversation about how it might apply. Get in touch.

This article is part of a series. The first article sets out the economic framework for AI and the labour market. The second applies it at the regional level. The fourth presents evidence on how AI is already showing up in labour market data.