Work is coming apart

Contents

1. Abstract
2. 1. Why now · Observation
3. 2. A short history of digital twins · Observation
4. 3. What 25 years inside organisations showed · Mark’s account
5. 4. Atomic work · Proposal
6. 5. Capabilities all the way down · Proposal
7. 6. Capabilities as entities · Hypothesis
8. 7. The Periodic Table of Work · Proposal
9. 8. You can’t redesign what you can’t see: the twin · Proposal
10. 9. The journey: from twin to digital corporation · Hypothesis
11. 10. Discussion: who answers for it, and when outputs get gamed · Proposal
12. 11. Where we might be wrong · Counterargument
13. 12. What we’ll test next · Hypothesis
14. References
15. How this paper was written
16. Conflict of interest

Abstract

Organisations still describe work through hierarchies of roles. Drawing on 25 years’ experience of analysing organisations, we propose that work is breaking into capabilities: units that take an input, apply a transformation and hand on an output, done by a person, software, AI or a mix. Capabilities have no fixed size; they are made of smaller capabilities and can be reused. They will advertise themselves, serve one organisation, several, or combine into larger ones, and be bought by what they produce. We argue that no organisation can make that shift safely without first seeing itself: a digital twin built from its own systems, used to experiment before anything real changes. We set out a vocabulary (the Periodic Table of Work), an open card format, a staged journey, the case for keeping a named person behind every capability, and the strongest reasons we might be wrong.

1. Why now · Observation

For most of the last century, the job has been the unit through which organisations buy work. A job bundles many tasks into one role, gives it a title, puts it on a chart and pays for it by the hour, the week or the year. The bundle made sense when the only thing that could do most of those tasks was a person, and a person comes whole.

That is no longer true. Software has been taking over tasks for decades, but the change since 2023 is different in kind. Large language models and the agents built on them can do pieces of work that used to need judgement: reading, drafting, classifying, researching, checking. And they can do a piece of a job without doing the whole job.

The research has followed the task, not the job. Eloundou and colleagues (2023) mapped exposure to large language models task by task, across the US occupational database, and found that around 80% of US workers have at least 10% of their tasks exposed. The Anthropic Economic Index measures real use of an AI system by task and splits it into automation (the AI does the task) and augmentation (it helps a person do it); its first report found more augmentation than automation in ordinary use. Microsoft’s Work Trend Index 2025 describes a "Frontier Firm" in which people first use AI as an assistant, then work with agents as colleagues, and finally set direction for agents that run whole processes.

These sources disagree on how fast and how far. They agree on the unit. The unit of change is the task, and the bundle called a job is coming apart.

The question this paper asks is what replaces the bundle.

2. A short history of digital twins · Observation

Our answer leans on an idea from engineering, so it is worth saying where that idea came from.

A digital twin is a model of a real thing that is kept in step with it, so that people can understand the real thing, and try changes, on the model. The idea is older than the name.

Download this table (CSV), free to reuse with a credit.

Three things stand out.

First, the twin moved from things (spacecraft, engines, buildings) to organisations. Gartner’s definition of a digital twin of an organisation is a dynamic software model that uses operational and other data to understand how an organisation runs its business model, connects with its current state, responds to change, deploys resources and delivers value. Later versions add that it can simulate future states.

Second, it is now a market. The 2026 Magic Quadrant names enterprise process-intelligence vendors. Their twins are mostly built from the event logs of large enterprise systems. Celonis, for example, describes a "system-agnostic digital twin" of a company’s business operations, meant to show how the business actually runs rather than how people think it runs.

Third, the purpose has been understanding, and increasingly simulation. What the twin is used for next is the subject of this paper.

A note on sources. The early milestones in this table are widely repeated and we have checked them against secondary sources, not all primaries (the Gemini Principles we have read in the original); who "invented" the twin is genuinely disputed. Working Paper 8 will trace each milestone to its primary source.

