What Replaces the Org Chart

The new map of the organisation, unpacked layer by layer, with the capability mechanics, failure modes, and the questions to ask about your own company this week.

The org chart is obsolete.

That sentence will feel premature to anyone who spent this morning in a meeting about reporting lines, or last quarter redesigning a functional structure, or last year fighting through a reorganisation. The shape on the whiteboard is still the shape we all work inside. Boxes and lines. Functions and heads. People reporting to people.

It is still what nearly every company in the world uses to describe what it is. A hundred-year-old shape that has outlasted every technology wave since the telephone.

It is also the wrong object.

The hierarchy of people was built for a specific set of conditions. Coordination was expensive. Communication was slow. Institutional memory lived in the heads of the people who carried it. The only unit of operational work available was a human being. Every structural assumption in modern organisational design traces back to those four constraints. None of them holds anymore, and yet the map we use to draw companies still assumes all four.

What this piece is, and what it is not

This is not another reframe. You have read those pieces and if you are here, you have accepted the shift. Intelligence is becoming cheap and abundant. Agents are workers rather than tools. Context compounds. The knowledge layer is the moat. The work is moving from being done by humans to being built by them.

This is the next layer down.

What follows is the structure of the organisation that replaces the org chart: five operational layers, what each one actually contains, what a capability inside it looks like when agents operate it, what the common failure modes are, what connects them, and what questions to ask about your own company this week. It is the map I have been testing inside client work, inside venture builds, and against twenty years of operating experience across digital organisations at serious scale.

By the end of this piece you should be able to draw the shape on a whiteboard, defend it in a boardroom, and know what to do with it on Monday morning.

The written argument is the essay you are reading. The interactive diagnostic is the Company Map at agentnetwork.ai/company-map, which lets you place your own organisation on each layer in five minutes. The blueprint for actually building in this shape is Groundwork, which sits alongside it. Three complementary things at three levels of engagement: the depth, the diagnostic, the build.

The hypothesis, held briefly

Every company, regardless of industry or era, is an expression of five operational layers. The layers do not change. What changes is who, or what, operates each of them.

For the last century, every layer was operated by humans. That was not a design choice. It was a constraint. There was no other option.

The hypothesis this piece rests on is that each of these five layers is now operable by systems designed and governed by humans, rather than staffed by humans. Not as replacement. As substrate. The humans do not leave. They move. They stop operating the layers and start building, managing, and teaching the systems that operate them.

The fundamentals do not change. A company still needs to exist legally. It still needs to know things. It still needs to stay solvent. It still needs to produce something the market wants. It still needs to be led. Those five imperatives are as old as commerce. What changes is the machinery inside each one.

The five layers

Every organisation, if you strip away the org chart and look at what it actually does, has the same five operational layers underneath.

The naming matters. These are not functions. They are not departments. They are not process areas. They are layers, and the distinction is structural. A function sits inside a layer. A department spans parts of several. A process runs across them. The layers are the underlying architecture of the thing. Everything else is implementation.

The five, in order:

Foundation. What the company is.

Knowledge. What the company knows.

Infrastructure. What keeps the company solvent and compliant.

Functions. What the company does.

Leadership. How the company decides and evolves.

The rest of this section takes each one in turn and unpacks it properly: the layer itself, the old operating pattern, the new capability shape, the common failure mode, and the question to take into your own company.

Foundation

What it is. The legal, financial, and structural scaffolding that makes the company a company. Entity formation. Equity structure. Governance. Intellectual property. Positioning. The articles of association, the cap table, the board, the ownership of what gets made, the story the company tells the market about why it exists.

How it was operated. Lawyers drafted everything from scratch. Accountants configured every structure from first principles. Founders wrote positioning documents and pitched them until they stuck. Governance was built by assembling a board and hoping the culture held. Intellectual property was registered manually and defended reactively. The whole layer consumed months of expensive human time at the start of a venture’s life, then got revisited reluctantly every few years when something broke.

What a capability at this layer now looks like. A foundation capability is a package. Templates for every standard entity type in every standard jurisdiction, configured for the specifics of this venture. Equity structures drawn from proven patterns with the tax, vesting, and founder-protection implications already modelled. Governance frameworks installed with standing agendas, decision rights, and board cadences already built in. IP registration and monitoring running as a continuous background process. Positioning refined against live market signal, not set once and forgotten.

