How companies might organise work with hypothetical superintelligent systems.
Compare capability, decision rights and complete customer outcomes before choosing an operating arrangement.

A company is more than a collection of tasks waiting to be automated. It makes promises to customers, combines resources, decides which work to accept, handles disagreement and remains answerable when something goes wrong. If a hypothetical superintelligent system could perform much of its analytical work, the company would still need an operating arrangement that turns capability into authorised, accepted outcomes. That arrangement could change substantially, but its shape would not be determined by intelligence alone.
This guide compares several possible arrangements rather than predicting one inevitable future. It asks where decisions sit, which handoffs remain necessary, what happens when reviewers become the limiting resource, and when a company should keep a capability inside its boundary or obtain it from outside. A complete fictional firm makes the comparison concrete. A second company then changes the numbers and the allocation of responsibility so that readers can test the reasoning independently.
In the eduKate series, Super Intelligence is an editorial umbrella for studying advanced AI and its implications. Artificial superintelligence, or ASI, is used here for a hypothetical broadly superhuman intellectual capability. No current product is presented as demonstrated ASI. All company figures and future capability assumptions below are invented teaching inputs. They are not business forecasts, staffing recommendations, investment advice or instructions to give a present system uncontrolled authority.
Choose your reading route
Understand future operating models · Compare Corin’s arrangements · Test failures and repairs · Try the independent company case
Capability, authority and company design
- 1. Start with the customer promise and work backwards
- 2. Make the future assumptions visible before drawing the company
- 3. Separate advice, preparation, commitment and acceptance
- 4. Strategy could become more analytical without becoming automatic
- 5. Three organisational arrangements can use the same intelligence
The complete Corin comparison
- 6. The complete Corin company packet
- 7. Establish the reference before crediting a future system
- 8. Central approval preserves the original bottleneck
- 9. Divisions can preserve context while stranding capacity
- 10. Bounded routine release changes the decision load
- 11. Reconcile the company accounts before calling the change a success
Stress, boundaries and people
- 12. Stress the exception boundary instead of assuming it stays stable
- 13. Handoffs need evidence of a shared understanding
- 14. Make or buy concerns the firm's boundary as well as price
- 15. Shared intelligence can create shared mistakes
- 16. People could do different work without becoming irrelevant
- 17. Use released capacity to investigate strategy rather than invent revenue
- 18. Evaluate an operating arrangement as a complete claim
Independent practice and conclusions
1. Start with the customer promise and work backwards
The most useful starting point is a finished customer outcome. A company might promise a correctly produced order, an agreed delivery window and a way to resolve problems. A system that generates attractive designs has completed only one part of that promise. Someone must establish that the design matches the order, that production can deliver it and that the company is authorised to make any proposed change. Without those connections, faster analysis can coexist with disappointed customers.
An operating model explains how the organisation combines work, resources and decision rights to fulfil that promise repeatedly. It includes the division of labour, the flow of information, the authority to commit resources and the response to exceptions. An organisation chart shows some of these relationships, but not all. A customer request may pass through people who report to different managers, and a shared service may constrain teams that appear independent on the chart.
A future company could use a powerful system to investigate demand, compare designs, prepare plans and coordinate routine work. That possibility does not tell us who chooses its strategy or accepts its obligations. A system may identify an attractive market while the company lacks the resources or permission to serve it. It may recommend a faster delivery date that the production team cannot honour. The operating model has to connect the recommendation to the relevant commitments.
This is why the question differs from measuring productivity. Productivity asks how useful output relates to inputs under a defined boundary. Here the boundary and the internal arrangement are themselves under examination. Should review be pooled or attached to divisions? Which routine decisions can follow standing rules? Which choices require a responsible person? A productivity measure may help compare the resulting arrangements, but it cannot specify those arrangements on its own.
It also differs from asking who receives economic gains. A company can complete more work without raising every employee’s income, lowering every customer’s price or changing its ownership. Those distributional questions deserve their own analysis. In this guide, the first task is to establish what the hypothetical firm can actually deliver, under whose authority, using which resources. Only then can a financial comparison have a defensible meaning.
A good reader should therefore resist two shortcuts. One is assuming that every existing management layer must remain unchanged. The other is assuming that superior intelligence makes organisation unnecessary. The interesting possibilities lie between them: fewer avoidable handoffs, clearer decision rights, different team boundaries and better use of expertise, alongside new dependencies and a continuing need to detect, contest and repair mistakes.
Back to contents · Next: 2. Make the future assumptions visible before drawing the company
2. Make the future assumptions visible before drawing the company
A conditional scenario should state what the system can do, what it knows, how reliably it performs and what resources it consumes. These are separate assumptions. A system could reason extremely well about a supplied problem while lacking current information about a customer’s changed requirements. It could generate a strong plan while taking too long or using too much capacity for a particular workflow. Calling it superintelligent does not fill those gaps.
Capability concerns the work the system can perform. Reliability concerns how often and under which conditions the required result occurs. Observability concerns whether other participants can tell what has happened. Controllability concerns whether the permitted behaviour can be constrained and corrected. A speculative operating model should not silently assign perfect performance to all four merely because it assumes a large improvement in one. Each assumption changes the comparison.
Resources matter as well. The company may need computing, records, connectivity, specialist staff and physical equipment. A hypothetical system could reduce some coordination costs while increasing demand for another resource. For example, more candidate products might create additional testing and approval work. The future scenario becomes informative when it shows where the burden moves, rather than treating every activity outside model inference as free and instantaneous.
Authority is another independent variable. A system can be able to draft a contract without being authorised to bind the company. It can identify a cost reduction without being entitled to change a worker’s conditions. It can predict customer preferences without having permission to repurpose personal information. Those boundaries are not evidence of low intelligence. They describe the relationship between a technical participant and the people and institutions responsible for its use.
For the central case, we will grant unusually strong but bounded preparation capability. The hypothetical system produces the stipulated order materials within a known machine-time budget. It also identifies the exact exception counts supplied by the fictional packet. This is deliberately generous analytical support. We will not assume that it may rewrite customer terms, waive required review, change its own authority or create physical capacity. That separation allows us to study organisation rather than hide it inside a magical premise.
A later stress test changes the exception burden while preserving the system’s ability to identify it. The company then faces more legitimate decisions, not merely a less accurate model. This distinction is important. Even perfect recognition of every unusual case would not guarantee that an authorised team has enough time to resolve all of them. Intelligence can reveal a constraint without automatically removing the resource or authority that makes it a constraint.