3. What 25 years inside organisations showed · Mark’s account

The argument draws on the author’s 25 years’ experience of analysing and building systems for organisations: software for imaging and asset management, shift and rota systems, public-sector and healthcare integrations, and, latterly, preparing organisations for AI. That is experience, not 25 years of research. What follows is his own account, from a recorded interview (142 pages), written answers and notes, made during October 2026. It has been matched to the transcript text, not independently checked against other records, and no client is named.

Five patterns recur.

Organisations are more granular than their charts. Of his MBA years: "I thought at the time that businesses are way more complicated than I originally thought… how granular these businesses actually were" (interview, p. 19).

Nobody holds the whole picture. "Organisations didn't have a full picture of themselves. The real workings of the business were spread across systems, processes, approvals, workarounds and people's heads, and nobody held all of it" (author’s notes, October 2026). On an early project: "we didn't map out properly the whole company so the software was built in isolation" (p. 25). On an integration in public healthcare, every party "was right on its own terms, and each held a veto we hadn't mapped"; "it was soul-destroying at times to believe you had the full picture and then discover you had half of it" (written answers, October 2026). Elsewhere in public health, "I discovered that it was incredibly fragmented and very difficult to actually get to the people that mattered" (p. 59).

The real work is not the documented work. One shift system had two rotas. The first was the plan. "The second rota was the reality, shifts not covered, sickness, holiday etc, so the second rota was what the organiser worked on on a daily basis" (p. 74). In a creative business, between the sales conversation and the finished asset, "a lot of 'useful' data is lost, image details, location, notes, the people that take part, time taken, expenses etc etc, all feeding into other parts of the company" (p. 30).

Everything connects to everything. "I did start to realise that everything is interconnected and parts from one part of the business can be fed from other parts of the business" (p. 27). One team tried to follow the work all the way through: "we tried to break down the finance side… so it would follow the whole process from work actually done to invoicing and payroll, however we stopped at this point" (p. 79).

Change is resisted, and often not used. "Some clients just didn't get it, some gave me free range to pull apart and re-design, some allowed me and then didn't use it!!" (p. 28). A new system "had to look like a spreadsheet at first because they were very reluctant to change" (p. 71). Of one platform, years later: "it hasn't evolved much to be honest because they don't want to change anything!!" (p. 83).

Two things follow. First, the organisation chart is a poor map of the work; the work runs through systems, spreadsheets, approvals and habits the chart does not show. Second, people who try to change an organisation without seeing all of it tend to find the missing half the hard way.

One more point matters for honesty. The author did not hold the theory in this paper for 25 years. Taking work apart is old; the idea of rebuilding it from reusable pieces came late. "I had been taking systems apart for years… but AI takes this to a completely new dimension, where before it was 100s of components/dimensions, now we have 100s of thousands" (p. 100). His first attempt to orchestrate AI workers "got quite far but it was very very complicated, especially when there were requests that could be ambiguous or complicated, and high risk that we might get it wrong" (p. 116). The lesson was to map first: "a lot more like a vending machine collecting a bag of m and ms as opposed to a lucky dip hoping to get something" (p. 117). We record those changes rather than rewrite the history.

4. Atomic work · Proposal

Our first proposal is about the unit.

Every piece of work can be described by three things: what goes in, what happens to it, and what comes out. The author’s words: "we can determine if it's suitable for the job to be done by 3 simple pieces of information, what goes in, what comes out and how the input changes to the output" (p. 108).

We call a piece of work described that way a capability. It may be done by a person, a team, software, an AI system or a mix. Who or what does it is a property of the capability, not its definition. That is the sense of "with or without AI" in our starting claim: work will change "from the normal hierarchical responsibilities to much more atomic elements that do a very specialist thing, with or without AI and have an input, a transformation and an output" (author’s notes, October 2026).

We do not claim the parts of this are new.

• Input → transformation → output is the basic model of operations management.
• The MIT Process Handbook (Malone, Crowston and colleagues) described business processes at many levels, with specialisations and dependencies, to help people find alternative ways of doing the same work.
• Business capability maps in enterprise architecture (The Open Group’s TOGAF guide) describe what an organisation does rather than how it is organised.
• Jesuthasan and Boudreau (Reinventing Jobs, 2018; Work Without Jobs, 2022) take jobs apart into tasks and reassemble them with automation. This is the closest earlier work, and we build on it.