Underneath the capability sit the familiar components. Canonical knowledge about the business (what it does, who owns what, how it is structured). Domain knowledge about how foundation work should be done (the playbooks, the jurisdictional patterns, the governance standards). Task agents that file, register, and monitor. Operator agents that hold standing responsibility for governance cadence. Advisor agents that weigh in on the harder decisions (restructures, cross-border implications, IP strategy). Evaluations that check every filing and every document against the standards the company has set for itself.

The failure mode. Founders treat foundation as sunk cost and stop paying attention. The capability drifts. The articles do not match the current reality of the company. The cap table has undocumented side letters. Governance is a calendar invite that never quite lands. IP is registered but not defended. Then something happens, a raise, an acquisition offer, a dispute, and the foundation layer turns out to be weaker than anyone realised.

The question. What would it take to turn your foundation layer from a one-time setup into a living capability that maintains itself and surfaces issues before they become crises?

Knowledge

What it is. Everything the company knows. The market it operates in. The customers it serves. The product it sells. The regulatory context it works within. The decisions it has already made and why. The patterns of how things are done. The lessons learned the hard way. The things a new hire would need to know on their first day, and the things a long-tenured employee could not articulate but would be lost without.

How it was operated. It lived in people’s heads. Some of it got written down, usually in documents nobody updated. Most of it moved through conversation, meetings, Slack threads, email chains, and the quiet transfer that happens when someone joins a team and absorbs how things work. When people left, they took the knowledge with them. When organisations grew, the knowledge fragmented across silos that did not talk to each other. Every organisation I have worked inside at scale, including some very well-run ones, was fundamentally leaking knowledge all the time. It was the cost of doing business.

What a capability at this layer now looks like. The knowledge capability is the substrate everything else depends on. It holds four distinct kinds of content, each with its own governance and access pattern. Canonical knowledge is the stable bedrock: what the company is, who the customers are, what the brand stands for, how the company is structured. Operational knowledge is live state: what is in the pipeline this week, what a customer said on Tuesday, what changed overnight. Domain knowledge is how things get done: the playbooks, the patterns, the reusable intellectual capital that used to live in senior heads. Experiential knowledge is what the company has learned: decisions and their outcomes, edge cases and their resolutions, failures and what they taught.

Each kind gets stored differently. Canonical in structured, versioned files. Operational in live systems that read the state in real time. Domain in playbooks the agents can reference as context. Experiential in a growing archive of decisions, outcomes, and lessons, structured so it can be retrieved and reused.

The agents at this layer are mostly operator agents. They maintain the knowledge continuously. They watch for stale content. They surface relevant context when other capabilities need it. They nudge humans when something important has not been captured. The evaluation signal is whether the right knowledge surfaces at the right moment, which is harder to measure than it sounds and is where most knowledge systems quietly fail.

The failure mode. Companies confuse storage with organisation. They dump documents into a shared drive and call it a knowledge layer. What they actually have is a landfill. Agents cannot find what they need. Context arrives stale. The knowledge that matters most, the experiential layer, never gets captured at all because there is no routine practice for writing down what was learned. The capability looks like it exists but it does not compound. This is the most common mistake I see.

The question. If your best people left tomorrow, what percentage of what they know would survive? And if you had to get an agent productive in your business this week, could you hand it a knowledge base and trust what it read?

Infrastructure

What it is. What keeps the company solvent and compliant. Banking. Accounting. Tax. Contracts. Data protection. Payroll. Legal obligations. Reporting requirements. The scaffolding that determines whether the company legally and financially exists tomorrow morning.

How it was operated. A finance team closed the books manually every month. A lawyer drafted contracts from templates they kept updating. A compliance officer tracked obligations on a spreadsheet and fought for attention at the board table. An operations lead chased payments, reconciled accounts, flagged late invoices, and absorbed the mental load of knowing what needed to happen by when. The layer was always understaffed and always the thing that got deprioritised until something broke.

What a capability at this layer now looks like. The infrastructure capability runs as a continuous background process rather than a periodic human intervention. Books reconcile in real time. Tax obligations surface the moment they become relevant, with the implications already modelled. Contracts generate from templates, get drafted, get negotiated within defined parameters, and surface for human review only when they step outside those parameters. Compliance runs as monitoring rather than panic, with obligations tracked against a live map of what the company is legally required to do in every jurisdiction it operates in.

The knowledge at this layer is heavily regulatory and procedural. The rules of every jurisdiction the company operates in. The specific obligations the company has signed up to. The contracts in force and their terms. The patterns for handling standard situations. The thresholds that trigger human review.

