What Stays Scarce in a World of Agents
General knowledge is shared, and every agent will have it. What stays scarce is what belongs to someone: six things, four of them sources of competitive advantage and two that decide who answers. The contest is over who holds each of them.
Key question. If every AI agent knows the same things, where does competitive advantage come from?
Key finding. In my forecast for 2032, when every agent knows the same things, what stays scarce is what belongs to someone: context, authority, accountability, process, purpose and culture. Four of them are sources of competitive advantage. Two, authority and accountability, decide who answers.
This is the last piece in a four-part series on agentic impact on core systems. Application Equilibrium set out the theory. Seven Domains ran it on multifamily. Will AI Agents Replace Systems of Record? argued that the core systems keep their records and their access, and that the contest moves to the working role. The last question is the person: who brings the agent, who owns what it learns, who gets paid, and what the workforce looks like.
Picture 2032. Agents are in every core system, the product that holds a company's central records, and in every employee's hands. They draft, reconcile, schedule, route and file at a speed no team matches. Agents drive the how of each task. People drive the what: what the organization is for, what it will answer for and what it refuses to do.
One recording of Accounting 101, by the best professor on earth, is enough for everyone. No one needs a second. Agents are that recording. Each arrives with comparable general knowledge, improves on a schedule, and gets an upgrade at the same moment as every other user of the product. Models do differ today. I assume the gap narrows. Same recording, same tools. What differs is what each firm does with them.
I make four points:
Six things stay scarce because each belongs to someone.
The agent's layers have different owners.
The person arrives first, and money can collect at four control points.
The entry rung of the workforce is where the early evidence shows up.
Six Things Stay Scarce Because Each Belongs to Someone
The world of AI agents is changing fast, and part of that change is convergence. The best models are catching up to each other. Stanford's 2026 AI Index found that on Arena, a public ranking of models, the top four companies were separated by fewer than 25 points as of March 2026. In early 2023 OpenAI's best model led Google's by 205.¹ A ranking is only a rough guide, but the direction is clear. Every agent will know about the same things.
So what stays scarce? Six things: context, authority, accountability, process, purpose and culture. Each one belongs to someone. Not in the legal sense. One party holds it, and one party answers for it.
The six differ in kind. Context is fragmented across systems. Authority is granted. Accountability needs a person who answers. Process takes years to redesign. Purpose is a choice. Culture is built by people over time.
Scarce thing | Whose it is |
Context | The firm and its core vendors, who contest what is published |
Authority | The enterprise and the core, which grant it |
Accountability | A named person, who answers for the work |
Process | The firm, which has to redesign it |
Purpose | The organization, which chooses it |
Culture | The people who make it |
Context. A model knows what a ledger is. It does not know this company's ledger, its customers, its contracts or its open orders.
That lives in the core systems, which the firm runs and the vendors build. Each core is built around an object no other domain is built around. The agent has to ask the core for it, and the core decides what to publish. So context is contested between the firm and its core vendors. It is the clearest case of the contest over who. The firm's intellectual property (IP) and trade secrets belong here too, and a trade secret can be protected in law.
In the property management systems I worked with, the exchange I saw most often was the system sending out the rent roll and the unit inventory and taking back guest cards and resident ledger charges. Application Equilibrium shows the same pattern between CRM and ERP, where only the booked order crosses to the ledger.
Authority. Permission is granted, not generated. An agent acts with the authority of a role someone else defines. The enterprise and the core keep that authority.
Microsoft's Entra, its identity service, requires at least one sponsor for each agent identity, a business representative accountable for the agent's purpose and lifecycle, without technical administrative access.² Yardi's June 2026 release says agent access through its Claude connector follows existing user permissions.³
Authority is scarce as a requirement, not as a source of advantage. It attaches to a named role and person that a model cannot supply.
Accountability. This is a separate claim from authority. Someone must stay responsible for decisions and outcomes even when an agent does the work, and a name on a form does not meet it. A sponsor who cannot see or supervise every action of a largely autonomous role still answers for it. Microsoft's guidance is to maintain sponsorship so it survives a sponsor moving or leaving.² A role with no person in it still needs a human owner of record.
In the software organizations I worked in, accountability rested with the people responsible, and the severity of an issue moved it up the organization. A low-priority bug stayed with the developer, the QA engineer and the development manager. A more significant issue climbed in proportion to its magnitude and to whether customers were affected.
