Rethinking the Org Chart for the Age of AI

Ask most organisations how AI has changed their structure, and the honest answer is: it hasn't, really. New tools, new habits, the same boxes and lines on the same chart. For a meaningful slice of that group, that's not a problem — it's simply not where the constraint sits. For a growing number of others, running AI through an unchanged structure is exactly where things are starting to strain, and no amount of "use the tool more responsibly" is going to fix a shape that was never built for this volume of judgment work.

This piece is about telling those two situations apart, and — for the organisations where structure genuinely is the constraint — laying out a few concrete shapes worth experimenting with, rather than assuming there's one correct answer everyone should be racing toward.

Where the same chart is fine

A good amount of what AI is doing inside organisations today is narrow, repetitive, and digital: sorting and tagging records, drafting standard letters, first-pass data entry, answering routine queries, pulling together a report that follows the same template every month. None of that work was ever running through a long chain of judgment and sign-off — it was mechanical to begin with, and AI just does the mechanical part faster and more cheaply. There's no structural mismatch to fix, because the structure was never really the bottleneck for that kind of work. The honest, sufficient outcome there is margin: the same output, or a bit more of it, at lower cost — and that's a perfectly good result, not a smaller version of some bigger transformation the organisation somehow failed to attempt.

It's worth being direct about the fact that not everyone agrees with that framing. Some recent research argues the opposite — that AI layered onto unchanged structures only ever produces incremental gains, and that organisations "will see incremental gains" at best unless they treat AI as "a catalyst to redesign work" altogether (BCG, "Making AI Productivity Deliver Real Value," 2026). That's a reasonable position, and probably the right call for a lot of functions. But it's also worth resisting the opposite overcorrection: treating every AI deployment as a failure unless it triggers a redesign. If a workflow was narrow and repetitive before AI arrived, incremental may simply be the honest ceiling — and there's nothing wrong with banking it.

Where the same chart is already the problem

The picture changes once AI's reach extends into synthesis, first-pass analysis, or client-facing judgment — the kind of work that used to be rate-limited by how fast a skilled person could produce a draft. AI collapses that production step, but the review and judgment step behind it doesn't get any faster to do — if anything it gets harder, because someone now has to check not just for typos but whether the reasoning holds up and the output actually fits the situation. Researchers studying AI-assisted teams describe exactly this: the bottleneck relocating downstream to review, testing and integration, with the people left holding it moving faster and, by their own account, less carefully (CIO, 2026). A recent industry paper on the same pattern is blunter still: "faster output does not automatically produce better outcomes," and the now-familiar finding that most generative-AI pilots fail to show a measurable return traces back to exactly this — production got cheaper, judgment didn't (Seramount, 2026).

There's a telling gap in the data on how widespread this is. McKinsey's 2026 survey of more than 10,000 leaders finds 88% of organisations already experimenting with AI, but only 14% with leaders consistently championing it around a clear strategy — and separately estimates that roughly three-quarters of current roles will need to be reshaped as AI embeds into everyday work (McKinsey, "The State of Organizations 2026," via AI to ROI, 2026). Put together, that's a lot of organisations pointing a genuinely different tool at their existing structure and hoping the gap closes itself.

This isn't hypothetical — it's already visible inside plenty of organisations running AI through an unchanged structure. Middle layers sized for a slower flow of work are being buried under review queues they weren't built to clear, and it isn't only a headcount problem. It's also a generational one: the people most fluent with AI tools tend to be the most recent hires, who learned them from day one — but fluency with a tool isn't the same as the judgment needed to check what it produces, and that judgment is normally built over years of doing the underlying work the slower way. Implementation research on this describes the same pattern: employees become fluent with the tools quickly, but often lack the experience to evaluate the outputs, so "speed becomes the proxy for competence, and evaluation becomes optional" (Seramount, 2026). The practical result: the people best placed to use the tool are often the least equipped to catch when it's wrong, and the people equipped to catch it — through years of pre-AI experience — are frequently the ones who've used it least. That's a structural mismatch, not something a training afternoon fixes.

It's also worth a caveat before reaching for a new shape: McKinsey's broader research on this cautions that flattening hierarchies by itself tends to yield diminishing returns — the bigger lever is redesigning the end-to-end process, not just relabelling the boxes on a chart. Shape is necessary, in other words, but it isn't sufficient on its own. Every model below assumes it's paired with an honest look at the workflow running through it.

Three shapes worth experimenting with

None of what follows is a template to adopt wholesale. It's closer to a menu — three structural patterns that different organisations, or different parts of the same organisation, are already experimenting with as they work through what AI actually changes about their work. Treat each the way you'd treat any other pilot: try it with one function or one team, watch what happens to backlog and decision quality, and adjust before scaling it further.

Model A - The calibrated rectangle

This option keeps the basic shape of today's organisation, but resists an assumption that's easy to smuggle in without noticing: that the bottom layer's output is a fixed, ever-growing flood, and the only available lever is scaling the middle layer to keep pace with it. Do that and nothing has really been redesigned — it's the same review backlog with more reviewers standing in front of it, an expensive way to run an organisation in headcount and in the sheer number of drafts and "new ideas" generated for nobody in particular.

The more useful version treats the bottom and middle layers as one system to calibrate together, not two independent dials. It starts at the bottom: the standard for AI-augmented work should be "produce what's genuinely decision-ready," not "produce as many drafts, options and angles as the tool will allow." Get that discipline right, and the middle layer isn't absorbing an uncapped firehose — it's absorbing a deliberately calibrated volume of practical, executable work, sized to what the organisation can actually act on. Only once that's in place does it make sense to ask whether the coordination layer also needs more capacity, and usually less of it than a naive "match the new volume" calculation would suggest. The resulting organisation ends up looking less like a classic pyramid and more like a rectangle — a wide base of AI-augmented producers, and a genuinely thicker, better-resourced middle doing synthesis and exception-handling, sized to calibrated output rather than raw volume.