The agents are mostly task agents (reconciling, filing, chasing) and operator agents (the compliance operator holds standing responsibility for the entire regulatory posture). Advisor agents get called in for the hard ones: cross-border tax implications, unusual contracts, data protection edge cases. The evaluation signal is twofold: nothing is missed, and nothing is done wrong. Both matter. One without the other is insufficient.

The failure mode. Companies treat infrastructure as the thing to automate first because it feels the most mechanical. They deploy agents before they have written down what good looks like. The result is fast, confident execution of the wrong thing. An agent that files a return incorrectly at speed is worse than a human who files it correctly slowly. Infrastructure is the layer where evaluations have to exist before the automation does, not after. Companies that skip this sequence end up with a mess that is harder to fix than the original inefficiency.

The question. Which parts of your infrastructure layer are you still treating as periodic human intervention when they could be running as continuous background capability? And do you have the evaluation discipline in place to automate them safely?

Functions

What it is. What the company does. Product. Marketing. Sales. Customer success. Operations. People. Partnerships. The outputs the market sees and pays for. This is the layer most org charts describe, which is why most org charts are both familiar and insufficient.

How it was operated. Each function was a team. The team had a head, some managers, some individual contributors, and a set of tools they used to do the work. Marketing was a marketing team running marketing software. Sales was a sales team running a CRM. Customer success was a customer success team running a support platform. The function was the team plus the tools. The team carried the knowledge. The tools supported the execution. The output was a product of both.

What a capability at this layer now looks like. This is where the shift is most visible and most misunderstood. A function is no longer a team with tools. It is a capability made of four components: agents doing the work, knowledge feeding them context, workflows connecting them, and evaluations measuring whether the work is good.

A marketing capability has task agents handling campaign execution, content generation, and performance reporting. It has operator agents running standing responsibilities: the content operator, the campaign operator, the analytics operator. Each one monitors, escalates, and coordinates within its domain. Advisor agents get pulled in for strategic questions, brand decisions, unusual campaigns. The knowledge feeding them includes canonical brand, market context, customer segments, and domain playbooks for how good marketing gets done in this specific company. The workflows connect the components: campaign brief to draft to approval to deployment to measurement to learning. The evaluations score outputs against defined standards of good, catch regressions, and feed the learning back into the knowledge layer.

The humans in this function are not the doers. They are the shapers. They build the capability, manage its output, and teach it what good looks like. One person in this role replaces the operational capacity of a team many times their size, not because the person is heroic but because the capability underneath them is doing the work.

The same pattern applies to sales, customer success, operations, people, partnerships. Different domain knowledge, different specific agents, same structural shape.

The failure mode. Companies deploy agents on top of the existing functional shape without redesigning the function. The marketing team adds an AI assistant. The sales team adds a prospecting bot. The support team adds a deflection agent. Each one is useful and none of them changes the shape of the function. Productivity improves at the margin. The cost structure does not. Six months in, the company has more tools, roughly the same headcount, and a vague sense that it should have gone further. This is the single biggest waste of the current moment and it is happening at scale right now.

The question. If you were designing your marketing function from scratch today, would you start with a team and add tools? Or would you start with a capability and assign humans to build, manage, and teach it?

Leadership

What it is. How the company decides and evolves. Strategic rhythm. Board cadence. Cross-functional decisioning. Feedback loops. The capacity to read the environment and adjust. Leadership is the layer with the smallest headcount and the highest leverage. Get it right and the other four layers align. Get it wrong and they drift.

How it was operated. An executive team met weekly. A board met quarterly. Strategy got set annually and revisited when something happened. Decisions flowed up through hierarchy, got filtered through politics and agendas, and arrived at the top distorted by the journey. Leaders worked with the information they could get rather than the information they needed. The quality of decisions was bounded by the quality of what reached the decision-maker, which was often a long way from ground truth.

What a capability at this layer now looks like. The leadership capability does not replace humans making decisions. It radically changes the information underneath them. Ground truth becomes achievable. Every signal, every action, every decision: visible, queryable, real.

The capability runs on the knowledge layer but faces the leadership team specifically. It synthesises what is happening across the business continuously. It surfaces anomalies before they become incidents. It runs scenario analysis on decisions before they are made. It tracks every decision that has been made, the reasoning behind it, and the outcome, so the company can learn from its own decision-making over time.

The agents at this layer are heavily advisor. They do not drive the work. They support the humans driving it. They provide analysis, options, framing, and challenge. They ask the questions the humans should be asking. They flag the things the humans are not seeing. Task agents run the operational parts (board packs, meeting prep, decision logs). Operator agents hold responsibility for specific standing inputs (the strategy operator, the performance operator, the culture operator). The evaluation signal is whether decisions are being made on better information than they would have been otherwise. Measurable, if you are serious about measuring it.

