Will AI Agents Replace Systems of Record? A 2032 Forecast
Picture 2032: every core is headless and employees bring their own agents. Work backward and the cores still grant the access. The contest is over who holds how the role is performed.
Key question. When AI agents become the main users of enterprise software, do they change which vendor controls the system of record, or only how people reach it?
Key forecast. I forecast that through 2030, agents do not replace systems of record. They change how people reach a core, not who controls it. Plainly, ERP and CRM vendors keep control of the records and the access. That control does not guarantee their pricing power.
An agent changes the equilibrium of a domain only if it forms a core around an object of its own. The candidate is how a role is performed: the role record, its calibration, and the process and rigor built around them. Whether that record is a new core object is an open test. 2030 is the horizon of the claim. The 2032 scenario below is the picture it works back from.
This is piece 3 in a four-part series on agentic impact on core systems. Application Equilibrium set out the theory. Seven Domains ran it on multifamily, where AI so far strengthens whoever already holds the object. Here I ask whether agents change that, and work backward from 2032 to see who holds what when every core is headless.
The series: 1, the theory (Application Equilibrium); 2, the multifamily example (Seven Domains); 3, the pressure test (Will AI Agents Replace Systems of Record?); 4, what holds (What Stays Scarce in a World of Agents). You are reading 3.
A core is a product built around one core object, the record around which it organizes its process and data, such as the ledger in enterprise resource planning (ERP) or the account in customer relationship management (CRM). Its system of record is the authoritative history of that object. The title's systems of record are the core applications themselves, as the market uses the phrase.
Enterprise software keeps changing shape. It ran on premise, moved to the cloud, and is now going headless. A headless system keeps its data and its rules, and the screen becomes optional. AI agents, software that acts on someone's behalf through a program interface, read and write it directly through an application programming interface (API). That is what headless ERP and headless CRM mean.¹
I make three points:
An agent is a new path to a core, not yet a new core.
The cores keep the system of record and the access, and grant it to agents on their terms.
The contest moves to the role, where nobody yet holds the role record.
An Agent Is a New Path to a Core, Not Yet a New Core
Application Equilibrium traced the surfaces people have used to reach a core: commands, terminal screens, windows, pages. Surfaces changed, and none of them moved the object.
One step is different in kind. Salesforce launched its own web API in 2000, and eBay and Amazon followed.² The web API serves programs, not people. Every other step is a surface for a person.
The agent is the first step that is both. It acts for a person through the program interface, in plain words. It is also the first surface a person chooses that acts across many cores, though choosing is not owning.
In each earlier step I traced, the application vendor built the surface: its terminal screens, SAP's graphical user interface, its web pages, its mobile app. That was one surface per core. An agent is one surface across many cores, and the person picks it.
The break I cite took years. Salesforce was founded in 1999,³ and it spent years building before the account made CRM a domain of its own.
So why would an agent move faster? Its users are already in place. In a May 2026 vendor survey of 1,020 workers by Resume Now, 76 percent used AI tools they found and signed up for themselves.⁴ Those are AI tools, not agents with enterprise access. The figure shows readiness to adopt, not agents in use.
Users already in place remove the slow part of an agent's rise. They do not supply a new core, and a new core requires a new object. For an agent the candidate is the role record, the history of how a role is performed. Someone has to build process and rigor around it, and the cores have to be unable or unwilling to build it themselves.
The reason can be commercial. A core paid by the seat earns less when an agent does the work and removes seats. Customer service shows both sides. Sierra, which sells customer-service agents, charges for outcomes, not seats.⁵ Intercom, a customer-service platform sold by the seat, adds a per-resolution fee on top, so some cores can adapt.
Adopting a surface is quick because it is superficial. It can mask complexity or automate a process, and it takes less user training than the process underneath. Moving the system of record behind it is not quick. In the ledger decisions I have seen, the CFO signs that the record is right, and moving it puts that signature at risk.
So an agent could break an equilibrium faster than any surface before it. If it does not hold a record of its own, the theory holds: a surface alone does not move a domain.
The scenario: it is 2032
The scenario holds five facts fixed:
Every major core provides three things: agents of its own, a door for outside agents that publishes units of work, defined actions such as approving an invoice, not only data, and the terms for both.
The enterprise and the core grant each agent's access and keep the system of record. No agent grants itself any.
Pricing counts what the customer controls, not seats.
Employees choose the agent they work with.
The employer sponsors each agent in a role.
In this scenario agents have not replaced systems of record by 2032.
Fact 2 carries the conclusion. Evidence against it would be cores that let agents grant themselves access, or stop gating and pricing it. I hold it because the strength of a core's equilibrium is its object and the audit trail around it. A customer that needs one answerable record has the same reason to keep access gated as the core does.
What stays open is who holds the role record.