Back to contents · Next: 3. Separate advice, preparation, commitment and acceptance
3. Separate advice, preparation, commitment and acceptance
Four stages are useful for thinking about a future company. Advice proposes what could be done. Preparation creates something ready for consideration or use. Commitment changes what the company is obligated or authorised to do. Acceptance establishes that the promised result has been delivered under the relevant conditions. A single software interface may display all four, but the organisational meanings remain different.
Consider a proposed change to a custom order. The system might advise using a different format because it would be easier to produce. It could prepare the revised artwork and a cost comparison. The customer and responsible company representative would still need to accept the change where their agreement requires it. Only after the agreed output is checked and delivered can the company count that order as completed. An excellent draft does not substitute for the missing agreement.
The distinction also applies to internal budgets. A system can compare projects using supplied costs, but selecting a project may commit money and other people’s time. A manager may have authority only within a defined budget or category. Giving an analytical tool access to the numbers does not grant it the same authority. A coherent operating model identifies which decisions are reversible, which create external obligations and which require a particular responsible role.
Current governance guidance already treats these relationships as organisational questions. The NIST AI Risk Management Framework’s Govern function includes documented responsibilities, communication, executive responsibility for AI-risk decisions and differentiated human–AI roles. This is a limited factual reference to the 2023 voluntary framework. It does not certify the future arrangements in this article or establish the legal requirements of a particular company.
A hypothetical superintelligent adviser might be better than a human at comparing options while still not being the legitimate owner of a decision. Conversely, a person with formal authority may need substantial assistance to understand the evidence. A useful arrangement connects the two: strong analysis supports accountable judgment, and the authorised decision is recorded in a form the rest of the company can use without guessing.
Acceptance should be equally concrete. Does complete mean prepared, internally checked, released, received or accepted by the customer? Different teams may otherwise report the same order at different stages as if it were the same result. The central case defines completed orders narrowly enough to count. That prevents a faster design team, a busy reviewer and an enthusiastic sales team from each claiming a separate success before the customer promise has actually been fulfilled.
Back to contents · Next: 4. Strategy could become more analytical without becoming automatic
4. Strategy could become more analytical without becoming automatic
Strategy concerns which problems the company chooses to solve and how it intends to sustain a useful position. A stronger analytical system might compare more possibilities, combine evidence from more sources or expose inconsistencies in a plan. Those contributions could improve strategic discussion. They would not make demand, customer trust or resource availability certain. A scenario that assumes a clever plan automatically creates paying customers has already assumed much of the outcome it claims to explain.
The organisation must distinguish an attractive opportunity from an authorised experiment and a committed offer. A proposed product can remain at the research stage while the company tests whether people need it. An experiment should have a purpose, bounded resources and a way to recognise an unpromising result. A committed offer must connect to the capacity and quality the company can actually supply. More ideas can make that discipline more important rather than less.
Decision speed is not the only objective. Some choices benefit from rapid action; others benefit from waiting for information, hearing affected people or comparing alternatives. A system capable of producing a recommendation immediately does not prove that immediate commitment is wise. The operating model should preserve useful deliberation while removing avoidable delay. It should explain the reason for a pause rather than treating every pause as a failure of automation.
Strategic authority may also be distributed. Product specialists may understand a customer problem, operations staff may understand feasibility, and leadership may hold responsibility for the company’s direction and commitments. A future system could help reconcile their evidence without erasing their different responsibilities. The question is which disagreement concerns facts, which concerns uncertainty and which concerns values or priorities. Those require different kinds of resolution.
There is a risk of confusing more detailed forecasts with more reliable ones. A model can describe a richly specified future even when key inputs are uncertain. A responsible comparison keeps alternative scenarios alive and identifies what would change the decision. If demand is the uncertain variable, producing another elaborate staffing chart may add little. The next useful activity could be a limited demand test rather than further optimisation of an unverified plan.
In the fictional company below, demand is supplied rather than predicted. That is a deliberate boundary. It lets us examine internal organisation without claiming to solve market strategy at the same time. A later strategic exercise will ask what the company should investigate before expanding. It will not infer a new market simply because the hypothetical system can prepare additional orders more quickly.
Back to contents · Next: 5. Three organisational arrangements can use the same intelligence
5. Three organisational arrangements can use the same intelligence
The first arrangement is a central approval office. One pooled review team sees every order before release. This can offer a consistent view of company commitments and allow reviewers to share capacity across categories. It can also become a queue through which even routine work must pass. Its performance depends on the volume and complexity of decisions, not only the speed of the systems preparing them.
The second arrangement assigns review to separate product divisions. Each division combines local knowledge with authority over its own work. This may improve context and accountability for a particular customer group. However, unused capacity in one division may not help another if reviewers are not authorised or qualified to move across the boundary. The aggregate staff-time total can conceal an overloaded division alongside a quieter one.
The third arrangement uses bounded routine release with an exception team. Orders satisfying a declared, pre-authorised specification can proceed through a validated routine path. Custom work and exceptions remain with responsible reviewers. A separate audit examines some routine outputs, and authority remains outside the system that prepares them. This arrangement can reduce repetitive approvals, but it relies on a justified boundary between routine and non-routine work.
None of these is universally best. A central office may be appropriate when decisions are tightly connected or review cannot be delegated. Divisions may be valuable when knowledge is specialised and local consequences differ. Bounded routine release may be useful when the permitted class can be defined and checked reliably. A company might combine the three across different products instead of selecting one pattern for everything it does.
The comparison should hold capability assumptions constant when the purpose is to examine organisation. If one option receives a more powerful system, more staff and fewer quality obligations, its apparent advantage cannot be attributed to its organisational structure alone. In the central case, all future arrangements receive the same hypothetical preparation capability and the same total reviewer budget. The distribution of work and permission changes, which is precisely what we want to study.
We also retain a fictional present-style reference arrangement. It uses human preparation and central review. This does not claim that all current firms work that way or that all future firms will automate the same activities. It provides a stable comparison that can expose an uncomfortable but useful result: buying stronger preparation capability may increase expense while leaving the customer outcome constrained by an unchanged approval system.
Back to contents · Next: 6. The complete Corin company packet
6. The complete Corin company packet
Corin is an invented company producing non-safety-critical printed information packs for fictional exhibitions. Its study week contains 120 orders: ninety standard and thirty custom. All orders are available at the beginning of the week. An order is counted as completed only when its agreed content is prepared, all required review is finished, it is released through an authorised route and the customer receives the stipulated accepted pack within that week. Partial orders do not count.
A standard order earns 200 study credits and consumes eighty credits of variable production and delivery resources. A custom order earns 320 and consumes 140. Revenue is recognised only for completed orders in this simplified exercise. Variable production cost is incurred only for completed orders; no separate variable charge is assigned to prepared but held work. Common weekly staff and overhead costs are 6,000 credits, remain payable in every arrangement and already include the staff capacities below. Taxes, financing charges and other unlisted costs are zero by assumption.