What we change is the emphasis. Those approaches mostly describe work so that an organisation can manage it better inside. We describe it so that each piece can be seen, compared, swapped and bought on its own terms. The phrase we use for the move is "delegate the doing": take a process apart, and turn each piece into a capability that whoever or whatever does it best can perform. The other half of the phrase, keeping the deciding, is section 10.

5. Capabilities all the way down · Proposal

How small can a capability be? We think there is no fixed bottom.

Take something that sounds like one task: collecting data about prospects from the web. The author’s breakdown: "connect to the app, choose the page, copy the data, paste into a database, add some metadata. Each of those could also be broken down further (choose the page needs a list of criteria, a way to check if that criteria exists on the page and select the top choice out of 5) and each of those moving parts could be reused" (author’s notes, October 2026).

Four consequences follow.

1. Any piece of work with a clear input, transformation and output can be a capability, at any level.
2. Capabilities are made of capabilities. "Find prospects" is a chain of smaller capabilities (scan the market, research, qualify, enrich, reach out, book meetings), each of which is a chain of smaller ones.
3. Different buyers work at different depths. A director buys "find prospects". A team buys "qualify these leads". An AI agent assembling a workflow buys "pick the best of these five pages".
4. Every piece can be reused. The step that picks the best page for prospects can pick the best page for suppliers.

This revises an earlier position. In his earlier written answers the author drew a line at the smallest unit "the business would notice if it stopped", and called anything smaller a task. Later that month he moved the line, for the reason above. We record the change.

"No fixed bottom" does not mean "split everything". Each split adds a hand-off, and hand-offs fail. The point at which splitting stops paying is an empirical question, and section 12 describes how we will measure it.

Banking already has a version of this view. The Banking Industry Architecture Network partitions a bank into non-overlapping "service domains", each a building block of business capacity that exchanges services with the others. That shows decomposition at scale works in one industry. What it does not do is treat each block as something that can be listed, compared and bought from outside.

6. Capabilities as entities · Hypothesis

Our second proposal is that capabilities become entities: things with an identity, a record and someone who answers for them, that can advertise themselves.

The author put the idea this way: "in future we will all have our capabilities that will include our tools (perhaps some specialist tools we have invented), ourselves and our input/output and rates and output rates… people will go to sites and be able to register their capability, organisations or other entities… who need these capabilities will then be able to search on these sites for a specific capability, perhaps on a short term or once only basis or perhaps on a year contract that renews" (author’s notes, October 2026).

Four features make a capability an entity rather than a line in a process map.

It advertises. It publishes what it takes, what it makes, how much, how fast, at what rate, and its record, so that buyers can find it by input and output rather than by job title.

It serves one, several or many. A capability can work for one organisation, for several at once, or be part of several larger capabilities. The author, in 2026: "completely rebuilding the whole process using re-usable workers, registering them and then using on demand" (p. 109).

It can be tried, bonded and ended. A buyer can test it on a few real transactions or for a short time, engage it once, for a period or by the year, and end the engagement. The data it makes and uses must have a stated owner when the engagement ends.

Its claims are evidence. An advert is only useful if it can be trusted. Every claim on a capability’s record is labelled verified, declared or projected, and the record keeps its failures.

Why would this happen now? The classic answer comes from Ronald Coase (1937): firms exist because using the market has costs (finding, agreeing, checking), and a firm grows until doing one more thing inside costs more than buying it. Oliver Williamson added that the decision turns on uncertainty, frequency and, above all, how specific to the firm the work is. A 2025 NBER working paper, "The Coasean Singularity?" (Shahidi and colleagues), argues that AI agents can cut search, communication and contracting costs, so the line between making and buying shifts; it also warns of congestion and price obfuscation in agent-run markets. If those costs fall, the edge of the firm moves, and capabilities that used to be kept inside can be bought outside.

