The value chain · what could be built

Businesses that could be built

The value chain page judges where each part's economics are heading, and names what would turn a passing shortage into a lasting profit. Each business below is one that could make such a turn happen: what it would sell first, what it would build up, why the largest AI companies would not simply give it away, and what would prove it wrong. Each is a business described, not a company named: the companies listed with it are the ones the tracker has already placed in the same part of the map, and some may be building something like it.

A few words recur. An agent is an AI program that carries out a task on its own, such as filing a claim or booking a purchase, rather than only answering questions. A lab is a company that builds the largest AI models, such as OpenAI or Anthropic, and a cloud is a company that rents out computing, such as Amazon or Microsoft; both can bundle extra services into what they already sell. A roll-up is a firm that buys many small businesses in one trade and runs them together, and a system of record is the software where a firm keeps the official copy of its work, such as a law firm's case files.

A rent is profit beyond what it takes to keep a business going, earned because something holds rivals off. The profit shown on each business comes from the rent rule the futures section uses. Who keeps the profit follows David Teece, an economist who studied why the firm that invents something often fails to keep the returns, and found that when an invention is easy to copy, whoever owns what customers need to use it collects instead. How large it is follows the kind of rent and how long it lasts, from none through thin, moderate and fat to monopoly-like; moderate is set where William Nordhaus, an economist who measured how much of an invention's value its maker keeps, found it usually sits: a small share. The answers about each case are the owner's judgement; the result follows the rule. Some businesses keep no lasting profit at all; they are listed anyway, because knowing what will not hold is as useful as knowing what will.

What the powers mean

Hamilton Helmer, a strategy consultant and investor who studied why some firms keep earning more than their investors could get elsewhere, found that an advantage lasts only when a benefit sits behind a barrier that stops rivals copying it, and named the kinds that recur:

a cornered resource: preferential access to something others need: a patent, a licence, a scarce input, a rare team
branding: buyers pay more for the same thing because of what they believe about who made it
network economies: each user makes the product more valuable to the others, so the leader's lead feeds itself
process power: a way of working that improves cost or quality and takes a rival years to learn
scale economies: unit costs fall as volume grows, and a rival cannot match them without matching the volume
switching costs: a customer would lose money, time or work by moving to a rival

For a call on each of these, and what each future on the outlook's grid would do to it, see what to build, future by future.

How a passing shortage becomes a lasting profit, and the two ways it failsA model, not a measurement

The problemA shortageSomething the chain is short of today. Alone it passes: once relieved, the shortage moves elsewhere.Nobody can cheaply prove that an agent finished a task correctly, so buyers keep humans checking every result

First stepOne narrow thing to sellWhat the business sells first, to the buyers who feel the shortage most.Check one high-volume workflow end to end and bill only for results that pass

What it builds upSomething rivals cannot quickly copyA record, a standing or a habit that grows with each customer and still holds rivals off after the shortage eases.The definition of done for each workflow, and a record of how often agents pass and fail that grows more useful with every buyer

Why that profitA rentProfit beyond what keeps the business going, sized by the rent rule drawn below.A fat profit, kept by the maker. The failure record grows more useful with each buyer, but the checking method itself is easy to copy

Why a lab would not just bundle itFails if a larger firm gives it awayA lab or a cloud can fold a service into what it already sells. Each card says why this one would be left alone.A buyer wants the checker to be independent of whoever sold the agent

Wrong ifFails if the shortage ends another wayEach card names the event that would undo it, so the reader can watch for it.Agents become cheap and reliable at checking their own work, so buyers stop paying for an outside check

a step every card describes, under the card's own labelwhat stops the turnwordsthe worked example, in its record's own words

A drawing of the argument each card makes, not a measurement. The example is Outcome verification with pay-per-verified-result billing, the opening record in the list; every other card fills the same boxes. It does not show how likely any step is: nothing on this page is scored.

21 businesses · 0 in categories the tracker has not mapped yet · profit by the rent rule: 3 monopoly-like · 6 fat · 8 moderate · 1 thin · 3 no lasting profit · 4 kept by incumbent firms

The rent rule, drawn

The rule answers in turn who keeps a profit and how large it is. Each business is drawn as its number in the list, shown here and on its card, boxed by who would keep its profit. The number is only its place in the list: nothing here is ranked or scored.

1 Outcome verification with pay-per-verified-result billing
2 Independent sign-off for AI in regulated work
3 A model that keeps learning inside one customer's work
4 Know-your-agent and delegated-authority registry
5 Tamper-proof audit trail for what agents did
6 AI warranty backed by a pool of loss data
7 Graded expert-work records pooled across buyers
8 A router that chooses models from pooled outcomes
9 Escrow that pays an agent only when its work passes a check
10 The standard way agents get permission into business software
11 Deployment that leaves behind the platform the customer runs on
12 Capturing how a firm really works, for its agents
13 A trade's AI system of record, with the approvals only that trade can give
14 An AI roll-up that moves its clients to fixed fees
15 A licensed firm that sells the finished job and stands behind it
16 Moving computing to where power is available
17 Training and placing the engineers who deploy AI
18 Routing that keeps a firm inside approved models and fails over when one is cut off
19 Tuning open models to frontier quality for one industry
20 Incident response and forensics for agent failures
21 A firm's own map of what its data means, kept on its own servers for every agent to read

Who keeps the profit: the rule asks its questions in order, and the earliest that fits decidesA model of the rule; each placing is the owner's judgement

Can the maker stop others copying it, by patent, secrecy or know-how that is hard to learn?
Kept by the maker2 6
If not: does the maker itself own what customers need in order to use it, something only a few hold?
Kept by the maker1 3 5 7 11 12 13 14 15 18 20 21
If not: does someone else own that?
Kept by incumbent firms4 9 10 16
Otherwise: it is easy to copy and nothing scarce stands in the way.
Competed away to users: no lasting profit8 17 19

the maker would keep the profit (the owner's judgement, by the rent rule)someone other than the maker would keep it (here, incumbent firms)it would be competed away to usersnits place in the list, not a rank; select it to read the card

The questions are the rent rule's, after David Teece, read top to bottom; the answers given for each business are the owner's judgement, and the result follows the rule. It shows who would keep a profit, not how large it is (the next figure) or whether the business gets built.