In the present-style reference arrangement, human preparation takes twelve minutes for each standard order and thirty for each custom order. The preparation team has 2,400 paid minutes available. Review capacity is a separate pool of 1,200 paid minutes. Printing and delivery together can complete up to 130 orders, with sufficient materials for every order in the packet. Preparation, review and physical production are separate capacities and cannot be exchanged merely by adding their minutes together. The packet uses aggregate work budgets: different orders may move through overlapping stages, and there are no additional intraweek sequencing, setup or transport-time constraints. Once the stated budgets and permissions are satisfied, completion within the week is stipulated.
The future system needs four machine minutes per standard order and twelve per custom order. Its weekly budget is nine hundred machine minutes. Human supervision of preparation uses 120 minutes from the preparation team’s existing 2,400-minute budget. This supervision is separate from the review work described next. A fixed technology-service charge of 1,200 credits per week applies to every future arrangement and does not apply to the present-style reference. There is no additional setup charge within this operating-week comparison.
The supplied workload has nine standard exceptions and twelve custom exceptions. Therefore there are eighty-one routine standard orders, nine exceptional standard orders, eighteen non-exceptional custom orders and twelve exceptional custom orders. The packet stipulates that these labels are correct and available before review. Every manually reviewed order needs six minutes for release review. Each exception needs an additional twenty-four minutes of resolution, for thirty minutes total. This includes all correction and rechecking effort in the base case.
Custom orders always require human release under Corin’s stipulated customer arrangements. Standard exceptions also require it. Only the eighty-one routine standard orders may use a standing, pre-authorised automatic release rule in the bounded arrangement. Those routine outputs meet the agreed specification by assumption; no real-system reliability claim follows. Nine of them receive a six-minute audit each, selected without additional sampling overhead. The audit is additional work, not a substitute for the stated permissions.
For divisional comparison, standard and custom review teams each receive six hundred of the same total 1,200 review minutes. They cannot transfer time or authority during the study week. For central review, the full 1,200 is pooled. For bounded routine release, all remaining release reviews and the nine audits use the pooled 1,200. General governance work has already been budgeted outside these available minutes, and its cost is included in the common 6,000. No arrangement may silently remove a required review to claim completion.
Back to contents · Next: 7. Establish the reference before crediting a future system
7. Establish the reference before crediting a future system
Human preparation of all 120 orders requires ninety times twelve plus thirty times thirty minutes. That is 1,080 + 900 = 1,980 minutes, within the available 2,400. Preparation therefore has 420 minutes remaining. Printing and delivery can accommodate 130 orders, exceeding the 120-order demand. Neither is the binding capacity in this particular reference week. Those conclusions follow from the supplied packet rather than a general statement about printing businesses.
Review tells a different story. All 120 orders need six minutes, creating 720 minutes of basic release work. Twenty-one exceptions need an additional twenty-four minutes each, creating 504 minutes. Total review demand is 1,224 minutes, twenty-four more than the available 1,200. A report that counts prepared designs would announce 120 completed jobs. Under the defined customer outcome, the company cannot complete all 120 through this central-review arrangement.
One feasible response is to hold one exceptional standard order. It removes thirty minutes of review, leaving 1,194 minutes and six minutes unused. Completed output is then eighty-nine standard plus thirty custom orders, or 119. The held order is explicitly unmet demand for the week. There is no claim that this is a universally fair way to choose whose order waits. The example is an identified feasible schedule, not an undisclosed priority policy.
No schedule can complete all 120 because that would require 1,224 minutes. Since the example completes 119, the maximum count is exactly 119. Multiple schedules can reach that count, including holding a custom exception instead. They can have different revenue and customer consequences. Counting maximum completed orders is therefore a narrower objective than choosing the best business response. A company would still need an authorised policy for the consequences of delay.
For the displayed standard-order hold, revenue is 89 × 200 + 30 × 320 = 27,400 credits. Variable cost is 89 × 80 + 30 × 140 = 11,320. Subtracting common fixed cost gives an operating surplus of 27,400 − 11,320 − 6,000 = 10,080. The word surplus refers only to this simplified packet. It is not a valuation of the company, a cash forecast or evidence that leaving an order unfulfilled is acceptable.
The reference establishes the problem that organisation must solve. It is not lack of drafting time or physical output capacity. It is a small review shortfall and a decision about which commitment cannot be completed. A more capable preparation system may still be useful for other reasons, but it should not receive credit for solving this particular problem unless the comparison shows how review, authority or the customer promise changes.
Back to contents · Next: 8. Central approval preserves the original bottleneck
8. Central approval preserves the original bottleneck
Under the hypothetical system, preparation needs ninety times four plus thirty times twelve machine minutes. The total is 360 + 360 = 720, leaving 180 of the nine hundred machine minutes unused. Human preparation supervision takes 120 minutes, leaving 2,280 of the preparation team’s paid minutes available for other authorised work. The model has transformed the preparation activity in the scenario. That does not mean every other stage has changed with it.
Central approval still sends every order to the pooled reviewers. The same 720 basic minutes and 504 exception minutes remain, so total review demand is still 1,224. The maximum completed count remains 119. Faster preparation has not created the missing review minutes, changed the custom-order commitments or granted a new release authority. This is a strong example precisely because the future system has been given very capable preparation performance.
Using the same exceptional-standard hold makes the financial comparison like for like. Revenue remains 27,400 and variable cost remains 11,320. Common cost remains 6,000, but the new technology fee adds 1,200. The weekly surplus becomes 8,880, which is 1,200 lower than the reference. The prepared materials arrive through a different process, yet the customer count does not rise and the supplied costs do not fall.
That result does not prove the technology has no value. Corin now has substantial preparation capacity available for other work, and the system might provide analytical benefits not measured in this packet. However, none of those possibilities can be entered as revenue or cash savings without specifying an actual use and its consequences. The common staff bill remains paid. A claimed wage saving would contradict the exercise unless a separate, authorised scenario changed that cost.
Centralisation may nevertheless have reasons beyond throughput. Reviewers could need a shared understanding of unusual commitments or consistent interpretations of customer terms. A pooled team can also use spare capacity across categories, something the divisional arrangement cannot do in the base week. The problem is not that central review is inherently foolish. It is that a firm should understand what it buys, what it constrains and which work genuinely requires it.
The conditional lesson is specific: if preparation capability improves while the same people must make the same release decisions with the same time budget, the approval bottleneck can remain. A future operating model must address the distribution and legitimacy of decisions as well as the production of drafts. Replacing the word manager with the word agent would not change any of these quantities or obligations.