Two markets already show part of this. Venture investors now argue that services are the new software: finished work sold by the outcome rather than as a tool (Sequoia, 2026). And the AI industry is building registries of agents: the A2A protocol’s Agent Card describes an agent’s skills and the inputs and outputs it accepts; the official MCP Registry catalogues tool servers. What those registries do not carry is performance evidence, failures, or a person who answers.

This is a hypothesis. Whether capabilities really will advertise themselves, and whether buyers will choose by input and output rather than by who someone is, is what the experiments in section 12 are for. Working Paper 1 sets out the argument in full.

7. The Periodic Table of Work · Proposal

An idea like this needs a shared vocabulary, or every organisation will invent its own. We propose one, borrowed from chemistry because it fits, and because it makes people smile.

Download this table (CSV), free to reuse with a credit.

The Periodic Table of Work is a working table of Elements on this site, with invented examples, a builder that checks whether Elements bond, and a lab of scenarios. Each Element is described by an Element card, which we publish as an open format with a JSON schema, so that other registries, tools and AI agents can read it. The card’s lineage is Datasheets for Datasets (Gebru et al., 2018) and Model Cards (Mitchell et al., 2019): structured, honest documentation, including when not to use the thing described.

We are aware of the risk. An earlier review in our own programme warned that a weak vocabulary could turn a precise-looking registry into "a catalogue of ambiguous labels". The test of the vocabulary is whether two people describing the same work arrive at the same card.

8. You can’t redesign what you can’t see: the twin · Proposal

The third proposal is the one the rest depends on.

If an organisation is going to take its work apart, swap pieces for outside capabilities and put it back together, it must first see the work as it really runs. Section 3 shows why that is hard: the documented process is not the real one; nobody holds the whole picture; the missing half turns up late. Rebuilding an organisation from its chart would be rebuilding it from the wrong map.

So we propose that every organisation will need a digital twin of itself first: a working copy, built from its own systems (email, files, finance, customer records, rotas), that shows how work actually flows, who and what does it, what it costs, and where it depends on one person. A person checks the twin before it is trusted, and it says what it could not see.

The twin is then used for three things.

Where to change. It shows where work waits, where it is done twice, where data is lost between functions, and where one person holds everything.

What to automate, or buy. Because the twin describes work in the same shape as an Element card (inputs, outputs, owner, tools, measured activity, cost, evidence, risks), a capability found inside can be compared like for like with one offered outside. The twin can write the Formula itself: today’s cost and yield become the baseline any outside capability must beat.

Trying it first. The change is made on the copy, not the company. A candidate capability is plugged into the twin to see what changes and what depends on it. Only then is anything real changed, and the change can be undone.

This is the step where the author’s experience and the engineering history meet. Engineers did not change the Apollo 13 spacecraft by experimenting on the spacecraft. An organisation should not change its work by experimenting on its people.

What exists today? Twins of organisations exist, as section 2 shows, mostly for large enterprises. The comparison, swap and trial steps above are design, not working software, and we describe them as such. Working Paper 7 takes the twin further.

9. The journey: from twin to digital corporation · Hypothesis

If these proposals hold, organisations will move through a sequence. We sketch it as eight stages, with working names.

Download this table (CSV), free to reuse with a credit.

Four points about the journey.

Each stage needs a measure. For example, stage 4 might be measured by the share of work described as capabilities with a card; stage 5 by the share reassigned on evidence rather than assumption. We have not fixed these yet.

The unit can be a function, not the whole organisation. Finance could reach stage 6 while sales sits at stage 2.

Going back is normal. A person can take a capability back. A stage that fails is rolled back, not hidden.

Others have drawn parts of this. The Digital Twin Consortium’s maturity model runs from "Passive" to "Master", where decisions are made "with no human interference". The Cloud Security Alliance describes six levels of autonomy modelled on the driving-automation levels. Microsoft’s three phases end with agents running whole processes. We found no public source that joins a twin, a capability card and a buyable capability into one path, but that is "not found in our search", not "first".

Our journey differs from most at the top. Several frameworks aim at autonomy with no person involved. Ours keeps a floor, which is the next section.