The numbers

Business	Can the maker stop copying?	What customers need to use it	Held by	Result
1. Outcome verification with pay-per-verified-result billing	no	something only a few hold	the maker	A fat profit, kept by the maker
2. Independent sign-off for AI in regulated work	yes	something only a few hold	the maker	A monopoly-like profit, kept by the maker
3. A model that keeps learning inside one customer's work	no	something only a few hold	the maker	A moderate profit, kept by the maker
4. Know-your-agent and delegated-authority registry	no	something only a few hold	incumbent firms	A monopoly-like profit, kept by incumbent firms
5. Tamper-proof audit trail for what agents did	no	something only a few hold	the maker	A fat profit, kept by the maker
6. AI warranty backed by a pool of loss data	yes	something only a few hold	the maker	A fat profit, kept by the maker
7. Graded expert-work records pooled across buyers	no	something only a few hold	the maker	A moderate profit, kept by the maker
8. A router that chooses models from pooled outcomes	no	nothing scarce	platforms	Competed away to users
9. Escrow that pays an agent only when its work passes a check	no	something only a few hold	incumbent firms	A monopoly-like profit, kept by incumbent firms
10. The standard way agents get permission into business software	no	something only a few hold	incumbent firms	A moderate profit, kept by incumbent firms
11. Deployment that leaves behind the platform the customer runs on	no	something only a few hold	the maker	A moderate profit, kept by the maker
12. Capturing how a firm really works, for its agents	no	something only a few hold	the maker	A thin profit, kept by the maker
13. A trade's AI system of record, with the approvals only that trade can give	no	something only a few hold	the maker	A fat profit, kept by the maker
14. An AI roll-up that moves its clients to fixed fees	no	something only a few hold	the maker	A moderate profit, kept by the maker
15. A licensed firm that sells the finished job and stands behind it	no	something only a few hold	the maker	A fat profit, kept by the maker
16. Moving computing to where power is available	no	something only a few hold	incumbent firms	A moderate profit, kept by incumbent firms
17. Training and placing the engineers who deploy AI	no	nothing scarce	the maker	Competed away to users
18. Routing that keeps a firm inside approved models and fails over when one is cut off	no	something only a few hold	the maker	A fat profit, kept by the maker
19. Tuning open models to frontier quality for one industry	no	nothing scarce	the maker	Competed away to users
20. Incident response and forensics for agent failures	no	something only a few hold	the maker	A moderate profit, kept by the maker
21. A firm's own map of what its data means, kept on its own servers for every agent to read	no	something only a few hold	the maker	A moderate profit, kept by the maker

How large the profit is: what holds rivals off, against how long it lastsA model of the rule; each placing is the owner's judgement

Which number is which business

The number is only its place in the list, not a rank.

1 Outcome verification with pay-per-verified-result billing
2 Independent sign-off for AI in regulated work
3 A model that keeps learning inside one customer's work
4 Know-your-agent and delegated-authority registry
5 Tamper-proof audit trail for what agents did
6 AI warranty backed by a pool of loss data
7 Graded expert-work records pooled across buyers
8 A router that chooses models from pooled outcomes
9 Escrow that pays an agent only when its work passes a check
10 The standard way agents get permission into business software
11 Deployment that leaves behind the platform the customer runs on
12 Capturing how a firm really works, for its agents
13 A trade's AI system of record, with the approvals only that trade can give
14 An AI roll-up that moves its clients to fixed fees
15 A licensed firm that sells the finished job and stands behind it
16 Moving computing to where power is available
17 Training and placing the engineers who deploy AI
18 Routing that keeps a firm inside approved models and fails over when one is cut off
19 Tuning open models to frontier quality for one industry
20 Incident response and forensics for agent failures
21 A firm's own map of what its data means, kept on its own servers for every agent to read

shortmediumlong

Nothing holds rivals off

no lasting profit

Scarcity: it holds something that cannot be expanded

thincompeted away: keeps none of this17 19

moderate7 14 16 20

fat6

Switching costs: leaving would cost the customer

thin12

moderate3 10 11 21

fat5 13

Scale or network: it gets better or cheaper as it grows

moderate

fat1competed away: keeps none of this8

monopoly-like4 9

Regulation: a rule or a licence keeps others out

moderate

fat15 18

monopoly-like2

how long before the shortage moves →

thinmoderatefatmonopoly-likethe maker would keep the profit (the owner's judgement, by the rent rule)someone other than the maker would keep it (here, incumbent firms)it would be competed away to usersnits place in the list, not a rank; select it to read the card

Each cell is the size the rent rule gives, darker for larger; the cells are the rule's and are categories, not measurements. Which cell a business sits in is the owner's judgement. A dashed mark, set apart beneath a line, is a business the first figure sends to users: it would sit in that cell if anyone kept its profit, and keeps none of it. Short means the shortage is expected to move within a few years, medium within a decade or two, long only after decades. The rule is described in the method.

The numbers

Business	What holds rivals off	How long	Result
1. Outcome verification with pay-per-verified-result billing	Scale or network	medium	A fat profit, kept by the maker
2. Independent sign-off for AI in regulated work	Regulation	long	A monopoly-like profit, kept by the maker
3. A model that keeps learning inside one customer's work	Switching costs	medium	A moderate profit, kept by the maker
4. Know-your-agent and delegated-authority registry	Scale or network	long	A monopoly-like profit, kept by incumbent firms
5. Tamper-proof audit trail for what agents did	Switching costs	long	A fat profit, kept by the maker
6. AI warranty backed by a pool of loss data	Scarcity	long	A fat profit, kept by the maker
7. Graded expert-work records pooled across buyers	Scarcity	medium	A moderate profit, kept by the maker
8. A router that chooses models from pooled outcomes	Scale or network	medium	Competed away to users
9. Escrow that pays an agent only when its work passes a check	Scale or network	long	A monopoly-like profit, kept by incumbent firms
10. The standard way agents get permission into business software	Switching costs	medium	A moderate profit, kept by incumbent firms
11. Deployment that leaves behind the platform the customer runs on	Switching costs	medium	A moderate profit, kept by the maker
12. Capturing how a firm really works, for its agents	Switching costs	short	A thin profit, kept by the maker
13. A trade's AI system of record, with the approvals only that trade can give	Switching costs	long	A fat profit, kept by the maker
14. An AI roll-up that moves its clients to fixed fees	Scarcity	medium	A moderate profit, kept by the maker
15. A licensed firm that sells the finished job and stands behind it	Regulation	medium	A fat profit, kept by the maker
16. Moving computing to where power is available	Scarcity	medium	A moderate profit, kept by incumbent firms
17. Training and placing the engineers who deploy AI	Scarcity	short	Competed away to users
18. Routing that keeps a firm inside approved models and fails over when one is cut off	Regulation	medium	A fat profit, kept by the maker
19. Tuning open models to frontier quality for one industry	Scarcity	short	Competed away to users
20. Incident response and forensics for agent failures	Scarcity	medium	A moderate profit, kept by the maker
21. A firm's own map of what its data means, kept on its own servers for every agent to read	Switching costs	medium	A moderate profit, kept by the maker

A sequence of bets

each stage opens only once its gate is passed

NowIndependent sign-off for AI in regulated workOpens the next stage when: One regulator or insurer recognises the tests
NextA model that keeps learning inside one customer's workOpens the next stage when: Customers grant rights to their work traces
LaterOutcome verification with pay-per-verified-result billingOpens the next stage when: Enough verified outcomes to price a warranty
EndgameA licensed firm that sells the finished job and stands behind itNeeds: Capital and an insurer to carry the risk

The businesses at a glance

Where the businesses sit on the chain and how many companies the map already holds in each part, how many kinds of firm are marked with a need each would meet and, folded beneath, what each would rely on.

Where each business would sit on the value chain, and how many companies the map already holds thereChart of each record's own placing

Which number is which business

The number is only its place in the list, not a rank.