Back to contents · Next: 9. Divisions can preserve context while stranding capacity
9. Divisions can preserve context while stranding capacity
Now divide the review budget into two six-hundred-minute teams. The standard division receives ninety orders and nine exceptions. Its basic review takes ninety times six, or 540 minutes. Exception resolution adds nine times twenty-four, or 216. Total required standard review is 756 minutes. The custom division needs thirty times six plus twelve times twenty-four, or 180 + 288 = 468 minutes.
The totals still add to 1,224, exactly as before. The distribution changes the operational result. Standard review exceeds its local budget by 156 minutes, while custom review has 132 minutes unused. The company cannot simply net these two numbers if its reviewers lack cross-division authority or qualification during the study week. Unused time is not a transferable resource until the arrangement makes it transferable.
The standard division can complete eighty-four orders by holding five standard exceptions and one routine standard order. Five exceptions remove 150 minutes, and one routine order removes six, reducing review from 756 to six hundred exactly. The custom division completes all thirty within its 468 minutes. The company therefore completes 114 orders: eighty-four standard and thirty custom, despite owning the same aggregate review budget as the central arrangement.
Why can the standard division not complete eighty-five? That would mean holding only five orders. The largest review requirement of any standard order is thirty minutes, so five holds can remove at most 150 minutes. The division needs to remove 156. At least six orders must be held. Because the displayed schedule achieves that minimum number of holds, eighty-four is the maximum standard-order count under the stated rules.
Revenue for this displayed outcome is 84 × 200 + 30 × 320 = 26,400 credits. Variable cost is 84 × 80 + 30 × 140 = 10,920. With 6,000 common cost and 1,200 technology cost, surplus is 8,280. Compared with central future approval’s 8,880, this is six hundred lower. Five fewer completed standard orders each contribute 120 credits after their variable costs, reconciling the difference.
Again, the conclusion is bounded. Divisions might perform better in a setting where local knowledge reduces resolution effort or where pooled review makes more mistakes. Those benefits are not supplied here, so they cannot be assumed to cancel the stranded capacity. The case shows how an organisation boundary can affect delivery even when the analytical system, total staff minutes, demand and nominal company resources are unchanged.
Back to contents · Next: 10. Bounded routine release changes the decision load
10. Bounded routine release changes the decision load
The third arrangement uses Corin’s stipulated standing authority for routine standard work. Eighty-one standard orders can proceed without individual human release. All thirty custom orders and nine standard exceptions still require it, making thirty-nine manually released orders. Their basic review takes 39 × 6 = 234 minutes. The twenty-one exceptions still require 504 additional minutes. Manual release and resolution together therefore consume 738 minutes.
Nine audits add 9 × 6 = 54 minutes, bringing the total to 792. That is within the 1,200-minute pool and leaves 408 minutes. Machine preparation still uses 720 of nine hundred minutes, supervision uses 120 of 2,400 human preparation minutes, and physical production uses 120 of 130 order slots. All 120 orders can now be completed under the packet’s defined permissions and quality assumptions.
The improvement is not permission discovered by the model. Corin’s arrangement supplies the standing authority in advance for a carefully identified class. The hypothetical system does not choose to remove review because the queue is inconvenient. Nor does the nine-order audit make every other order permissible. The legitimacy of routine release and the evidence supporting its use are separate conditions, both stipulated for this teaching comparison.
An unbounded proposal would say that the system should release all 120 because it prepared them correctly. That violates the requirement for human release on thirty custom orders and nine standard exceptions. It would produce thirty-nine unauthorised releases in the fictional rule system. Accurate preparation is insufficient to make those commitments acceptable. The comparison distinguishes capability from authority using an explicit count rather than a general warning.
The remaining 408 review minutes are capacity, not automatically savings. Corin could reserve them for a change in demand, use them for deeper investigation or assign other authorised work compatible with the reviewers’ skills. One illustrative allocation is 120 minutes for evaluating strategic assumptions, 180 for incident reserve and 108 for skills development. These sum to 408. They are a possible use of time, not extra revenue already earned.
This arrangement works because the packet supplies a stable routine class, trustworthy labels, appropriate standing permission and sufficient review for everything outside that class. If any condition changes, the conclusion needs revision. A future company could become less approval-heavy in some activities while becoming more demanding about how categories, evidence and accountability are maintained. Reducing repetitive decisions is different from eliminating responsibility.
Back to contents · Next: 11. Reconcile the company accounts before calling the change a success
11. Reconcile the company accounts before calling the change a success
With all 120 orders completed, revenue is ninety times 200 plus thirty times 320, giving 18,000 + 9,600 = 27,600 credits. Variable cost is ninety times eighty plus thirty times 140, giving 7,200 + 4,200 = 11,400. Subtract common cost of 6,000 and future technology cost of 1,200. The bounded arrangement’s weekly operating surplus is 9,000 credits.
Compared with central future approval, the bounded arrangement completes one additional standard order and increases surplus from 8,880 to 9,000. The increase is 120 credits, equal to the standard order’s 200 revenue minus eighty variable cost. Compared with the divisional future arrangement, it completes six additional standard orders and increases surplus by 720. These reconciliations show that the financial differences are connected to the counted outcomes rather than separate unexplained claims.
Compared with the present-style reference, however, bounded routine release produces a lower surplus: 9,000 rather than 10,080. Completing one more standard order contributes 120, while the new technology service costs 1,200. The net difference is negative 1,080. This is an important result. The packet demonstrates an organisational route to fulfilling all demand, but it does not demonstrate a higher weekly surplus than the reference under the supplied prices and fixed costs.
A company might still value meeting the additional promise, preserving preparation capacity or improving other work. The case does not assign prices to those benefits, so it cannot decide whether they justify the difference. It also cannot count avoided penalties or retained future customers, because none are supplied. A defensible decision records the demonstrated result and the missing valuation instead of quietly inventing an economic benefit to make the preferred technology win.
The time comparison must be equally careful. Reference preparation uses 1,980 human minutes, while future preparation supervision uses 120. The difference is 1,860 human minutes, or thirty-one hours, available within the same paid preparation budget. That is not a comparison between human and machine minutes as if they were interchangeable units. It is a comparison of the stated human activity before and after, with machine use separately reported as 720 minutes.
A complete company memo would therefore show at least three outcomes: accepted customer orders, operating surplus under the supplied accounting, and the amount and type of staff capacity available for other work. It would not compress them into a single claim that the company is better. The best arrangement depends on the objective and the value of consequences that the toy packet deliberately leaves unpriced.