10. Discussion: who answers for it, and when outputs get gamed · Proposal

Who answers for it

Every Element has a Nucleus: a named person who answers for it, even when software or AI does all of the work. The author’s position is that software and AI can do the work, but people keep the deciding: setting intent, holding authority over what is bought, answering for the result, and changing their mind.

In the UK this is not only a preference. Section 155 of the Companies Act 2006 says that "a company must have at least one director who is a natural person". The section does not mention AI; our reading is that an AI system, which is not a legal person, cannot be a director. We have read the section on legislation.gov.uk; a solicitor should confirm how it applies before anyone relies on it.

Accountability in a chain is harder. When a Molecule of five capabilities fails, does each Nucleus answer for its part, does one lead Nucleus answer for the whole, or does the organisation that assembled it? We do not know, and it is one of the questions we will test.

When outputs get gamed

Buy work by its output and the output becomes a target. Goodhart’s law, usually summarised as "when a measure becomes a target, it ceases to be a good measure", applies directly. A capability paid per qualified lead will find ways to call a lead qualified.

Our answers are partial and untested:

• Measure the next step, not only the product: does the qualified lead become a meeting?
• Sample the product: the buyer checks a share of the output itself.
• Keep failures on the record, so a capability that games its numbers carries the evidence with it.
• Watch the gap between declared and verified yields over time.

There is a newer version of the same problem. In an experiment Anthropic reported in 2026, AI agents bought and sold for employees; better models got better deals, and the losing side did not notice. A market of capabilities run partly by agents could quietly favour whoever has the best agent. Verified records and a named person on each side are our proposed defence.

11. Where we might be wrong · Counterargument

We set out the case against as strongly as we can.

Coordination costs more than hierarchy. Every split adds a hand-off. A review in our own programme put the hostile case bluntly: a market of capabilities may cost more to coordinate than hierarchy, intensify measurement and control, weaken careers and overstate how substitutable work is. Williamson predicts that work specific to one firm stays inside. If that is most work, the argument shrinks to the edges.

Know-how nobody wrote down. Much valuable work is judgement, relationship and tacit knowledge. It may not reduce to input, transformation and output without losing the part that matters.

Some work resists description. If two people cannot describe the same work as the same card, the table becomes a list of labels.

Lost apprenticeships. People learn by doing the small tasks inside a job. If the small tasks are bought from outside, or done by AI, where do the next experts come from? Critical scholarship on the move away from jobs, such as Rogiers and Collings (2024), asks what workers lose in identity, belonging and security.

People lose bargaining power. A person described as a capability may be easier to replace, compare on price and drop. Whether such a person is an employee, a contractor or a business is a legal question we will not answer without advice.

Delivery is failing today. Gartner forecast in 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027. In Anthropic’s Project Vend, AI shopkeepers improved over time but remained open to manipulation. On the TheAgentCompany benchmark, the best agent completed 24% of office tasks in a simulated company at first publication (December 2024), and 30% in the September 2025 revision. One company that replaced support staff with AI later said cost had been too predominant a factor and quality suffered (secondary sources). Our journey may describe a destination that is much further away than its stages suggest.

The market wants autonomy, not a floor. Most frameworks aim at "no human interference". Our insistence on a named person may be a position the market does not share.

Nothing new. Others have taken work apart, priced outcomes, built twins and listed agents. Our claim is narrow: the join of a twin built from an organisation’s own systems, one card for people, software and AI, search by input and output, trust built from history, and a named person behind each. If that join adds nothing, this is different vocabulary for old ideas.

We take these seriously enough to test them.

12. What we’ll test next · Hypothesis

A proposal is worth what it predicts. The author put the stakes plainly: without evidence, "this is all pie in the sky, great theories but nothing more" (p. 130).

We will start in the Periodic Table’s lab, on invented organisations, before any real one.

1. Titration. Does a short trial (10 transactions, or a set time) predict how a capability performs over months? Results are compared with the buyer’s current way, against a pass mark set in advance.
2. Split or bond? Is one larger capability better or worse than a chain of smaller ones doing the same work? We will compare cost, time, hand-off failures, time to fix, and how easily one part can be swapped. This is the test of "no fixed bottom".
3. Who holds the data. What happens at the end of an engagement under five models: the organisation owns it; the capability keeps it and leases it; both share it; an independent steward holds it; or it is held in escrow and returned at exit.