I expect the same rule to hold for agents. Severity decides who answers. A named person answers for what an agent does.
Process. Process is how a firm redesigns its work around agents. The applications underneath shape it. Thomas Davenport argued about ERP in 1998 that an enterprise system imposes its own logic on a company's strategy, organization and culture.⁴
A core's published units of work are the steps an agent can call. An agent cannot redesign a process it cannot reach, so process and context are linked.
McKinsey's 2025 survey points the same way. Of the 25 attributes it tested, redesigning workflows had the biggest effect on whether a company saw an earnings impact from gen AI. Only 21 percent of respondents whose organizations use gen AI said they had fundamentally redesigned at least some workflows.⁵
History says redesign is slow. General purpose technologies "enable and require significant complementary investments," often intangible, and measured productivity lags until the investments are made.
When factories swapped electric dynamos in for steam engines, "productivity barely improved." Growth leapt in the 1920s, four decades after electricity was commercialized, once factories were laid out around the flow of materials.⁶ The knowledge is free. The redesign takes years.
Purpose. An organization's purpose is what it is for, what it will answer for and what it refuses to do. In 2032 an agent carries out a strategy. It does not choose one.
Nicholas Carr argued in 2003 that a technology everyone has stops being an edge. It becomes a cost of doing business, and what makes a resource strategic is scarcity.⁷
Carr wrote about IT infrastructure. I apply it to agents.
Every firm will have comparable agents. Agents shorten the life of differentiation that sits in features, because a feature is easier to copy than a culture. A stated strategy can be copied. A purpose cannot, because it is shown by what the firm refuses to do.
Culture. Culture is how people work together in a firm and how much of themselves they put in. Agents cut administration and automate the routine. They supply neither a culture nor engagement.
Engagement is scarce before any agent arrives. Gallup's 2026 global workplace report puts it at 20 percent of employees, down from 21 percent in 2024.⁸ The link to agents is mine.
General knowledge is open to every rival. A culture and a team are not. Both lose their edge when key people leave.
The strongest objection is that an agent with access to a firm's systems can acquire the firm's context. It can read what a core publishes. But the core decides what to publish, so the context the agent gets is granted, not generated, and it stays specific to whoever grants it.
The Agent's Layers Have Different Owners
Authority and accountability decide which layer belongs to the employer, because the employer grants the one and answers for the other.
A friend of mine has carried the same custom Excel ribbon from job to job. The ribbon is the useful picture. It is not his knowledge, which every colleague has. It is the kit he built, and it goes where he goes.
An AI agent that a person uses at work has three layers: persona, role and activity. Each has a different owner. The role layer is where authority and accountability sit. Persona and activity are what a person builds and carries.
Persona is how the agent sounds and what it tolerates. It belongs to the person. Role is what the agent is authorized to do and answers for. It belongs to the employer. Activity is the skill for one task, such as the ribbon. It belongs to whoever built it and to the market that sells it.
I call an agent's activity-layer items its kit: workflows, custom agents and skill files.
Layer | What it is | Who owns it | Does it travel? |
Persona | How the agent sounds and what it tolerates | The person | Partly. In the import flows I read, a summary moves and the corrections behind it do not |
Role | What the agent is authorized to do, and the role record, the history of decisions, corrections and approvals in the job | The employer | No. It stays with the role |
Activity | The skill for one task: workflows, custom agents, the ribbon | Whoever built it. The employer if built on the job. A contractor keeps it absent an assignment | Mostly not. The kit does not move |
The layers table is my model, not a statement of law or product behavior. Memory can move technically and still not move contractually or in a form that keeps it useful. The kit is subject to employment agreements and licenses.
Property operations show the role layer at work. A leasing role means answering prospects, scheduling tours and following up. RealPage says its own AI Leasing Agent is working across more than 175 management companies, and Yardi's June 2026 release says data access through its Claude connector follows existing user permissions.³
In both cases the agent carries a role the employer defined, and the property management system grants the access. Seven Domains maps the vendors behind each role.
The Excel ribbon is an activity-layer item, a kit of skills one person built for a task. Memory moves easily. The kit does not.
In March 2026 Anthropic launched a memory import, and Google followed with a Gemini import. Memory moves as a summary the old agent writes.⁹
What stays behind is the kit, the instructions and files a person built for a task. A custom GPT is one.