Back to contents · Next: 12. Stress the exception boundary instead of assuming it stays stable
12. Stress the exception boundary instead of assuming it stays stable
Change one consequential workload assumption. Keep total demand at ninety standard and thirty custom orders, but suppose there are now thirty standard exceptions and eighteen custom exceptions. All other preparation, capacity, price and authority assumptions remain unchanged. The hypothetical system identifies the new labels correctly. This is not a test of hallucination; it is a test of whether the organisation can handle a more demanding mix of legitimate decisions.
There are now sixty routine standard orders eligible for bounded release. The thirty standard exceptions and all thirty custom orders require human release, giving sixty manually reviewed orders. Basic review consumes 60 × 6 = 360 minutes. Forty-eight exceptions add 48 × 24 = 1,152 minutes. Retain the nine six-minute routine audits, adding fifty-four. Total review demand becomes 1,566 minutes, exceeding the 1,200-minute pool by 366.
The system can still prepare every order within 720 machine minutes because preparation times are unchanged by stipulation. It can identify exactly what needs attention. Yet the company cannot complete every order under the bounded arrangement without changing another resource or the service promise. This is a case in which greater awareness reveals more required work. Treating the additional exceptions as inconvenient noise would destroy the meaning of the review boundary.
If capacity cannot change and every retained order must satisfy the same review requirement, at least thirteen orders must be held to maximise count. Twelve thirty-minute holds remove 360 minutes, leaving 1,206, which is still six too many. Thirteen remove 390, leaving 1,176. One feasible schedule holds thirteen standard exceptions, completes seventy-seven standard plus thirty custom orders and retains the fixed nine audits. Output is then 107, with twenty-four review minutes unused.
That maximum-count argument does not determine which customers should wait. A company needs an authorised priority and communication policy. The example chooses standard exceptions only to make one schedule explicit. A different permitted choice could change revenue, costs and the consequences of delay. The key diagnosis is the 366-minute shortfall. A stronger model, a more confident explanation or an unchanged audit sample does not by itself supply those minutes.
A targeted repair is to obtain 420 additional qualified, authorised review minutes for the stress week at a stipulated all-inclusive cost of 1.50 credits per minute. That adds 630 credits and increases capacity to 1,620. All 120 orders then fit, with fifty-four minutes remaining. Surplus becomes 9,000 minus 630, or 8,370. The repair is a supplied scenario, not a recommendation to hire on those terms or a claim such reviewers are available in any real market.
Back to contents · Next: 13. Handoffs need evidence of a shared understanding
13. Handoffs need evidence of a shared understanding
A handoff is not complete merely because a file moved between systems. The next participant needs to know what the work means, which version is authoritative, which assumptions remain open and what decision has actually been made. A future system might transmit information rapidly while still leaving an ambiguity unresolved. In a company, that ambiguity can become duplicated work or an unintended commitment when different teams act on different interpretations.
Consider a wholly fictional extension to Corin’s ordinary week. A preparation record refers to version B of a customer’s layout, but the release packet displays version A. The system has not changed the agreed content; it has failed to carry the correct reference through the handoff. The appropriate result is to stop that affected release and reconcile the source with the responsible people. A plausible-looking version is not a substitute for an agreed version.
Suppose the teaching packet identifies ten such affected orders and assigns eighteen additional review minutes to each correction and recheck. The extra work is 180 minutes. Added to the bounded base-week load of 792, total review becomes 972, leaving 228 of the original 1,200 minutes. The company can complete the week under these supplied conditions, but its incident reserve is partly consumed. Calling the error harmless because capacity happened to absorb it would miss that cost.
This extension is separate from the heavier-exception stress week. Combining them would require adding their burdens explicitly and checking whether the same orders and minutes overlap. A report should never quietly reuse the same reserve for several simultaneous incidents. The model’s capacity to describe many repairs does not establish that the organisation can carry them out concurrently. Resource accounting must follow the actual scenario being claimed.
The broader lesson is about the content of a handoff. It should preserve a stable identity for the customer outcome, the relevant source, the authorised decision, the current status and the unresolved exception. These are functional requirements, not a request to reproduce a particular workplace integration procedure. Their purpose is to make it possible for a different responsible participant to continue the work without inventing the missing meaning.
Evidence of a shared understanding can reduce the need for some coordination meetings while increasing the quality of those that remain. A meeting may be unnecessary when all parties can inspect the same completed decision. It may be essential when objectives conflict or an exception lacks an authorised answer. A future company should remove redundant transmission, not erase the moments when people genuinely need to resolve what the company is promising.
Back to contents · Next: 14. Make or buy concerns the firm's boundary as well as price
14. Make or buy concerns the firm's boundary as well as price
A company might build an internal capability, buy a service or combine both. Stronger AI could change the effort required to maintain internal software, prepare specifications or compare offers. That could move the boundary between in-house work and purchased services. It would not make one boundary universally best. The company still needs to consider responsibility, maintainability, quality, continuity and the resources required to make each arrangement workable.
Use a separate prospective component exercise for Corin, not an extra charge to add to the previous weekly accounts. Two mutually exclusive ways of obtaining a scheduling component meet the same stipulated service and authority requirements. Internal development costs 9,000 credits upfront, then three hundred weekly plus 0.50 for each completed order. Buying the component costs seven hundred weekly plus two credits per order, with no setup fee. All other comparison costs are equal and excluded. Compare post-setup operating weeks; no implementation duration is supplied, so this module does not determine a calendar launch date.
At 120 orders per operating week, internal running cost is 300 + 0.50 × 120 = 360 credits. Purchased running cost is 700 + 2 × 120 = 940. The internal route is 580 cheaper each operating week after setup. Dividing 9,000 by 580 gives approximately 15.52 weeks. Fifteen whole weeks save 8,700, still short of setup. Sixteen save 9,280, exceeding setup by 280 under the simplified undiscounted assumptions.
The horizon changes the preferred cost result. Over twelve operating weeks, internal total is 9,000 + 12 × 360 = 13,320, while buying costs 12 × 940 = 11,280. Buying is 2,040 cheaper over that horizon. Over twenty-four weeks, internal total is 17,640 and buying costs 22,560, making internal provision 4,920 cheaper. Neither result proves the cheaper option has the better quality or risk profile; equivalence was explicitly stipulated for this arithmetic.
Now reduce demand to sixty orders weekly. Internal running cost becomes 330, while buying becomes 820, a difference of 490. The simple setup recovery point moves to about 18.37 weeks, so nineteen whole operating weeks are needed to exceed setup. This sensitivity shows why a make-or-buy decision needs an expected use and time horizon. Technical ability to build a component does not alone establish that building it is a useful allocation of company resources.