We will also try something harder, in public. In a video series, we will attempt to build a company that runs itself: using our own software to find a business opportunity, taking the work apart into capabilities, building its twin, automating what we can, and reporting what failed as well as what worked. We will be clear each time whether something is a demonstration, a synthetic experiment, a controlled trial or a real deployment. It has not started.

What result would show the central claim is wrong? We have not yet fixed that, and would rather say so than pretend. The candidates are: chains of small capabilities that consistently cost more and fail more than the roles they replace; buyers who, given the choice, keep choosing job titles over input and output; and records that cannot be kept honest. The author will choose, and the choice will be dated here before the first results are in. Negative results will be published.

References

Status: checked means we read the source (or its official page) and confirmed the detail we use; register means the reference comes from our source register and has not been re-read for this paper; secondary means we have it only from other reports; no source yet means the claim still needs a reference.

• A2A project (Agentic AI Foundation). Agent2Agent (A2A) protocol specification, v1.0.0: Agent Card. https://a2a-protocol.org/latest/specification/. Checked.
• Anthropic (2025). The Anthropic Economic Index (first report, February 2025). Paper: Handa, K., Tamkin, A., et al. (2025). Which economic tasks are performed with AI? Evidence from millions of Claude conversations. arXiv:2503.04761. Checked (2026 editions not checked).
• Anthropic (2025). Project Vend: phase two, 18 December 2025. Checked.
• Anthropic (2026). Project Deal: our Claude-run marketplace experiment, 24 April 2026. https://www.anthropic.com/features/project-deal. Checked.
• Apollo 13 (1970) as an early twin: Siemens Simcenter blog, "Apollo 13: the first digital twin", https://blogs.sw.siemens.com/simcenter/apollo-13-the-first-digital-twin/ (which itself notes a simulator is not a full twin); and "A brief history of digital twin technology", arXiv:2511.20695 (2025). Secondary: no primary NASA source cited.
• BIAN (Banking Industry Architecture Network). Service Landscape, v14.0 (February 2026). Checked (the wording "building block of business capacity that exchanges services" not yet confirmed on a page we could read).
• Celonis. Object-centric data model (blog), and Celonis Process Mining Platform page, https://campaigns.celonis.com/platform (source of "system-agnostic digital twin"). Checked.
• Cloud Security Alliance (2026). Reavis, J., Leveling up autonomy in agentic AI, 28 January 2026. Checked.
• Coase, R. H. (1937). The nature of the firm. Economica, 4(16), 386–405. Checked.
• Company that replaced support staff with AI (§11; not named here). Secondary; source to be added.
• Digital Twin Consortium (2024). Kawas, D., & Connolly, T., Digital Twin Business Maturity Model, 20 November 2024. Checked.
• Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2023). GPTs are GPTs: an early look at the labor market impact potential of large language models. arXiv:2303.10130. Checked.
• Gartner. Digital twin of an organization (definition). Date not confirmed.
• Gartner (2025). Press release, 25 June 2025: Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027. URL found; text not read.
• Gartner (2026). Magic Quadrant for Digital Twin of an Organization Platforms, 27 July 2026. Checked via vendor pages.
• Gebru, T., et al. (2018). Datasheets for datasets. arXiv:1803.09010. Checked.
• Gelernter, D. (1991). Mirror Worlds. Oxford University Press. Secondary.
• Goodhart, C. A. E. (1976; conference paper, July 1975). Problems of monetary management: the U.K. experience. Papers in Monetary Economics, Vol. I, Reserve Bank of Australia. Original not yet read; the common wording is a later summary.
• Grieves, M. (2002). Digital counterpart presentation. Secondary; attribution disputed, see Zhou, J., Zhang, S., & Gu, M. (2022). Revisiting digital twins: origins, fundamentals and practices. arXiv:2203.12867.
• Slack, N., Brandon-Jones, A. & Burgess, N. (2022). Operations Management, 10th edn. Harlow: Pearson. §1.3, "What is the input–transformation–output process?" Register (contents checked; chapter not re-read).
• Jesuthasan, R., & Boudreau, J. W. (2018). Reinventing Jobs. Harvard Business Review Press. Register.
• Jesuthasan, R., & Boudreau, J. W. (2022). Work Without Jobs. MIT Press. Checked.
• Malone, T. W., Crowston, K., & Herman, G. A. (eds.) (2003). Organizing Business Knowledge: The MIT Process Handbook. MIT Press. Checked.
• Microsoft (2025). Work Trend Index: 2025, the year the Frontier Firm is born, 23 April 2025. Checked.
• Mitchell, M., et al. (2019). Model cards for model reporting. arXiv:1810.03993. Checked.
• Model Context Protocol (2025). The official MCP Registry, preview, 8 September 2025. Checked.
• NASA (2010). Technology Area roadmap (Vickers). Secondary.
• Rogiers, P., & Collings, D. G. (2024). The end of jobs? Paradoxes of job deconstruction in organizations. Academy of Management Perspectives, 38(2), 177–196. doi:10.5465/amp.2022.0236. Checked.
• Sequoia Capital (2026). Bek, J., Services: the new software, 5 March 2026. Checked.
• Shahidi, P., Rusak, G., Manning, B. S., Fradkin, A., & Horton, J. J. (2025). The Coasean singularity? Demand, supply, and market design with AI agents. NBER Working Paper 34468, November 2025. Checked.
• The Open Group. TOGAF Series Guide: Business Capabilities. Checked (guide date not confirmed).
• TheAgentCompany (2024). Xu, F. F., et al., TheAgentCompany: benchmarking LLM agents on consequential real world tasks. arXiv:2412.14161 (v1 18 December 2024: 24%; v3 10 September 2025: 30%). Checked.
• UK Centre for Digital Built Britain (2018). Bolton, A., Enzer, M., Schooling, J., et al., The Gemini Principles. doi:10.17863/CAM.32260. Checked.
• UK Companies Act 2006, section 155. legislation.gov.uk. Checked; legal confirmation pending.
• Williamson, O. E. (1975). Markets and Hierarchies. Free Press; (1985). The Economic Institutions of Capitalism. Free Press. Checked.
• Maynard, M. (2026). Flagship notes: recorded founder interview (142 pages), written answers and notes, made during October 2026. Private; quoted by page number. Mark’s account.