1 Outcome verification with pay-per-verified-result billing
2 Independent sign-off for AI in regulated work
3 A model that keeps learning inside one customer's work
4 Know-your-agent and delegated-authority registry
5 Tamper-proof audit trail for what agents did
6 AI warranty backed by a pool of loss data
7 Graded expert-work records pooled across buyers
8 A router that chooses models from pooled outcomes
9 Escrow that pays an agent only when its work passes a check
10 The standard way agents get permission into business software
11 Deployment that leaves behind the platform the customer runs on
12 Capturing how a firm really works, for its agents
13 A trade's AI system of record, with the approvals only that trade can give
14 An AI roll-up that moves its clients to fixed fees
15 A licensed firm that sells the finished job and stands behind it
16 Moving computing to where power is available
17 Training and placing the engineers who deploy AI
18 Routing that keeps a firm inside approved models and fails over when one is cut off
19 Tuning open models to frontier quality for one industry
20 Incident response and forensics for agent failures
21 A firm's own map of what its data means, kept on its own servers for every agent to read

1 Physical inputs and compute supply1 sit here

1.1 Power and sites41 on the map16

1.2 Data-centre infrastructure19 on the map

1.3 Semiconductor supply chain30 on the map

1.4 Memory, storage and networking18 on the map

1.5 Compute providers32 on the map16

1.6 Compute utilization and control10 on the map16 8

2 Models and capability productionnone sit here

2.1 General-purpose models37 on the map19

2.2 Modality-native models43 on the map

2.3 Specialized models26 on the map

2.4 Model compression and adaptation3 on the map

3 Data, training and verification inputs4 sit here

3.1 Data origination16 on the map7

3.2 Data rights and exchange12 on the map

3.3 Data transformation12 on the map3

3.4 Synthetic data8 on the map

3.5 Training environments10 on the map

3.6 Reward and verification systems7 on the map1 7

3.7 Post-training and learning infrastructure20 on the map3 19

3.8 Evals and assurance17 on the map2 3 5 6 7 19

4 Runtime, serving and orchestration4 sit here

4.1 Inference supply29 on the map18 19

4.2 Inference optimization10 on the map8 16

4.3 Routing and gateways13 on the map8 18 9

4.4 Context and memory21 on the map21 10 12

4.5 Agent runtimes14 on the map10

4.6 Tools and actuation13 on the map10 4 21

5 Trust, control and transaction infrastructure6 sit here

5.1 Security16 on the map20 5

5.2 Identity and authority11 on the map4 5 9 10 18

5.3 Provenance and audit9 on the map5 2 4 6 13 18 20

5.4 Outcome verification8 on the map1 6 8 9 15

5.5 Settlement and billing5 on the map9 1 4

5.6 Liability and risk transfer4 on the map6 2 13 15 20

6 Deployment and organizational diffusion3 sit here

6.1 Technical implementation30 on the map11 12 17 21

6.2 Context engineering6 on the map12 3 11 17 21

6.3 Managed AI operations1 on the map1 11 12 17

6.4 Adoption and change8 on the map11

6.5 Talent production40 on the map17

7 Applications and operating models3 sit here

7.1 Horizontal copilots26 on the map

7.2 Functional enterprise applications20 on the map

7.3 Vertical workflow software58 on the map13 14 15

7.4 Autonomous services89 on the map15 13 14

7.5 Sensor–model–actuator networksnot yet mapped

7.6 AI-native rollups106 on the map14

7.7 Consumer AI12 on the map

8 Adoption and economic outcomesnone sit here

8 Outcomes (indicator-only)not yet mapped

na boxed number sits in this partna plain number touches this part without sitting in itthe maker would keep the profit (the owner's judgement, by the rent rule)someone other than the maker would keep it (here, incumbent firms)it would be competed away to usersn on the mapcompanies the market map already lists in that parta part no business here sits in or touches

Every part of the market map, in the map's order from physical inputs to applications, with each business in the part its record names and a plain number wherever it touches another. Where a business sits is the judgement in its record. An empty part means no business was written for it, not that none could be. The small count in each part is the number of companies the map lists there, and the part's name links to that list: these are companies in the same part of the map, not companies building this business, and a small count may mark where the tracker's coverage stops, not an empty market.

The numbers

Part of the chain	Sits here	Touches	Companies on the map	Independent	Bought, being bought or closed
1.1 Power and sites	none	Moving computing to where power is available	41	41	0
1.2 Data-centre infrastructure	none	none	19	17	2
1.3 Semiconductor supply chain	none	none	30	28	2
1.4 Memory, storage and networking	none	none	18	17	1
1.5 Compute providers	none	Moving computing to where power is available	32	32	0
1.6 Compute utilization and control	Moving computing to where power is available	A router that chooses models from pooled outcomes	10	9	1
2.1 General-purpose models	none	Tuning open models to frontier quality for one industry	37	37	0
2.2 Modality-native models	none	none	43	41	2
2.3 Specialized models	none	none	26	24	2
2.4 Model compression and adaptation	none	none	3	3	0
3.1 Data origination	Graded expert-work records pooled across buyers	none	16	16	0
3.2 Data rights and exchange	none	none	12	12	0
3.3 Data transformation	none	A model that keeps learning inside one customer's work	12	12	0
3.4 Synthetic data	none	none	8	8	0
3.5 Training environments	none	none	10	10	0
3.6 Reward and verification systems	none	Outcome verification with pay-per-verified-result billing; Graded expert-work records pooled across buyers	7	6	1
3.7 Post-training and learning infrastructure	A model that keeps learning inside one customer's work; Tuning open models to frontier quality for one industry	none	20	19	1
3.8 Evals and assurance	Independent sign-off for AI in regulated work	A model that keeps learning inside one customer's work; Tamper-proof audit trail for what agents did; AI warranty backed by a pool of loss data; Graded expert-work records pooled across buyers; Tuning open models to frontier quality for one industry	17	16	1
4.1 Inference supply	none	Routing that keeps a firm inside approved models and fails over when one is cut off; Tuning open models to frontier quality for one industry	29	27	2
4.2 Inference optimization	none	A router that chooses models from pooled outcomes; Moving computing to where power is available	10	10	0
4.3 Routing and gateways	A router that chooses models from pooled outcomes; Routing that keeps a firm inside approved models and fails over when one is cut off	Escrow that pays an agent only when its work passes a check	13	13	0
4.4 Context and memory	A firm's own map of what its data means, kept on its own servers for every agent to read	The standard way agents get permission into business software; Capturing how a firm really works, for its agents	21	21	0
4.5 Agent runtimes	none	The standard way agents get permission into business software	14	14	0
4.6 Tools and actuation	The standard way agents get permission into business software	Know-your-agent and delegated-authority registry; A firm's own map of what its data means, kept on its own servers for every agent to read	13	13	0
5.1 Security	Incident response and forensics for agent failures	Tamper-proof audit trail for what agents did	16	14	2
5.2 Identity and authority	Know-your-agent and delegated-authority registry	Tamper-proof audit trail for what agents did; Escrow that pays an agent only when its work passes a check; The standard way agents get permission into business software; Routing that keeps a firm inside approved models and fails over when one is cut off	11	10	1
5.3 Provenance and audit	Tamper-proof audit trail for what agents did	Independent sign-off for AI in regulated work; Know-your-agent and delegated-authority registry; AI warranty backed by a pool of loss data; A trade's AI system of record, with the approvals only that trade can give; Routing that keeps a firm inside approved models and fails over when one is cut off; Incident response and forensics for agent failures	9	9	0
5.4 Outcome verification	Outcome verification with pay-per-verified-result billing	AI warranty backed by a pool of loss data; A router that chooses models from pooled outcomes; Escrow that pays an agent only when its work passes a check; A licensed firm that sells the finished job and stands behind it	8	8	0
5.5 Settlement and billing	Escrow that pays an agent only when its work passes a check	Outcome verification with pay-per-verified-result billing; Know-your-agent and delegated-authority registry	5	5	0
5.6 Liability and risk transfer	AI warranty backed by a pool of loss data	Independent sign-off for AI in regulated work; A trade's AI system of record, with the approvals only that trade can give; A licensed firm that sells the finished job and stands behind it; Incident response and forensics for agent failures	4	4	0
6.1 Technical implementation	Deployment that leaves behind the platform the customer runs on	Capturing how a firm really works, for its agents; Training and placing the engineers who deploy AI; A firm's own map of what its data means, kept on its own servers for every agent to read	30	30	0
6.2 Context engineering	Capturing how a firm really works, for its agents	A model that keeps learning inside one customer's work; Deployment that leaves behind the platform the customer runs on; Training and placing the engineers who deploy AI; A firm's own map of what its data means, kept on its own servers for every agent to read	6	6	0
6.3 Managed AI operations	none	Outcome verification with pay-per-verified-result billing; Deployment that leaves behind the platform the customer runs on; Capturing how a firm really works, for its agents; Training and placing the engineers who deploy AI	1	1	0
6.4 Adoption and change	none	Deployment that leaves behind the platform the customer runs on	8	8	0
6.5 Talent production	Training and placing the engineers who deploy AI	none	40	40	0
7.1 Horizontal copilots	none	none	26	24	2
7.2 Functional enterprise applications	none	none	20	20	0
7.3 Vertical workflow software	A trade's AI system of record, with the approvals only that trade can give	An AI roll-up that moves its clients to fixed fees; A licensed firm that sells the finished job and stands behind it	58	58	0
7.4 Autonomous services	A licensed firm that sells the finished job and stands behind it	A trade's AI system of record, with the approvals only that trade can give; An AI roll-up that moves its clients to fixed fees	89	87	2
7.5 Sensor–model–actuator networks	none	none	0	0	0
7.6 AI-native rollups	An AI roll-up that moves its clients to fixed fees	none	106	106	0
7.7 Consumer AI	none	none	12	12	0
8 Outcomes (indicator-only)	none	none	0	0	0