The firm’s boundary also determines who maintains the component and who explains a failure. If the company buys, its customer promise does not disappear when the supplier disappoints it. If it builds, it inherits continuing work rather than a permanently free asset. These responsibilities can matter even when the simple costs are close. The complete decision should name the duties that cross the boundary, while leaving wider supplier bargaining and market concentration to the separate input-power guide.
Back to contents · Next: 15. Shared intelligence can create shared mistakes
15. Shared intelligence can create shared mistakes
A common analytical system can help teams use consistent information and compare plans. It can also spread one mistaken assumption through many activities. If sales, preparation and scheduling all rely on the same incorrect interpretation, their apparent agreement is not independent confirmation. The company may have three matching reports and only one underlying judgment. A future operating model should distinguish consistency from corroboration.
Independent checks need not repeat every task from scratch. They can examine a different evidence source, verify a consequential constraint or compare an output with a separately maintained acceptance requirement. The relevant question is whether the check can reveal the kind of failure being considered. Asking the same system to restate its conclusion more confidently may improve readability without adding meaningful independence.
Current generative AI provides a reason to take this distinction seriously without claiming that future systems must have identical weaknesses. The NIST Generative Artificial Intelligence Profile, published in 2024, identifies risks including confabulation and describes risk-management considerations for generative systems. That source concerns present technology and assurance practice. The future case here assumes particular capabilities only for the tasks explicitly stated; it does not generalise from them to universal trustworthiness.
A company also needs a workable response when shared analytical services are unavailable. At a conceptual level, some commitments may continue through established alternatives, while new or unusual work may pause. The appropriate arrangement depends on the service and its consequences. This article does not provide operational security instructions or a universal fallback plan. It asks whether the company’s promises remain consistent with what it can verify and authorise under changed conditions.
The cost of maintaining independent knowledge should be visible. If every person loses the ability to interpret the firm’s own records, the company may become less able to challenge an output or change suppliers. Preserving expertise consumes time, but so does recovering from an arrangement nobody understands. The right comparison includes the capability to diagnose and correct a problem, not only the speed of routine production on a favourable day.
For hypothetical ASI, the challenge could become sharper rather than vanish. A system may produce conclusions that people cannot easily reproduce unaided. The response cannot be to infer legitimacy from complexity. The company would need assurance methods appropriate to the consequences, clear limits on unverified claims and institutions able to contest consequential decisions. This is a conditional governance question, not an assertion that ordinary review can fully validate any conceivable superintelligent system.
Back to contents · Next: 16. People could do different work without becoming irrelevant
16. People could do different work without becoming irrelevant
If preparation and routine coordination become less labour-intensive, human work may shift toward understanding customers, selecting goals, resolving exceptions, evaluating evidence and maintaining relationships. That is a possible direction, not a guaranteed employment outcome. The number of roles, their pay and their quality depend on demand, ownership, institutions, costs and the choices organisations make. The central packet deliberately keeps staffing costs fixed so that released time is not misrepresented as an automatic wage saving.
A future operating model should ask whether the remaining work is learnable and sustainable. Concentrating every difficult exception on a small group can make their work more demanding even while total task volume falls. A reviewer who sees only unusual, ambiguous or contested cases may need different training and support from someone checking a predictable mix. Fewer routine decisions can therefore improve throughput while changing the experience of work in ways a simple time total misses.
The OECD’s 2023 employer and worker survey report studies current AI use in selected finance and manufacturing settings and discusses training, consultation and worker outcomes. Its associations are not proof that a particular consultation process causes every favourable result. The relevant lesson for this article is to investigate how workers experience a change rather than assume that an efficiency calculation fully describes it.
Employee knowledge can also improve the model of the company. People may know why an apparently redundant step exists, which customer requirements frequently change or which data field is unreliable. Consultation is useful when that knowledge can change a proposed arrangement. It is less useful when the future chart is already fixed and people are asked only to endorse it. Better intelligence should make room for more relevant evidence, including evidence from the people doing the work.
Learning pathways matter because expertise is often developed through ordinary tasks before people handle difficult ones. If all entry-level practice disappears, the organisation needs another way to help future reviewers acquire judgment. Simulations, supervised work and explicit teaching may help, but their effectiveness must be evaluated. A company cannot assume an unlimited supply of experienced exception handlers while removing the processes through which experience develops.
The humane question is therefore concrete: what will each group be asked to do, how will they learn it, what support will they have and how can they raise a concern? Those questions remain relevant even if a machine performs many cognitive activities better than a person. A company is still an organisation of commitments among people, not merely a container for the most capable available calculation.
Back to contents · Next: 17. Use released capacity to investigate strategy rather than invent revenue
17. Use released capacity to investigate strategy rather than invent revenue
Corin’s future arrangements leave substantial preparation-team time available. That creates an opportunity to investigate a new offer, but it does not establish a profitable offer. The company could use some of that time to understand unmet customer needs, test a revised design or improve its own evidence. A disciplined strategic process distinguishes the value of learning from the later value of a service that customers actually accept.
Suppose the company considers adding multilingual exhibition packs. Before entering any new revenue in the accounts, it would need evidence about demand, appropriate language quality, review requirements, production costs and authority to use the supplied content. A capable system may help prepare examples or identify questions. It cannot make those missing conditions true merely by writing a business plan. The initial output should be a testable proposition with explicit uncertainties.
For this guide, no new numerical demand is invented for that offer. The central packet remains fixed at 120 existing orders. This prevents spare time from being counted twice, once as a saving and again as income from an unspecified new product. If a separate expansion scenario is later supplied, it should include the extra review burden, physical capacity and costs, not only the additional selling price.
A useful experiment also needs a stopping condition. The company might decide that the offer cannot progress until independent language review establishes an acceptable result within the available time. If the evidence is weak, stopping is a successful learning outcome even though it produces no launch. The future system’s ability to generate more variants does not justify an indefinite cycle that never resolves the uncertainty relevant to the decision.
This illustrates a broader organisational possibility. Stronger intelligence could make companies more experimental, with smaller commitments before larger ones, if it lowers the cost of preparing and comparing evidence. It could also make them more overconfident if polished plans are mistaken for validated opportunities. The operating arrangement determines how ideas encounter constraints, disagreement and independent checks before the company commits resources or promises delivery.
The strategic benefit of released capacity is therefore conditional on what the organisation does with it. Time used for a relevant experiment, better customer understanding or staff learning may be valuable. Time spent producing more untested presentations may not be. Measuring the next accepted result, or the uncertainty genuinely resolved, keeps the company connected to useful progress without forcing every learning activity into an immediate revenue claim.
Back to contents · Next: 18. Evaluate an operating arrangement as a complete claim
18. Evaluate an operating arrangement as a complete claim
A real evaluation would begin with a defined service, an explicit comparison and a record of the assumptions being tested. It should include all relevant stages, including preparation, review, release, delivery and correction. The purpose is not to copy Corin’s numbers into a real company. It is to use the same discipline to discover which parts of a proposed arrangement are supported and which remain uncertain.