The failure mode. Leaders use the capability as a dashboard and nothing more. They read the outputs but do not change how they decide. The meetings look the same. The politics still shape what surfaces. The advisor agents get ignored when their analysis is inconvenient. The capability provides ground truth but the organisation does not yet know how to act on it. This is a maturity problem, not a technology problem, and it is the one most leadership teams are about to encounter.

The question. What percentage of the decisions your executive team made last quarter were made on the information you actually needed, versus the information that happened to reach you? And what would change if the answer was ninety percent instead of thirty?

The connective tissue

The five layers are not silos. They share machinery, and the machinery running through all of them is as important as the layers themselves.

Four pieces of connective tissue run across the whole map.

Context engineering. The discipline of deciding what the agents see, what they remember, how the knowledge gets structured, and how it flows between layers. It is not prompt engineering. It is information architecture for agents, and it is becoming the most valuable skill in the building. The companies furthest along treat it as a dedicated practice.

Evaluation infrastructure. The scoring, regression testing, and feedback loops that tell you whether the capabilities are working. Without evaluation, capabilities drift. Memory that is not measured decays. You are not running a capability. You are hoping. Tools like Braintrust have turned this from artisan craft into real infrastructure. It is the discipline with the highest return in the new stack and the one most companies skip.

Orchestration. How capabilities coordinate across layers. The sales capability needing brand context from marketing. The support capability needing product data from operations. The leadership capability pulling ground truth from every layer at once. Orchestration is what turns five separate capabilities into a coherent organisation. Frameworks like LangGraph and the Claude Agent SDK provide the patterns. The challenge is less technical than architectural: deciding how the layers actually talk to each other.

Governance. The policies, permissions, and safeguards that determine what the capabilities are allowed to do on their own, what requires human review, and what is off-limits entirely. This is the layer most companies are underinvested in and it is the one regulators are about to look at hardest. Governance done well is invisible. Governance done badly is the story that ends up on the front page.

These four run through every layer. A finance capability has context, evaluation, orchestration, and governance. A marketing capability has context, evaluation, orchestration, and governance. If any of the four is missing or weak, the layer built on top of it cannot hold.

This is why the map is three-dimensional rather than flat. Five layers sitting on four connective disciplines. The companies redesigning around this shape are the ones thinking in both dimensions at once.

Why now, compressed

The hypothesis depends on agents being capable enough to operate these layers, and cheap enough to make the economics work. I have unpacked this at length in previous pieces. Here is the state of the stack in one passage, as of this month.

Frontier models sustain multi-step execution now. Claude Opus 4.7, released in April 2026, hits 87.6 percent on SWE-bench Verified. The trajectory is steepening, not flattening. Orchestration frameworks are production-ready: LangGraph at 120,000 stars, OpenClaw at 345,000, Hermes at 95,000 in seven weeks. The Model Context Protocol is the universal standard for how agents plug into tools. The memory layer has matured into a named discipline. Evaluation has turned from an afterthought into infrastructure. Notion’s AI team went from three issue fixes a day to thirty by building evaluation into their development loop. Block runs an operating model that produced a million dollars of gross profit per employee in 2025, projected to hit two million in 2026. McKinsey runs twenty-five thousand agents alongside forty thousand humans. Gartner predicts forty percent of enterprise applications will embed agents by the end of this year, up from five percent twelve months ago.

It is still early. The frameworks are improving weekly but they are not finished. Agents hallucinate. Context gets lost. Integration is harder than the demos suggest. Anyone telling you this is frictionless has not built it. Anyone telling you it is not happening has not been inside it.

Both things are true at once.

What you would expect to see if this is right

A hypothesis worth the name makes predictions. Here are the ones this map generates.

Gross profit per employee should rise sharply at the companies furthest along. It has. Block is the canonical public example. There are many others that have not published the numbers yet.

Hiring patterns should change, particularly at the junior end, as operational execution shifts from humans to systems. They have. Senior roles expanding, junior roles quietly unfilled. Snap announced a sixteen percent headcount reduction in April 2026 and cited AI-generated code explicitly. The pattern is now documented across the UK and US.

New disciplines should emerge that did not exist three years ago, and practitioners should become disproportionately valuable. They have. Context engineering. Agent evaluation. Capability design. Knowledge stewardship. None of these were job titles eighteen months ago. All are scarce roles now.