How many kinds of firm are marked with a need each business would meetChart of this site's judgement

1 Outcome verification with pay-per-verified-result billingA check for work that has none

12 of 12

9 Escrow that pays an agent only when its work passes a checkIdentity, permissions and spending limits for agents; Selling the result, and settling it

7 of 12

2 Independent sign-off for AI in regulated workSign-off for regulated work

5 of 12

5 Tamper-proof audit trail for what agents didA record of what the agent did, and who approved it

5 of 12

6 AI warranty backed by a pool of loss dataInsurance priced for an agent's mistakes

5 of 12

15 A licensed firm that sells the finished job and stands behind itSign-off for regulated work

5 of 12

3 A model that keeps learning inside one customer's workContext that stays with one customer

4 of 12

4 Know-your-agent and delegated-authority registryIdentity, permissions and spending limits for agents

4 of 12

10 The standard way agents get permission into business softwareIdentity, permissions and spending limits for agents

4 of 12

11 Deployment that leaves behind the platform the customer runs onHelp putting it to work

4 of 12

14 An AI roll-up that moves its clients to fixed feesSelling the result, and settling it

4 of 12

17 Training and placing the engineers who deploy AIHelp putting it to work

4 of 12

13 A trade's AI system of record, with the approvals only that trade can giveThe trade's system of record

3 of 12

12 Capturing how a firm really works, for its agentsWriting down how the firm works

1 of 12

20 Incident response and forensics for agent failuresSomeone to call when an agent goes wrong

1 of 12

21 A firm's own map of what its data means, kept on its own servers for every agent to readWriting down how the firm works

1 of 12

Not on that page's grid

7 Graded expert-work records pooled across buyers8 A router that chooses models from pooled outcomes16 Moving computing to where power is available18 Routing that keeps a firm inside approved models and fails over when one is cut off19 Tuning open models to frontier quality for one industry

No need on that page names these businesses. That is a gap in the grid, not a finding that nobody would need them: some sell to labs, clouds and data-centre operators, which that page does not describe.

Needs on that page that no business here meets

Training people when the junior work is gone

4 of 12

People on call

2 of 12

Running autonomy in stages

2 of 12

Running robots and driverless fleets

3 of 12

kinds of firm marked, on the kinds-of-firm page, with a need this business would meet (a judgement)kinds of firm with a need that no business here meetsthe whole bar is every kind of firm that page describes, not every kind there isthe maker would keep the profit (the owner's judgement, by the rent rule)someone other than the maker would keep it (here, incumbent firms)it would be competed away to usersnits place in the list, not a rank; select it to read the card

A count of the kinds of firm on what happens to each kind of firm that are marked with a need this business would meet; a kind is counted once however many of its needs point here. Where a need names more than one business, each is given every kind with that need, so businesses that share a need share a bar and the bars cannot be added. Breadth of need is not size of profit: a business may be needed by many kinds of firm and still keep nothing, as the dashed marks show. It counts kinds of firm, not firms, customers or money. The needs, the marks and the link from a need to a business were drafted by a model for that page in October 2026 and are that page's judgement, so the bars are hatched; the owner reviewed them in October 2026, after his review of the records below.