Capability evidence should match the task. A model that prepares one successful example has not demonstrated stable performance over the intended workload. A test dominated by routine cases may not represent the difficult decisions concentrated in an exception team. The company needs an account of what was evaluated, what was excluded and what happens outside the tested scope. Rejected outputs and unresolved cases belong in that account rather than disappearing from the denominator.
Authority evidence is different. It concerns who may make a decision, what can be delegated, which commitments need approval and how the decision is communicated. A technically correct action can still exceed the system’s role. The NIST AI RMF discussion of human–AI interaction emphasises distinguishing human roles and responsibilities. That supports asking the question; it does not answer every jurisdiction-specific permission issue for the company.
Resource evidence should include the work introduced by the arrangement. Audits, incident handling, training, maintaining alternative routes and investigating disputes require capacity. An evaluation that counts only automated task time can make a fragile organisation look efficient. Conversely, treating all checking as waste can remove the very activity that makes the customer outcome trustworthy. The task is to understand which checking is necessary and how its burden changes under each arrangement.
The company should also preserve a way to revise the arrangement. A new customer mix, revised model, changed staff capability or different supplier service can invalidate an earlier conclusion. Review dates alone are insufficient if nobody knows what evidence would trigger a change. The stress test in this article gives one example: when exception burden exceeds review capacity, the company must obtain capacity, revise commitments through the proper process or hold work rather than conceal the shortfall.
Evaluation should end in a bounded decision: continue under the tested scope, revise a particular assumption, gather a missing form of evidence or stop. It should not conclude that the company is now generally autonomous. Operating arrangements concern specific work, people and consequences. A successful change in one product line may provide useful learning for another, but transfer still requires checking the new capability, resource and authority conditions.
Back to contents · Next: 19. Independent company packet: Marlow's two product teams
19. Independent company packet: Marlow's two product teams
Marlow is a separate invented company producing customised display cards. Its study week has 150 orders: one hundred simple and fifty custom. All must be completed within the week to count. Simple orders earn 160 credits and use sixty of variable resources; custom orders earn 280 and use 120. Common fixed weekly cost is 7,000 credits, and the hypothetical system costs another 1,400. All stated prices include the costs relevant to this exercise; taxes and other unlisted charges are zero.
The future system prepares a simple order in five machine minutes and a custom order in fifteen. Its weekly machine capacity is 1,500 minutes. Physical completion capacity is 160 orders. Human release-review capacity is 1,100 minutes after all general governance work has been budgeted. Each manually released order uses five minutes. An exception adds fifteen more minutes, making twenty total. There are fifteen simple exceptions and twenty custom exceptions, correctly identified by assumption.
Marlow’s customer arrangements require human release for all custom orders and all simple exceptions. The eighty-five routine simple orders may use a validated standing-release rule. In the bounded arrangement, ten routine simple orders receive five-minute audits, adding fifty minutes. In the central arrangement, every order receives manual release and no separate routine audit is added. The total review budget remains 1,100 in both cases.
For a divisional variant of central review, the simple team has five hundred minutes and the custom team six hundred. They cannot share review authority or capacity during the week. Preparation quality, exception counts, machine times, prices and physical capacity are the same across arrangements. As with Corin, the question is how the organisation distributes work and decisions, not which option secretly receives the best technology.
First calculate machine preparation load and show whether physical capacity is sufficient. Then calculate central-review load, its shortfall and the maximum completed-order count if orders are indivisible. Give one feasible schedule achieving that count. Next test the divisional arrangement, explain why aggregate review minutes can mislead, and find a feasible maximum-count outcome. Do not invent new reviewers or waive a required approval.
For bounded routine release, calculate manual release work, exception resolution, audit time and remaining review capacity. Reconcile full-completion revenue, variable cost and surplus. Finally, change the workload to thirty simple exceptions and thirty custom exceptions, keeping everything else fixed, including the ten audits. Find the new review load and the extra capacity needed to complete all orders. Explain which conclusion depends on arithmetic and which depends on the stipulated authority and quality conditions.
Back to contents · Next: 20. Complete answers for the Marlow packet
20. Complete answers for the Marlow packet
Machine preparation needs 100 × 5 + 50 × 15 = 500 + 750 = 1,250 minutes. It fits within 1,500, leaving 250 machine minutes. Physical capacity of 160 exceeds the 150-order demand by ten. These checks eliminate preparation and physical output as binding constraints in the base packet, without claiming they would never bind in a different product mix or company.
Central review requires 150 × 5 = 750 basic minutes and 35 × 15 = 525 exception minutes, totalling 1,275. The shortfall against 1,100 is 175 minutes. Holding eight exceptions can remove at most 160, which is insufficient. Holding nine removes 180, leaving 1,095. Therefore the maximum count is 141 orders. One feasible schedule holds nine simple exceptions, completing ninety-one simple and fifty custom orders with five review minutes unused.
In the divisional arrangement, simple review needs 100 × 5 + 15 × 15 = 725 minutes against five hundred. Custom review needs 50 × 5 + 20 × 15 = 550 against six hundred. The simple team has a 225-minute deficit while the custom team has fifty unused minutes. Eleven held simple exceptions remove 220 minutes, still insufficient; twelve holds can remove at least 225. Holding eleven exceptions and one routine simple order removes exactly 225.
The simple team therefore completes eighty-eight orders within five hundred minutes, and the custom team completes fifty within 550. Total output is 138. At least twelve simple orders must be held because eleven orders cannot remove more than 220 minutes. The displayed schedule proves the bound can be reached. The lower output than pooled central review comes from the non-transferable division boundary, not a change in the total staff budget or model capability.
Under bounded routine release, fifteen simple exceptions plus all fifty custom orders require human release: sixty-five orders at five minutes give 325. The thirty-five exceptions add 525, and the ten audits add fifty. Total review is nine hundred minutes, leaving two hundred. All 150 orders fit the machine, physical, review and stated permission limits. The eighty-five routine releases are allowed because the packet grants standing authority and quality conditions, not because the arithmetic itself grants them.
Full revenue is 100 × 160 + 50 × 280 = 30,000 credits. Variable cost is 100 × 60 + 50 × 120 = 12,000. After 7,000 common fixed cost and 1,400 technology cost, surplus is 9,600. For comparison, the displayed central schedule’s nine missing simple orders reduce contribution by nine times one hundred, giving 8,700. The divisional schedule’s twelve missing simple orders give 8,400. Each financial difference reconciles to the identified missing outcomes.