Ownership follows the layer. Aaron Hall of Hall PC writes that a business "almost certainly owns" the skill files its employees create on the job, and that a contractor keeps ownership of their work without a written assignment. An employee keeps general skill and knowledge, but not the former employer's trade secrets.¹⁰
Akerman LLP adds that courts have not yet directly addressed whether AI-generated summaries of trade secrets count as misappropriation, and that the direction of the law suggests the test is derivation, not duplication.¹¹ These are law firms' commentary, not settled law and not legal advice.
The first true bring-your-own agents, agents a worker chooses and brings to the job, will be contractors and consultants. They already bring their own tools, and they keep what they build absent an assignment.
They also face client security rules, confidentiality terms and limits on where data can be processed. The kit they bring has to fit each client. Their working arrangement, not their ownership of tools alone, makes them the early case.
The Person Arrives First, and Money Can Collect at Four Control Points
Context and authority decide what an agent can reach. The core controls what is published. The enterprise grants what an agent is allowed to do. Together they shape which parties can charge for its actions.
The person's own subscription follows ordinary consumer habit, not context or authority.
How the person arrives
Bring-your-own AI, employees using AI tools the employer did not provision, is already common. Microsoft and LinkedIn reported in 2024 that 78 percent of AI users bring their own tools to work.¹²
In a May 2026 vendor survey by Resume Now of 1,020 workers, 76 percent used AI tools they found and signed up for themselves, and only 14 percent had an employer-paid subscription.¹³ The Microsoft and Resume Now figures both count AI tools, not agents with enterprise access.
The habit starts earlier. Pew found that 64 percent of US teens have used AI chatbots and 54 percent have used them for schoolwork.¹⁴ Universities add to it. Cal State announced ChatGPT Edu for more than 460,000 students on 23 campuses, and Syracuse adopted Claude for Education.¹⁵
Google offers eligible US college students a free year of its AI Pro plan, which renews at $19.99 a month as of October 2026.¹⁶ OpenAI is testing ads on its Free and Go tiers.¹⁶
The campus account does not follow the student. Chico State revokes access and permanently deletes the data, with "no option to export or transfer." WashU deletes the history if you leave. Cal Poly's license "was not purchased to cover alumni." A personal account's custom GPTs "can't be merged or exported" when it migrates into the campus workspace, and access is "permanently lost."¹⁷
The moment that matters is graduation. A student can export chats and import them into a personal account in the first week of a first job, and Anthropic and Google both offer that import.⁹
Employers answer by provisioning. Beri reports that Samsung banned external generative AI in 2023 after engineers leaked data. It reversed in June 2026 after a proof of concept with 2,500 employees across ChatGPT Enterprise, Gemini and Claude. OpenAI says Samsung deployed ChatGPT Enterprise and Codex, its coding agent.¹⁸
The person chooses an agent from a list the employer has sanctioned. The employer buys and governs the enterprise tier.
Who gets paid: four control points
The four control points are four places money can collect. They are not four settled revenue streams. They are a forecast of where money can collect, not a description of how it flows today.
Two of them can be the same company: Samsung's enterprise tier is the agent vendor's. A vendor that bundles governance and action costs into one enterprise subscription would turn four control points into fewer revenue pools.
The agent vendor collects from the person, through a subscription after the free year or through ads.
Identity and governance collect from the employer, as they did when employees brought their own phones. The phone precedent is an analogy. VMware paid $1.54 billion for AirWatch in January 2014.
Palo Alto Networks agreed in 2025 to buy CyberArk for $25 billion, naming AI agents in its reasoning. Cyera agreed in July 2026 to buy Oasis Security, which sells "non-human identity and agentic access governance," for $1 billion.¹⁹
Microsoft reached general availability of Agent 365, its product for governing agents, on May 1, 2026, at $15 per user per month standalone, as of October 2026.²⁰
The cores collect for the action, whoever sends it. As of October 2026, Salesforce's Agentforce pricing page lists Flex Credits, its prepaid units for agent actions, at $500 per 100,000 and $2 per conversation. ServiceNow says headless actions consume the same Assist currency, the units it meters its own AI features in.²¹
The enterprise tier, the agent the employer provisions, collects last, and the employer picks it. The bet is that it picks the brand the employee already uses. By Beri's account, Samsung ran three vendors against each other and bought one.