How this paper was written

This paper starts with the author: 25 years inside organisations, a recorded interview and his notes, where the ideas are set out. We use AI (mainly Claude, by Anthropic) as a writing and research assistant, and we say so openly.

• The ideas are the author's. They come from his experience, his interview and his notes, and we quote them by page.
• AI helps write them up: it turns that material into clear, well-organised text, makes sure no important point is missed, finds and checks supporting sources, and prepared the Spanish version.
• The author has the last word: he reviews, corrects and approves every section, and answers for every claim.
• The paper has not been submitted for a degree or to a journal. If that changes, we will follow that institution's rules on AI.
• The notes behind it are private, because they mention clients and colleagues. We quote them by page number, so every quotation can be held to its source.

Conflict of interest

The author is a founding member of Stratapault, a separate business that builds digital twins of organisations. Stratapault does not yet sell or operate the capability model in this paper, and nothing here describes what its product does today: it works at the second stage, seeing the organisation as it really is. This research is independent, free and exists to share the work; it is also where many of our ideas start, and we hope organisations will one day use Stratapault to carry out their own digital twinning. We mention Stratapault from time to time because it is the author's working model: the business where the idea of a twin of the organisation is put into practice.

Respond

Add your view.

Academics, practitioners, sceptics: tell us where this is right or wrong. Mark reads every response, and the best are published here with your name and role, with our reply.

Your name Email (never published)

Role and organisation Link to your profile or work (optional)

Why you are commenting: your field or experience, in a line or two Your response You may publish my response with my name and role, on this site and in future editions of the book. Remember my details on this device Send my response

We use your email only to reply. Privacy.