The numbers

Business or unmet need	Kinds of firm	Which	Through the need
1. Outcome verification with pay-per-verified-result billing	12	Accounting practice; Law firm; Software and IT services company; Bank or lender; Insurer or insurance broker; Hospital or medical practice; Customer-support and business-support firm; Management consultancy; Property manager or letting agency; Freight and logistics operator; Manufacturer; Retailer	A check for work that has none
9. Escrow that pays an agent only when its work passes a check	7	Accounting practice; Law firm; Software and IT services company; Bank or lender; Customer-support and business-support firm; Property manager or letting agency; Retailer	Identity, permissions and spending limits for agents; Selling the result, and settling it
2. Independent sign-off for AI in regulated work	5	Accounting practice; Law firm; Bank or lender; Insurer or insurance broker; Hospital or medical practice	Sign-off for regulated work
5. Tamper-proof audit trail for what agents did	5	Accounting practice; Law firm; Bank or lender; Insurer or insurance broker; Hospital or medical practice	A record of what the agent did, and who approved it
6. AI warranty backed by a pool of loss data	5	Law firm; Insurer or insurance broker; Hospital or medical practice; Management consultancy; Manufacturer	Insurance priced for an agent's mistakes
15. A licensed firm that sells the finished job and stands behind it	5	Accounting practice; Law firm; Bank or lender; Insurer or insurance broker; Hospital or medical practice	Sign-off for regulated work
3. A model that keeps learning inside one customer's work	4	Accounting practice; Law firm; Bank or lender; Customer-support and business-support firm	Context that stays with one customer
4. Know-your-agent and delegated-authority registry	4	Software and IT services company; Bank or lender; Property manager or letting agency; Retailer	Identity, permissions and spending limits for agents
10. The standard way agents get permission into business software	4	Software and IT services company; Bank or lender; Property manager or letting agency; Retailer	Identity, permissions and spending limits for agents
11. Deployment that leaves behind the platform the customer runs on	4	Accounting practice; Software and IT services company; Customer-support and business-support firm; Property manager or letting agency	Help putting it to work
14. An AI roll-up that moves its clients to fixed fees	4	Accounting practice; Law firm; Customer-support and business-support firm; Property manager or letting agency	Selling the result, and settling it
17. Training and placing the engineers who deploy AI	4	Accounting practice; Software and IT services company; Customer-support and business-support firm; Property manager or letting agency	Help putting it to work
13. A trade's AI system of record, with the approvals only that trade can give	3	Accounting practice; Property manager or letting agency; Freight and logistics operator	The trade's system of record
12. Capturing how a firm really works, for its agents	1	Management consultancy	Writing down how the firm works
20. Incident response and forensics for agent failures	1	Retailer	Someone to call when an agent goes wrong
21. A firm's own map of what its data means, kept on its own servers for every agent to read	1	Management consultancy	Writing down how the firm works
7. Graded expert-work records pooled across buyers		not on the grid	not on the grid
8. A router that chooses models from pooled outcomes		not on the grid	not on the grid
16. Moving computing to where power is available		not on the grid	not on the grid
18. Routing that keeps a firm inside approved models and fails over when one is cut off		not on the grid	not on the grid
19. Tuning open models to frontier quality for one industry		not on the grid	not on the grid
No business listed: Training people when the junior work is gone	4		Training people when the junior work is gone
No business listed: People on call	2		People on call
No business listed: Running autonomy in stages	2		Running autonomy in stages
No business listed: Running robots and driverless fleets	3		Running robots and driverless fleets

What each business would rely on to hold rivals off: the owner's judgement, as a table

Business	scale economies	network economies	switching costs	branding	a cornered resource	process power
1 Outcome verification with pay-per-verified-result billing
2 Independent sign-off for AI in regulated work
3 A model that keeps learning inside one customer's work
4 Know-your-agent and delegated-authority registry
5 Tamper-proof audit trail for what agents did
6 AI warranty backed by a pool of loss data
7 Graded expert-work records pooled across buyers
8 A router that chooses models from pooled outcomes
9 Escrow that pays an agent only when its work passes a check
10 The standard way agents get permission into business software
11 Deployment that leaves behind the platform the customer runs on
12 Capturing how a firm really works, for its agents
13 A trade's AI system of record, with the approvals only that trade can give
14 An AI roll-up that moves its clients to fixed fees
15 A licensed firm that sells the finished job and stands behind it
16 Moving computing to where power is available
17 Training and placing the engineers who deploy AI
18 Routing that keeps a firm inside approved models and fails over when one is cut off
19 Tuning open models to frontier quality for one industry
20 Incident response and forensics for agent failures
21 A firm's own map of what its data means, kept on its own servers for every agent to read
Businesses naming it	1	5	9	2	4	2

the owner judges the business would build this powerit does not

scale economies: unit costs fall as volume grows, and a rival cannot match them without matching the volume
network economies: each user makes the product more valuable to the others, so the leader's lead feeds itself
switching costs: a customer would lose money, time or work by moving to a rival
branding: buyers pay more for the same thing because of what they believe about who made it
a cornered resource: preferential access to something others need: a patent, a licence, a scarce input, a rare team
process power: a way of working that improves cost or quality and takes a rival years to learn

The powers are Hamilton Helmer's; which ones a business would build is the judgement in its record, in the list's order. The bottom row counts the businesses naming each. No business here relies on counter-positioning. A mark says what the business would rely on, not that it would succeed. These powers and the rent rule's account of what holds rivals off are separate judgements under separate schemes and need not match: a business can sit in the rule's regulation row and name switching costs here.