The catch with this model is supply, not concept. If many organisations reach for the same fix at once, the market for people who can competently review AI output — not just check formatting, but judge whether the reasoning and the context are right — gets competitive fast, and that is a genuinely scarce skill to hire for externally. The more promising source of supply is internal: people who built their judgment doing the work before AI arrived are often well placed to become exactly this kind of reviewer, since knowing what "good" looks like doesn't expire just because the tool changed. The honest catch is that not everyone in that pre-AI cohort will want to make the move, or be equally good at it — reviewing AI output well is its own skill, not simply the old job with a lighter workload. Even so, retraining part of the existing workforce into this role is likely to be more viable, and faster, than trying to hire the same profile from a market where every competitor is reaching for it too.

Model B - The flat pod

This option takes the opposite instinct: instead of adding capacity to keep pace with AI, use AI to remove the need for several of the layers in between. A small, self-reliant unit — sometimes just one or two experienced people — uses AI to handle the iterative heavy lifting: drafts, options, first-pass analysis, produced and refined in tight loops, with the humans providing direction, final judgment and quality, then spending the time that frees up on the client or the decision rather than on managing hand-offs. BCG points to exactly this pattern in AI-native companies already operating this way — small teams generating revenue per person that would once have needed an entire department — and argues more broadly that hierarchies will flatten as AI, overseen by humans, takes on coordination work that used to require a layer of managers (BCG, "How Companies Can Prepare for an AI-First Future," 2025). This model tends to suit organisations, or units within larger ones, that are small or specialised enough to keep everyone within reach of the actual decision — a position many small and mid-sized organisations are already in, arguably giving them a structural head start here that larger enterprises have to work harder to recreate.

The trade-off with this model is upfront cost, and it's easy to underestimate. Before a pod can run this way, someone has to genuinely understand the existing workflow end to end, work out which parts of it are worth automating and which aren't, and then redesign the roles around that new workflow — rather than simply handing the old job an AI tool and calling it done. That is real, front-loaded work, and it doesn't stop once the pod is live: workflows keep shifting as the business, the market and the AI tools themselves move, which means the pod needs ongoing access to people who can keep redesigning the workflow and the automation around it — in effect, in-house AI process engineers. That standing capability is the part organisations most often leave out when they cost this model: it isn't a one-off setup project, it's an ongoing function.

Model C - The portfolio organisation

The most honest answer for many organisations is probably not "pick one." Different functions inside the same company are being changed by AI to very different degrees, so it stands to reason they might end up in different shapes at the same time: a back-office or operations function where AI is mostly doing narrow, repetitive digital work stays close to its current pyramid and simply gets more efficient; a coordination-heavy function — planning, risk, client delivery at scale — moves toward the calibrated rectangle in Model A; and a small advisory, innovation, or specialist unit adopts the flat pod in Model B. Rather than a single redesign imposed top-down, the organisation becomes a deliberate portfolio of shapes, each matched to how much AI has actually changed the work running through it.

This model tends to fit best in large organisations made up of business units that don't interact with each other much to begin with — each unit can find the shape that suits its own work without much need to reconcile it with its neighbours. The harder part sits at the centre. Corporate functions — finance, HR, risk, reporting — now have to operate across three genuinely different structures at once, each with different review paths, different decision rights, and different data flowing through it. For a head office used to a single reporting line and a single set of processes, coordinating that can become a significant management challenge in its own right, and it's worth going in with eyes open about that cost rather than treating the portfolio simply as "the flexible option."

What doesn't change, whichever shape you pick

It's worth naming directly the anxiety that conversations about restructuring around AI tend to circle: that this is really just a staging post on the way to an organisation with no people left in it. All three models above say the opposite. In each one, the work AI has genuinely absorbed is the iterative, first-draft, high-volume kind. The work that stays firmly human is judgment — deciding whether an output is actually right, whether it fits this client's real situation, and who is accountable when it isn't. That isn't a comforting assumption bolted on afterwards; it's the actual design constraint every option here is built around. AI accelerates the tasks; people still have to interpret the result and decide what happens next.

None of this resolves itself by leaving the org chart untouched out of habit, and it doesn't resolve itself by restructuring the whole company because a diagram made a good point on LinkedIn, either. The organisations getting the most out of this are treating structure the way they'd treat any other part of an AI pilot: pick the function where the mismatch is actually showing up, try one of the shapes above with a clear before-and-after measure, and let what you learn — not a template borrowed from someone else's AI journey — decide what happens next.

It remains genuinely to be seen whether any one of these shapes wins out as this plays out across industries — and it may be years before there's enough evidence to say so with confidence. For what it's worth, our own leaning, particularly for small and mid-sized organisations, is toward the flat pod: it offers the most flexibility per person invested, and lets a small team produce far more without carrying the coordination overhead the other models do. That's a judgment call, not a conclusion — the right shape for any specific organisation still depends on the workflow sitting in front of it.

Sources

1. The AI productivity paradox: why your teams are busier, but not faster — CIO, 2026

2. The AI Productivity Paradox — Insight Paper — Seramount, March 2026

3. The State of Organizations 2026 — McKinsey & Company, 2026 (statistics via AI to ROI's summary of the report)

4. Building the AI-powered organization — McKinsey & Company, 2026

5. Making AI Productivity Deliver Real Value — Boston Consulting Group, 2026

6. How Companies Can Prepare for an AI-First Future — Boston Consulting Group, 2025

FYT Consulting — building practical, sustainable data and AI capability. This piece reflects our own analysis and synthesis of current research; where we've gone beyond the cited evidence into judgment or prediction, we've said so directly above.