Control point | Collects from | How | Example |
The agent vendor | The person | Subscription after a free year, or ads | Google AI Pro at $19.99 a month. Ads on ChatGPT Free and Go |
Identity and governance | The employer | Licenses, and acquirers' bets on the category | Agent 365 is a charged price. AirWatch, CyberArk and Oasis are acquirers' bets |
The core | The employer | A charge per action | Salesforce Flex Credits. ServiceNow Assist currency |
The enterprise tier | The employer | A seat for the agent the employer provisions | Samsung's ChatGPT Enterprise, chosen, Beri reports, after a three-vendor proof of concept |
Two gaps remain in the evidence. No source I found separates work use from personal use in consumer AI revenue. None shows an employer paying for an employee's chosen consumer agent.
The loyalty a person builds at 19 is worth money at 25 only if an employer buys it. That is still a forecast.
The Entry Rung of the Workforce Is Where the Early Evidence Shows Up
Stanford's Digital Economy Lab used payroll data from ADP, a payroll provider, on millions of US workers through June 2026. It found "no evidence of widespread, economy-wide job displacement."
It also found employment of workers aged 22 to 25 in AI-exposed occupations 19 percent below where it would be against their less-exposed peers, with no comparable gap for experienced workers.
The gap runs through reduced hiring, not separations. It concentrates where AI substitutes for tasks. Where AI complements workers, employment is flat or rising. The authors call the findings early and descriptive, not causal.²²
The rote work that taught juniors judgment goes to agents. The result is a missing bottom rung, not mass layoffs.
In the teams I worked on, new hires were trained, given manageable assignments and achievable goals, mentored and peer reviewed, as any new hire is. There was no magic.
Deliberate training and review built judgment in those teams. An organization has to keep doing both once agents take the rote work.
Accountability and entry-level work meet here. Accountability is the scarce thing, and entry work is where people learn to carry it. If agents take that work, organizations face a shortage of junior jobs and of future people qualified to supervise agents and answer for what they do.
Replenishing that supply becomes a management problem of its own. Organizations will need to redesign how work gets done and how people acquire the experience to supervise it. The mechanism is mine. The Stanford data show a hiring gap. They do not show that agents removed the tasks that teach judgment.
What to Do, and What to Watch
What to do. Name a human sponsor for every agent, and start by inventorying the agents already running. Put agent kits and memory into employment and contractor terms, and have counsel review them, because the ownership views here come from law firms' commentary. Protect the entry-level work that teaches judgment, and keep a person on it.
What to watch.
An enterprise tier adds import of personal context or skills.
A core publishes terms for a personal agent that differ from its terms for an enterprise-provisioned one.
A campus deal adds alumni conversion.
An employment agreement names agent memory and who owns it.
A court decides a departing-employee case involving AI-generated summaries of a former employer's information. Akerman says courts have not yet addressed the question.
Contractor engagement terms list agent kits as an asset.
Four things would change this view of what stays scarce:
The young-worker gap closes with no change in hiring practice.
An agent's calibration proves fully portable, which removes the ribbon.
One model gains a durable lead over the rest, which ends the premise that general knowledge is shared.
Firms with comparable agents compete on price and features, and no lasting difference traces to context, process, purpose or culture.
The recording can be shared. The ribbon, the permission and the practice cannot.
There are four seats in the name. The fifth one is yours.
¹ Stanford HAI, The 2026 AI Index Report, chapter 2, section 2.1, "Model Performance Converges at the Frontier": AI Index 2026, chapter 2. The chapter highlights say "Top model performance is converging, with 4 companies now clustered within 25 Elo points." Using the Arena leaderboard exported in March 2026, the section says the top four models are separated by fewer than 25 points, with Anthropic at 1,503, xAI at 1,495, Google at 1,494 and OpenAI at 1,481. It gives early 2023 as the contrast: OpenAI's top model at 1,322 against Google's at 1,117, a gap of 205 points. The report cautions that leaderboard standing may partly reflect adaptation to the platform rather than general capability alone. Read by me October 27, 2026 through the chapter PDF; the 205-point gap is my subtraction of the two scores.
² Microsoft Entra, owners, sponsors and managers of agent identities: Microsoft Learn.
³ Yardi's Virtuoso Connector for Claude, June 16, 2026: Yardi. RealPage Lumina, August 10, 2026, including the AI leasing agent figure: RealPage.
⁴ Thomas Davenport, "Putting the Enterprise into the Enterprise System," Harvard Business Review, July-August 1998: HBR. HBR's page is paywalled. The sentence paraphrases the article's argument, and a summary at maaw.info reports it.