The businesses

1Outcome verification with pay-per-verified-result billing
A fat profit, kept by the maker
Sits in 5.4 Outcome verification; touches Settlement and billing, Reward and verification systems, Managed AI operations
The problem
Nobody can cheaply prove that an agent finished a task correctly, so buyers keep humans checking every result
Sold to
Operations leaders who pay for agent work and must show it was done right
First step
Check one high-volume workflow end to end and bill only for results that pass
Charges for
A verified outcome
Why a lab would not just bundle it
A buyer wants the checker to be independent of whoever sold the agent
What it builds up
The definition of done for each workflow, and a record of how often agents pass and fail that grows more useful with every buyer (network economies)
Why that profit
The failure record grows more useful with each buyer, but the checking method itself is easy to copy
Wrong if
Agents become cheap and reliable at checking their own work, so buyers stop paying for an outside check
Already on the map
Bluejay, Cekura, Coval, Hamming AI, Judgment Labs, Raindrop and 2 more
Builds on the site's reading
- Testing AI is being folded into the labs' and clouds' platforms; an independent tester keeps a rent only where buyers need a judge who is not the seller
- Checking AI's work is sold as a category of its own.
Needs first
A buyer who already pays for agent work by the task, and a workflow whose finished result can be checked against records the buyer holds
2Independent sign-off for AI in regulated work
A monopoly-like profit, kept by the maker
Sits in 3.8 Evals and assurance; touches Liability and risk transfer, Provenance and audit
The problem
Regulated firms cannot use a model until someone other than the seller vouches for it
Sold to
Banks, insurers, hospitals and their regulators
First step
Private test suites for one regulated task, such as claims handling or credit decisions, accepted by one regulator or insurer
Charges for
A certification per model and task, renewed on each model change
Why a lab would not just bundle it
The whole value is that the tester does not sell the model
What it builds up
A name buyers and regulators insist on, and, if outside sign-off becomes required, a place rivals can reach only by earning the same standing (branding)
Why that profit
If a regulator names the tester, rivals need the same standing, which takes years
Wrong if
Regulators accept the labs' own testing, or no regulator requires an outside sign-off
Already on the map
Arize, Goodfire, Patronus, Braintrust, Confident AI, Galileo and 10 more
Builds on the site's reading
- Testing AI is being folded into the labs' and clouds' platforms; an independent tester keeps a rent only where buyers need a judge who is not the seller
- Checking AI's work is sold as a category of its own.
Needs first
One regulator or insurer willing to recognise the tests
3A model that keeps learning inside one customer's work
A moderate profit, kept by the maker
Sits in 3.7 Post-training and learning infrastructure; touches Data transformation, Evals and assurance, Context engineering
The problem
Everyday work leaves no reusable signal, so models stop improving once deployed
Sold to
Firms with a high volume of repeated expert work, such as claims, support or underwriting
First step
Turn live work into privacy-safe traces and a test set that refreshes itself
Charges for
A monthly fee per workflow that the model learns
Why a lab would not just bundle it
The customer's record helps only that customer, so a lab has little reason to build it for each one
What it builds up
A model that has learned the firm's work for months, which the firm would lose by moving to a rival (switching costs)
Why that profit
The lock-in is real once learning works, but the site reads tuning as commoditising until it does
Wrong if
Each new general model wipes out what the customer's model learned, so firms stop paying to keep it learning
Already on the map
Baseten, Prime Intellect, Adaption Labs, Adaptive ML, Applied Compute, Arcee AI and 13 more
Builds on the site's reading
- Tuning a model for one customer is becoming a feature of the clouds and the serving firms; a rent survives only where the tuned model keeps learning inside the customer's work
Needs first
Continual learning working in production
4Know-your-agent and delegated-authority registry
A monopoly-like profit, kept by incumbent firms
Sits in 5.2 Identity and authority; touches Provenance and audit, Settlement and billing, Tools and actuation
The problem
A counterparty cannot tell which agent acts for whom, and under what limits
Sold to
Platforms and firms that let outside agents act for people
First step
Issue and revoke agent credentials for one platform's partners
Charges for
An active agent identity per month
Why a lab would not just bundle it
Each lab issuing its own identities leaves no neutral register that counterparties across labs trust
What it builds up
A register that becomes more useful to everyone checking it as more agents are listed (network economies)
Why that profit
Existing identity firms already own the customer relationships and are extending to agents
Wrong if
Agents keep acting inside human sessions, so no separate agent identity is needed
Already on the map
Arcade, Delinea, Keeper, Keyfactor, Linx, Ory and 4 more
Builds on the site's reading
- Agent plumbing is mostly open source that labs and clouds buy or copy; the rent sits with whoever holds the connections into customers' systems
Needs first
One large platform that will require the credential from the agents it admits, and a way to tie each agent to the person or firm answerable for it
5Tamper-proof audit trail for what agents did
A fat profit, kept by the maker
Sits in 5.3 Provenance and audit; touches Identity and authority, Security, Evals and assurance
The problem
After an agent acts, nobody can reconstruct what it saw, decided and changed
Sold to
Compliance teams and auditors at firms that let agents act
First step
A decision log for one regulated workflow that auditors accept
Charges for
Logged agent actions, priced per seat of the audited team
Why a lab would not just bundle it
The log is only credible if the agent's maker cannot edit it
What it builds up
Years of audit history that cannot be moved without breaking the trail (switching costs)
Why that profit
Years of audit history are costly to leave, though the logging itself is easy to copy
Wrong if
Regulators accept each vendor's own logs
Already on the map
Credo AI, EQTY Lab, Monitaur, OpenBox AI, OpenOrigins, Saidot and 3 more
Builds on the site's reading
- Agent plumbing is mostly open source that labs and clouds buy or copy; the rent sits with whoever holds the connections into customers' systems
Needs first
Agents that act through connections which can be logged, and an auditor or regulator willing to say what record it will accept
6AI warranty backed by a pool of loss data
A fat profit, kept by the maker
Sits in 5.6 Liability and risk transfer; touches Outcome verification, Evals and assurance, Provenance and audit
The problem
Agent failures cannot be priced, so procurement stalls or insurers exclude them
Sold to
Firms buying agents for consequential work, and the insurers who cover them
First step
A warranty on one agent product, paid for by the vendor and backed by a reinsurer
Charges for
A premium per covered workflow
Why a lab would not just bundle it
A lab standing behind its own mistakes is a conflict buyers discount
What it builds up
The record of losses that prices the next policy better than any newcomer can (a cornered resource)
Why that profit
Loss data accumulates only with years of claims
Wrong if
Courts put agent losses on the model makers, so buyers need no separate cover
Already on the map
AIUC, Armilla AI, Mitigata, Munich Re (aiSure)
Builds on the site's reading
- Insurers write AI's mistakes out of their standard policies, so the cover has to be bought and priced on its own.
Needs first
A capacity partner willing to carry the risk
7Graded expert-work records pooled across buyers
A moderate profit, kept by the maker
Sits in 3.1 Data origination; touches Reward and verification systems, Evals and assurance
The problem
Expert data is made to order and belongs to the buyer, so the seller keeps nothing when the job ends
Sold to
Labs and large firms training models on expert judgement
First step
Sell graded work where the grading, not the answers, is licensed to many buyers
Charges for
A licence to a graded record per field
Why a lab would not just bundle it
Labs hire experts directly, but no single lab can grade across the whole market
What it builds up
A record of checked expert outcomes that improves every buyer's model and that no one buyer holds (a cornered resource)
Why that profit
The grading record is hard to rebuild, but the labs could bring the work inside
Wrong if
Labs bring expert work in-house and stop buying outside records
Already on the map
Mercor, Prolific, Bee Maps, Datacurve, Handshake AI, Human Archive and 10 more
Builds on the site's reading
- Expert-data sellers broker a skill the labs cannot yet make for themselves; what they sell belongs to the buyer, so their rent lasts only while the labs need people
Needs first
Experts and their employers agreeing that the grading may be licensed to many buyers, and at least two buyers who accept a shared record
8A router that chooses models from pooled outcomes
Competed away to users
Sits in 4.3 Routing and gateways; touches Compute utilization and control, Inference optimization, Outcome verification
The problem
Each firm guesses which model and venue to use for each task, and pays for the guess
Sold to
Platform teams spending on several model providers
First step
Save money on day one by sending routine work to cheaper models, and record which choices worked