In the changed workload, seventy routine simple orders remain eligible for standing release. Thirty simple exceptions plus all fifty custom orders require human release, giving eighty orders and four hundred basic minutes. Sixty exceptions add nine hundred minutes; ten audits still add fifty. Total is 1,350, exceeding 1,100 by 250. Additional qualified review capacity of 250 minutes would exactly meet the arithmetic requirement, but availability, authority and cost are not supplied, so no priced repair or revised surplus can be claimed.
The independent result is intentionally different from Corin’s stress repair. Marlow supplies a shortfall but no offer for extra review. A complete answer preserves that missing information rather than borrowing Corin’s rate or capacity. The technique transfers; the inputs do not. This is the central skill the exercise tests: carry the structure of reasoning to a new company while keeping its own facts, permissions, resources and unknowns intact.
Back to contents · Next: 21. A completed decision memo for Corin
21. A completed decision memo for Corin
Corin’s memo begins with a conditional conclusion. Under the supplied future capability, permission and workload assumptions, bounded routine release completes all 120 orders. Central approval completes at most 119, and the fixed divisional arrangement at most 114. The system’s machine preparation capacity is sufficient in all three. The difference comes from how human decisions are assigned and which routine releases are legitimately pre-authorised.
The memo does not conceal the financial result. Bounded release gives a 9,000-credit weekly surplus, above the two other future arrangements but below the 10,080-credit present-style reference schedule. It therefore supports a claim about complete service and available staff capacity, not an unconditional claim of higher profit. Any decision to pay for the future service would need to assess benefits not priced in the packet or change a separately stated assumption.
It also names the conditions that could defeat the preferred arrangement. The routine class must remain identifiable and appropriate for standing release. Custom and exceptional work must continue reaching authorised reviewers. Machine and physical resources must remain sufficient. Audits and correction work must be included rather than treated as free. If these conditions cannot be substantiated in a real company, the classroom result is not evidence that the same arrangement is suitable there.
For the heavier-exception week, the memo records 1,566 review minutes and a 366-minute shortfall. The supplied 420-minute repair restores complete service at an additional 630 credits, reducing surplus to 8,370. It does not describe the exception team as having failed because it raised more concerns. Correct recognition of a harder workload is useful information. The managerial decision concerns resources and commitments, not pressuring the system or staff to make the concerns disappear.
The make-or-buy comparison remains a separate decision. Internal provision becomes cheaper over a sufficiently long operating horizon under the stipulated equivalence, while buying is cheaper over twelve weeks. Corin should not use that component result to claim that the whole firm should vertically integrate or outsource everything. The boundary of each decision should follow the service, responsibilities and evidence that were actually compared.
Finally, the memo allocates further learning rather than assuming the analysis is finished forever. It calls for evidence about routine-release reliability, customer consequences of delay, the value of released preparation time and the feasibility of obtaining additional qualified review. Those are specific uncertainties that could change the decision. The company’s future organisation remains revisable, which is more useful than presenting one speculative chart as the final form of a superintelligent enterprise.
Back to contents · Next: 22. Questions that remain after the arithmetic works
22. Questions that remain after the arithmetic works
Would a superintelligent system eliminate managers?
The scenario does not establish that. It separates analytical preparation from authority, commitments, exception handling and accountability. Some coordination work might shrink or change, while other responsibilities remain or become more important. The number and form of managerial roles would depend on the company, its obligations and the capabilities actually demonstrated. A general prediction about the disappearance of management goes beyond the evidence provided here.
Why should people review work that a machine can do better?
Review can serve different purposes: verifying an output, making an authorised choice, resolving conflicting preferences or accepting responsibility for a commitment. Superior task performance may change which verification methods are useful, but it does not automatically transfer authority. A future arrangement should avoid meaningless rubber-stamping while preserving the decisions and safeguards that its consequences require. The appropriate method is a substantive assurance question, not a slogan about humans or machines being in charge.
Is a central company system always more efficient than specialised teams?
No. The central arrangement pools scarce review capacity in this packet, while divisions strand some of it. Another setting could give specialists a real advantage in interpretation or resolution time. A fair comparison would state those differences and recalculate the complete outcome. The case teaches how to test an organisational boundary, not that every company should erase all local teams or centralise every decision.
Do fewer approvals mean weaker governance?
Not necessarily. Repeated approvals can be redundant when a suitable standing rule is legitimate, well evidenced and maintained. They can also be essential where conditions or consequences demand individual judgment. Strong governance concerns the quality and accountability of the arrangement, not simply the number of approval clicks. Removing a required decision without authority is very different from deliberately establishing an appropriate routine path.
Why did the future arrangement earn less than the reference?
The fictional staff cost stayed fixed, the technology added a fee and the completed-order increase was small. Those inputs make the result possible even though preparation used much less human time. The example prevents released capacity from being mistaken for realised cash savings. A different outcome would require explicit changes in demand, cost, valuable redeployment or other consequences; it cannot be supplied by optimism alone.
What does the guide let a reader judge independently?
You should be able to state a company promise, separate capability from authority, identify which work consumes a particular review budget, test whether division boundaries strand capacity and reconcile accepted output with its costs. You should also be able to recognise a missing input instead of inventing it. Those abilities support careful discussion of hypothetical future companies while remaining useful for interpreting today’s claims about organisational change.
Back to contents · Next: 23. Continue with the adjacent question
23. Continue with the adjacent question
For the practical work of introducing current AI into an existing organisation, use the workplace Super Intelligence guide. Its present-day implementation question is distinct from this conditional comparison of future company arrangements.
For the bridge from faster tasks to useful output and economic measures, read Super Intelligence and productivity. For the separate question of who receives gains, continue with wealth, ownership and inequality.
For supplier dependence, data rights and credible alternatives outside the firm, read compute, data and capital. Return to What Is Super Intelligence? for the wider definitions and evidence framework.
Back to contents · Next: Sources and scope
Sources and scope
The NIST AI RMF1.0 Govern and human–AI interaction sections support the limited descriptions of organisational roles and responsibilities. The NIST2024 Generative Artificial Intelligence Profile supports the stated present-system risk context. The OECD2023 employer/worker survey report supports the bounded discussion of workplace research, training and consultation. None is evidence that an ASI firm currently exists or that the hypothetical arrangements have been deployed successfully.
Corin, Marlow, their products, customers, permissions, exception counts, capacities, prices, costs and results are original fictional teaching constructs. Model capability, correct exception identification and routine-output quality are assumptions, not measured AI performance. All arithmetic is conditional on the packet. Actual deployment, staffing, legal, contractual and financial decisions need their own evidence and appropriate professional judgment.
Back to contents · Return to the Super Intelligence master guide