⁵ McKinsey, The state of AI: How organizations are rewiring to capture value, March 12, 2025, 1,491 participants in 101 nations: McKinsey. The correlation analysis considered 25 attributes.
⁶ Brynjolfsson, Rock and Syverson, The Productivity J-Curve, NBER Working Paper 25148, October 2018: NBER. Tim Harford on Paul David and the dynamo, June 2, 2007, republished by History News Network: HNN.
⁷ Nicholas Carr, "IT Doesn't Matter," Harvard Business Review, May 2003, reprint R0305B: HBR. HBR's page is paywalled. The two quoted phrases are as they appear in the article, and the author's own page is nicholascarr.com.
⁸ Gallup, State of the Global Workplace 2026, published April 2026 with 2025 data, global engagement 20 percent of employees, 21 percent in 2024: Gallup. The page reports no findings on AI. The link between engagement and agents is my reading.
⁹ Anthropic: Claude Help Center, Import and export your memory, which carries no date; I read it October 20, 2026. It has the user ask the old assistant to write out its memories and paste the result into Claude. A news report dates the launch March 2, 2026: SiliconANGLE. Google, March 26, 2026: Google, which describes a prompt pasted into the old app, the reply pasted into Gemini, and a ZIP upload of chat history.
¹⁰ Aaron Hall, Hall PC, on who owns AI skill files: Hall PC. A law firm's view, not a court ruling, and counsel has not reviewed it for this piece.
¹³ Resume Now survey of 1,020 workers, May 2026, reported June 24, 2026: Allwork.Space, which gives the 76 percent and 14 percent figures, and Orange Leader, which gives the sample size and survey month.
¹⁴ Pew Research Center, How Teens Use and View AI, February 24, 2026, from a survey fielded September 25 to October 9, 2025: 64 percent of teens have used an AI chatbot and 54 percent have used one for schoolwork. The 64 percent also appears in Pew's December 9, 2025 report.
¹⁵ OpenAI, February 4, 2025: OpenAI and the CSU system, which gives ChatGPT Edu to more than 460,000 students on 23 campuses. Syracuse and Claude for Education: Syracuse University, September 22, 2025. The Cal State figures are as OpenAI reports them.
¹⁶ Google's student offer of a free year of AI Pro: Google. OpenAI's ad test for logged-in adults on the Free and Go tiers, February 9, 2026, updated August 11, 2026: OpenAI. ChatGPT Ads "available in over 40 countries," August 31, 2026: OpenAI. That page does not name the Go tier.
¹⁷ Cal Poly, existing ChatGPT accounts: Cal Poly. WashU, account migration: WashU. Chico State, ChatGPT Edu FAQ: Chico State.
¹⁸ OpenAI, June 21, 2026: Samsung Electronics brings ChatGPT and Codex to employees, for all Samsung Electronics employees in Korea and all Device eXperience employees worldwide. That page does not mention the 2023 ban, the reversal or the 2,500-employee proof of concept. Those rest on one secondary source: Beri.
¹⁹ VMware and AirWatch, January 2014: eWeek. The VMware figure is as eWeek reports it. Palo Alto Networks and CyberArk, July 30, 2025: Palo Alto Networks, about $25 billion in equity value; the release says just-in-time access gives AI agents "only the permissions they need." Palo Alto Networks completed the CyberArk acquisition on February 11, 2026: SEC filing. Cyera and Oasis Security: Oasis, July 28, 2026, a letter of intent, and Cyera, September 3, 2026, completion at $1 billion. The quoted description of Oasis is from SecurityWeek; neither company's page uses those words.
²⁰ Microsoft, May 1, 2026: Microsoft Agent 365, now generally available. The $15 per user per month price for the standalone product also appears on Microsoft's pricing page, listed as paid yearly.
²¹ Salesforce's Agentforce pricing page, which carries no date; I read it October 20, 2026. A news report dates the pricing announcement August 31, 2026: Let's Data Science, citing GuruFocus. ServiceNow, May 5, 2026: ServiceNow opens its full system of action to every AI Agent in the enterprise. The release says headless actions consume the same Assist currency as Now Assist and AI Agents. It gives no rate.
²² Stanford Digital Economy Lab, Brynjolfsson, Chandar and Chen, August 12, 2026: Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence and the news summary.
Comments