Charges for
A share of the saving
Why a lab would not just bundle it
Each lab routes to its own models; a neutral router routes to whichever wins
What it builds up
Choices learned from many firms' outcomes that beat any one firm's own (network economies)
Why that profit
The site reads choosing a model as being given away unless pooled outcomes create the network
Wrong if
One provider becomes best on price and quality for most tasks
Already on the map
OpenRouter, Portkey, Vercel, AI21 Labs, Cloudflare, Helicone and 7 more
Builds on the site's reading
- Choosing a model is being given away; a gateway keeps a rent only as the place where developers pay for many models
- Routers gain power
Needs first
Customers who will report which answers worked, and enough spending across several providers for the saving to matter
The version that could hold
See Routing that keeps a firm inside approved models and fails over when one is cut off
9Escrow that pays an agent only when its work passes a check
A monopoly-like profit, kept by incumbent firms
Sits in 5.5 Settlement and billing; touches Identity and authority, Outcome verification, Routing and gateways
The problem
Agents can find and order work from other agents, but nothing holds the money until the work is proven done
Sold to
Marketplaces and firms whose agents buy services from other agents
First step
Escrow that releases payment when the result passes an agreed check
Charges for
A fee per settled transaction
Why a lab would not just bundle it
Card networks move money but do not judge whether the work was done
What it builds up
The record of which agents deliver, and the billing and credit that buyers and sellers would lose by leaving (switching costs, network economies)
Why that profit
Payment incumbents hold the rails; the opening is the check that releases the money
Wrong if
Agent-initiated purchases stay a tiny share of payments
Already on the map
Catena Labs, Paid, ProRata, Skyfire, TollBit
Builds on the site's reading
- Choosing a model is being given away; a gateway keeps a rent only as the place where developers pay for many models
Needs first
Agent identity that a payment network accepts, and a check both sides agree on before money moves
10The standard way agents get permission into business software
A moderate profit, kept by incumbent firms
Sits in 4.6 Tools and actuation; touches Identity and authority, Agent runtimes, Context and memory
The problem
Every agent needs its own negotiated, audited link into each system a firm runs
Sold to
Firms connecting agents to their records, and the software vendors who hold them
First step
Connectors for one crowded category of business software, with permissions and an audit trail built in
Charges for
A connected system per month
Why a lab would not just bundle it
The software that keeps the records can refuse or charge outside agents, so a neutral standard needs both sides
What it builds up
The standard link other agents must use if it spreads, and each customer's wiring if it does not (network economies, switching costs)
Why that profit
Record-keeping software owners can capture this themselves
Wrong if
Labs and clouds ship their own connectors and the record-keepers let them in free
Already on the map
Arcade, Vapi, Bland, Browserbase, Composio, Cua and 7 more
Builds on the site's reading
- Agent plumbing is mostly open source that labs and clouds buy or copy; the rent sits with whoever holds the connections into customers' systems
- The software that keeps the records gains
Needs first
The vendors that hold the records letting outside agents in, and a permission scheme their security teams accept
11Deployment that leaves behind the platform the customer runs on
A moderate profit, kept by the maker
Sits in 6.1 Technical implementation; touches Context engineering, Managed AI operations, Adoption and change
The problem
Engineers who can turn messy data and unwritten habits into working systems are scarce
Sold to
Mid-sized firms that cannot hire their own AI engineers
First step
Forward-deployed engineers who build on the firm's own map of its data and operations
Charges for
A deployment project, then a platform fee
Why a lab would not just bundle it
Labs are buying consultancies to do this, which is the risk as well as the proof
What it builds up
A platform the customer's operations run on, which it would have to rebuild to leave (switching costs)
Why that profit
The services margin grows with headcount; only the platform left behind earns a rent
Wrong if
Models need so little fitting that self-serve deployment wins
Already on the map
Anthropic enterprise services venture, Palantir, The OpenAI Deployment Company, Airbyte, Ampersand Labs, Artium and 24 more
Builds on the site's reading
- Fitting AI into a firm earns a services margin that grows with headcount; a rent lasts only where the deployer leaves behind software the customer runs its work on
- Deployment staff pull the labs' margins down
Needs first
Engineers who can work inside a customer, and first customers alike enough that the second deployment reuses the first
12Capturing how a firm really works, for its agents
A thin profit, kept by the maker
Sits in 6.2 Context engineering; touches Technical implementation, Context and memory, Managed AI operations
The problem
The knowledge an agent needs sits in people's heads, not in any system
Sold to
Firms whose agents fail on unwritten rules and exceptions
First step
Interview and observe one team, and turn its unwritten rules into tests and instructions agents follow
Charges for
A captured workflow, then upkeep as it changes
Why a lab would not just bundle it
The work is specific to each firm and changes over time
What it builds up
The captured rules the firm's agents run on (switching costs)
Why that profit
Useful and needed now, but copyable and likely to be automated
Wrong if
Agents learn a firm's unwritten rules on their own from its records
Already on the map
Edra, Mimica, Skan AI, Sugarwork, Tessl, Unblocked
Builds on the site's reading
- Agent plumbing is mostly open source that labs and clouds buy or copy; the rent sits with whoever holds the connections into customers' systems
- Fitting AI into a firm earns a services margin that grows with headcount; a rent lasts only where the deployer leaves behind software the customer runs its work on
Needs first
A customer that will let its people be watched and interviewed at work, and agents already failing on unwritten rules, so the gain can be measured
The version that could hold
See Deployment that leaves behind the platform the customer runs on
13A trade's AI system of record, with the approvals only that trade can give
A fat profit, kept by the maker
Sits in 7.3 Vertical workflow software; touches Autonomous services, Provenance and audit, Liability and risk transfer
The problem
A general model can draft the work but cannot hold the trade's records or sign for it
Sold to
Firms in a licensed trade, such as law, medicine or insurance
First step
The tool the professional already works in, which becomes where the matter, patient or case lives
Charges for
A seat, then a share of the work it completes
Why a lab would not just bundle it
Labs lack the trade's records, licensed content and the approvals a regulator gives
What it builds up
The trade's records and approvals, which customers would lose by leaving (switching costs)
Why that profit
Holding the trade's records is sticky; packaging a general model is not
Wrong if
Vertical apps' prices fall with each better general model as customers move to the labs' own products
Already on the map
Abridge, Harvey, Hebbia, Legora, 5c Network, Accrual and 52 more
Builds on the site's reading
- A trade's own AI app keeps a rent where it holds the trade's workflow, records and approvals; where it only packages a general model, each lab release takes the margin
Needs first
Standing in the trade, through a licence or partners who hold one, and customers willing to move their records
14An AI roll-up that moves its clients to fixed fees
A moderate profit, kept by the maker
Sits in 7.6 AI-native rollups; touches Autonomous services, Vertical workflow software
The problem
A firm billed by the hour hands AI's saving straight back to the client at the next re-bid
Sold to
Clients of hour-billed trades such as accounting, brokerage and property management
First step
Buy owner-run firms and switch their clients to fixed fees before automating the work
Charges for
A fixed fee per client or per job
Why a lab would not just bundle it
The asset is the customer relationship and the licence, which software cannot buy
What it builds up
A way of running many firms that rivals take years to learn, and lower costs per client than any single firm (process power, scale economies)
Why that profit
The roll-up owns the complementary assets; the AI itself is bought in
Wrong if
Margins fall back to the trade's ordinary level as unmerged rivals buy the same tools
Already on the map
2X, Adaptive Innovations, Agent AI, Agent Vista, AIAI Holdings (Ai²), AIXIOM and 100 more
Builds on the site's reading
- A rollup keeps AI's saving only where the firms it buys are paid for results rather than hours; its asset is the customer, and the AI is bought in
Needs first
Capital and a pipeline of owners ready to sell
15A licensed firm that sells the finished job and stands behind it
A fat profit, kept by the maker
Sits in 7.4 Autonomous services; touches Liability and risk transfer, Vertical workflow software, Outcome verification
The problem
Buyers want one accountable party when consequential work goes wrong
Sold to
Small and mid-sized firms that want the work done, not another tool
First step
One narrow service in one licensed trade, priced per outcome
Charges for
A completed job
Why a lab would not just bundle it