New cross-layer roles should appear, people whose value is not inside a single function but in their ability to connect the layers. They are appearing. Chief of staff roles being redefined around capability orchestration. Capability leads holding responsibility across multiple functions. Context stewards who do not sit inside any one function but serve the entire knowledge layer. This is happening quietly and the market has not yet priced it.

Cost to start should fall sharply. It has. Five million pounds to seed three years ago, around one million today, materially less for some builds. The curve is still steepening.

Organisations mastering the new shape should outperform organisations that do not. They are, early. The companies running agent infrastructure at scale are posting margin expansion outside the normal distribution.

The transition should be uneven, with most organisations still in the old shape, a growing middle in hybrid states, and a small leading edge already operating in the new shape. This is exactly what the adoption curves look like.

Every prediction the map generates is already observable. The question is whether it is being read by the people responsible for drawing the new structure.

Where this could be wrong

A map that does not mark its fog is not trustworthy. Four things could slow or reshape the picture.

The frameworks could fragment rather than converge, leaving organisations carrying technical debt across incompatible agent stacks. The early signs point to convergence around MCP and a small number of orchestration patterns, but it is not locked in. A fragmented ecosystem would slow deployment and raise the cost of the transition.

The evaluation problem could prove harder than it currently looks, particularly for subjective, high-stakes, or regulated work. Many confident deployment stories involve tasks where good and bad are relatively easy to define. The murkier territories (creative judgement, strategic analysis, difficult interpersonal calls) may take longer to evaluate reliably, and the companies that get there first will have a disproportionate edge.

The regulatory environment could materially reshape what agents are allowed to do, particularly in finance, health, law, and the public sector. The EU AI Act is the most visible piece of this. More is coming, and the shape of it is not yet settled.

The organisational politics could hollow the shape out from inside. This is the risk most senior readers will quietly be worried about. A company can redesign around layered capabilities on paper and still fail to redesign the incentives, the power structures, and the budget flows underneath. If the new map gets drawn but the old incentives survive, the capabilities get built but the organisation does not change. The layers exist, the humans still behave as if functions were teams, the leverage never materialises. This is the failure mode that looks like progress and produces none.

The map still holds even if one or more of these plays out. The shape of the organisation is the shape of the organisation. The variable is how fast the transition runs and how cleanly the old shape gets let go. Every signal in the market right now says faster, not slower.

What this means for the reader

Look at the org chart you are currently using to describe your organisation.

Notice that it draws one layer, poorly, with reporting lines. Notice that four of the five layers that actually run your company are invisible on the page. Notice that the chart assumes every operational layer is staffed by humans, because when someone designed it, that was the only option available.

Now ask the real question. If you were designing your organisation from scratch today, knowing what agent infrastructure can currently do and the rate at which it is improving, would you draw the same picture?

Almost certainly not.

Here is what to do with the map this week.

Draw the five layers on a page. Foundation, knowledge, infrastructure, functions, leadership. Under each, write what your company currently has at that layer and who operates it. Be honest. Most of the cells will say humans and many of them will say humans and some tools they bought.

Now ask, for each layer: what is the first capability I would build if I were going to move this layer from staffed to operated? What knowledge would it need? What agents would do the work? What evaluations would tell me it was working? Who in my current team would build it, manage it, and teach it?

If you want the faster version, I have built an interactive Company Map at agentnetwork.ai/company-map. It takes the same five layers and lets you place your organisation on each one in about five minutes. It is free, it is designed for senior operators, and it will give you a starting point to take into your next leadership conversation.

You do not need to answer all of this right now. You do not need to start building this week. But if you do the exercise honestly, in either form, you will see the shape of what your organisation is about to become. You will see which layers are furthest from it and which are closest. You will see where the leverage is.

The map is not the terrain. But drawing the map is where the redesign starts, and the companies that do the drawing in the next six months will be operating in the new shape before the companies still arguing about reporting lines even notice the ground has moved.

This is the shape.

The build is Groundwork.

---

If this landed, subscribe to Ground Truth for the next pieces as they land. The Company Map at agentnetwork.ai/company-map is the interactive version of everything in this essay: five layers, four connective disciplines, a diagnostic that places your organisation on each one. Groundwork is the blueprint that sits alongside it, with the knowledge architecture, the file templates, and the order of operations for building in this shape. Both are free. Both are written for the people doing the work, not the people writing about it.

---

Craig Hepburn is an AI strategist and Perplexity Fellow. Twenty years building at the frontier of digital, from Microsoft and Nokia to Art Basel and UEFA. Now building at the frontier of agentic intelligence.