A lab would have to hold the licence and carry the liability
What it builds up
The customer relationship and the licence to do the work, which software firms cannot buy (a cornered resource)
Why that profit
The licence and the liability keep software vendors out, but other licensed firms can copy the model
Wrong if
Labs launch services in the same trades with their own warranties
Already on the map
Adaptive, AIOS, Akido, Alma, Arca, Arcos Law and 81 more
Builds on the site's reading
- A trade's own AI app keeps a rent where it holds the trade's workflow, records and approvals; where it only packages a general model, each lab release takes the margin
- A rollup keeps AI's saving only where the firms it buys are paid for results rather than hours; its asset is the customer, and the AI is bought in
Needs first
A licence or a licensed partner, insurance that covers work done by software, and one narrow service whose result can be checked
16Moving computing to where power is available
A moderate profit, kept by incumbent firms
Sits in 1.6 Compute utilization and control; touches Power and sites, Compute providers, Inference optimization
The problem
A site already wired to the grid is scarcer than the building or the money
Sold to
Data-centre operators, utilities and large compute buyers
First step
Shift flexible workloads to hours and places where power is spare, and sell that flexibility to the grid
Charges for
A share of the power saved or the capacity freed
Why a lab would not just bundle it
It needs both the grid's and the operator's data, which neither side holds alone
What it builds up
Agreements with utilities that let a site connect sooner (a cornered resource)
Why that profit
The scarcity is real while grid hookups lag; the rent goes to whoever holds the grid position
Wrong if
Interconnection queues shorten sharply
Already on the map
Anyscale, Emerald AI, Hammerhead, Niv-AI, Pado AI, QUTWO and 3 more
Builds on the site's reading
- A site already wired to the grid is scarcer than the building or the money
Needs first
Workloads that can wait or move, and a utility or grid operator that will pay for flexibility or connect a flexible site sooner
17Training and placing the engineers who deploy AI
Competed away to users
Sits in 6.5 Talent production; touches Technical implementation, Context engineering, Managed AI operations
The problem
Deployment talent, not models, keeps pilots from reaching production
Sold to
Deployment firms and enterprises short of deployment engineers
First step
A paid apprenticeship that places graduates in deployment roles
Charges for
A placement fee per engineer
Why a lab would not just bundle it
Labs hire the best themselves but cannot train the whole market
What it builds up
A network of placed alumni and the firms that hire them
Why that profit
The shortage is real now and will ease as training spreads
Wrong if
Deployment staff become easy to hire
Already on the map
ADaSci, Aggregate Intellect, AI Builder Club, AI Camp, AI Engineer, AI Makerspace and 34 more
Builds on the site's reading
- Fitting AI into a firm earns a services margin that grows with headcount; a rent lasts only where the deployer leaves behind software the customer runs its work on
Needs first
Employers that commit to hire graduates before the course starts, and working deployment engineers to teach it
18Routing that keeps a firm inside approved models and fails over when one is cut off
A fat profit, kept by the maker
Sits in 4.3 Routing and gateways; touches Identity and authority, Provenance and audit, Inference supply
The problem
Access to models splits by country and supplier, and a banned supplier breaks the product
Sold to
Firms and governments bound by rules on which models they may use
First step
A register of approved substitutes for each model, with automatic failover
Charges for
A fee per governed workload
Why a lab would not just bundle it
No lab will route customers to a rival when it is cut off
What it builds up
An approved-model register that becomes part of each firm's compliance (switching costs)
Why that profit
Rules create the need; the routing itself can be copied
Wrong if
Cross-border access to models widens again
Already on the map
OpenRouter, Portkey, Vercel, AI21 Labs, Cloudflare, Helicone and 7 more
Builds on the site's reading
- Choosing a model is being given away; a gateway keeps a rent only as the place where developers pay for many models
Needs first
Rules specific enough to encode, and substitute models tested as good enough for each governed task
19Tuning open models to frontier quality for one industry
Competed away to users
Sits in 3.7 Post-training and learning infrastructure; touches General-purpose models, Inference supply, Evals and assurance
The problem
Open weights are free but turning them into a reliable product takes scarce skill
Sold to
Firms that want frontier quality without depending on a closed lab
First step
Tune an open model for one industry's hardest tasks and prove it on that industry's tests
Charges for
A tuned model per industry, with upkeep
Why a lab would not just bundle it
Closed labs gain nothing from making open models better
What it builds up
The skill of tuning a model without breaking what it already knew (process power)
Why that profit
The site reads tuning as commoditising; the rent lasts only until clouds bundle it
Wrong if
Closed labs reopen a lasting lead over open models
Already on the map
Baseten, Prime Intellect, Adaption Labs, Adaptive ML, Applied Compute, Arcee AI and 13 more
Builds on the site's reading
- Open-weight labs give the model away, so they keep little of its value but cap what the closed labs can charge
- Tuning a model for one customer is becoming a feature of the clouds and the serving firms; a rent survives only where the tuned model keeps learning inside the customer's work
Needs first
Open models close enough to the frontier, and an industry test set the customer trusts
The version that could hold
See A model that keeps learning inside one customer's work
20Incident response and forensics for agent failures
A moderate profit, kept by the maker
Sits in 5.1 Security; touches Provenance and audit, Liability and risk transfer
The problem
After a public agent failure, nobody can say quickly what went wrong or who is liable
Sold to
Firms whose agents have caused, or could cause, a costly failure
First step
Retained responders who can reconstruct an agent's actions from its logs
Charges for
An annual retainer plus fees per incident
Why a lab would not just bundle it
The maker of the failed agent cannot investigate itself
What it builds up
A name insurers and courts trust after a failure (branding)
Why that profit
Scarce expertise after a shock; rivals follow once demand is proven
Wrong if
Two years pass with no material agent failure
Already on the map
7AI, Barndoor AI, Dropzone AI, Geordie, Gray Swan, Lakera and 8 more
Builds on the site's reading
- Insurers write AI's mistakes out of their standard policies, so the cover has to be bought and priced on its own.
Needs first
Agents that leave logs worth reading, and insurers or law firms that will name a responder in advance
21A firm's own map of what its data means, kept on its own servers for every agent to read
A moderate profit, kept by the maker
Sits in 4.4 Context and memory; touches Context engineering, Technical implementation, Tools and actuation
The problem
What a firm knows is scattered across systems whose tables no document explains, so every agent has to be taught the firm again
Sold to
Large firms in regulated fields whose data may not leave their own systems
First step
Answer hard questions over the firm's data where it already sits, writing down what each table and term means along the way
Charges for
A platform licence, then each agent and application that runs on the record
Why a lab would not just bundle it
A lab or a cloud would hold the record on its own systems, and a software vendor would describe only its own tables; the buyer wants one record across all of them, on its own servers and usable with any model
What it builds up
The written-down map of what the firm's data means and how its work is done, which every agent and application comes to depend on (switching costs)
Why that profit
Everything a firm builds comes to read the record, so replacing the platform that keeps it means rebuilding that work; but the knowledge in it is the customer's own, and a rival or a better model could rebuild the map from the same data, so the hold is weaker than holding a trade's records
Wrong if
The labs, the clouds or the firms that already sell a firm its data warehouse and business software ship a model of the firm as a built-in feature that buyers accept, or customers move the record to another seller at little cost
Already on the map
Exa, ByteRover, Chroma, Cognee, Compresr, Contextual AI and 15 more
Builds on the site's reading
- Fitting AI into a firm earns a services margin that grows with headcount; a rent lasts only where the deployer leaves behind software the customer runs its work on
- Agent plumbing is mostly open source that labs and clouds buy or copy; the rent sits with whoever holds the connections into customers' systems
Needs first
A customer with data it may not move and questions it cannot answer today, a capable model that can run where that data sits, and engineers who can work inside its systems
The version that could hold
See Deployment that leaves behind the platform the customer runs on

Drafted by a model from the site's readings and reviewed by Alex Brogan, who set the order and released every record. What each needs first, who would buy it and the next step were drafted by a model; a company is named as a buyer only where this site's deal ledgers record the purchase. These are conditions, not forecasts: nothing here is scored.