The Cores Keep the System of Record and the Access
Cores publish units of work. Two kinds of AI act on a core object. They map onto the 2000 buyer's question, integrated or best of breed: take the function into the vendor's suite, or attach a separate product to it.
Internal AI works for the application inside one vendor's boundary, on that vendor's data. That is integrated, because one vendor draws the line around everything the AI touches. External AI works for the user, reaching into each vendor's system through its API or its Model Context Protocol (MCP) server, the standard interface built for agents. That is best of breed, because many vendors sit on the other side of the line.
Traditional API access moved data in and out. Headless adds process activation, the ability to start what the core does, not just read what it holds. Whether those interfaces exist at all is the core's decision. It chooses what they expose, what they cost and which agents are reviewed to use them.
On the 2026 evidence from five vendors (Salesforce, ServiceNow, Workday, SAP and ServiceTitan), that control is more power to the core, not less. It is why the fight runs through terms and contracts.
The cores have answered two ways. Some opened the door and kept the terms. Salesforce went headless itself in April 2026.¹ ServiceNow says it opened its workflows and approvals to any AI agent in May 2026, with every action permission-scoped and audited. Workday says it offers outside agents defined business actions scoped to the user's identity and permissions.⁶
Others wrote the terms around the agent. SAP's April 2026 API policy prohibits API use for AI systems that plan, select or execute sequences of API calls, except through SAP-endorsed architectures.⁷
The April 2026 terms of ServiceTitan, which sells software to home-service contractors, say a partner may not let an AI system independently determine which endpoints to access. ServiceTitan also ended the integration of Podium, a customer-messaging vendor, for about 1,000 shared customers on about 30 days' notice.⁸
The line between the open cores and the gated ones is not openness. It is which agents the core has reviewed.
The cores' answer is taking shape as published units of work in some cases and a governed dispatcher in others, not a bare endpoint. SAP exposes its Joule agents to outside agents over the Agent2Agent (A2A) protocol, inbound as of October 2026, with two-way communication not yet supported.⁹ Workday and ServiceNow publish defined actions and flows the same way.⁶
The core publishes the unit and keeps everything else inside. The orchestration agent invokes the unit from outside. The core keeps its logic, security model and audit trail. I call it API II (Agent Protocol Interface), a wink at Gartner's ERP II. It is not an adopted standard.
Access terms change the contest. The party that sets the access terms can also sell an agent. So the question is whether the orchestration agent a person works through comes from a core or from a third party the core has to let in.
If the orchestration agent becomes where work begins, the application a person works in gets easier to swap. The cores' bargaining position shifts from the screen to the access terms. The system of record behind it stays hard to replace unless the agent also holds how the role is performed.
The enterprise and the core grant the access. Besides the record, the cores keep the security model: who can see what and do what. On the 2026 evidence, agents do not bypass it.
Workday and ServiceNow scope outside agents to the user's permissions,⁶ and ServiceTitan reserves the right to require human authorization for writes, if it permits AI use at all.⁸
An agent acts with the authority of a role the core defines. An agent that does not work within each core's security model risks being cut off, as in the Podium case above.
The cores stay. Foundation Capital wrote in December 2024 that systems of record would be reduced to commoditized storage. Satya Nadella said on a podcast reported that month that once the logic sits in an AI tier, people will start replacing the back ends.¹⁰ Bessemer's 2025 paper takes the longer view: a working thesis that systems of record will be reimagined within five to ten years, which runs past 2030.¹¹
As the theory piece argues, the system of record is the hard thing to move, and a commodity that works and holds the audit trail is the last thing a company wants to migrate. What commoditization removes is pricing power over storage. Headless costs the core its screen, not its system of record or its access terms. So the fight moves to access and the paper around it.
Control of access is leverage, not a guarantee. A core can keep its system of record and its terms and still lose pricing power, or the user's attention, to the orchestration agent above it.
Two earlier layers sat above the cores without displacing them: robotic process automation and integration platforms. Both reached the cores through the screens and APIs the cores allowed. The cores kept the record. The harder rival is a horizontal assistant that already holds a user's attention and memory. It moves a domain only if it holds the role record.
Employees choose, and the employer sponsors. Microsoft's Entra requires at least one named business sponsor for each agent identity.¹² I read that as the employer's answer to employees who bring their own agents. The employee picks the agent. The enterprise issues the identity and keeps the access.
Who picks the agent and who holds the role record are separate decisions. How the habit forms in school, what a person carries from job to job and who gets paid are the subject of What Stays Scarce in a World of Agents.
The Contest Is Over Who Holds the Role Record
In a headless world the cores keep the system of record and grant the access. The contest is over who holds the role record.
Above the cores sits the orchestration agent, the person's own agent, working across vendors. It takes over a job people have always done: the swivel-chair integration, the spreadsheet populated from one system and keyed into another. The person was the integration layer. The agent is the next level of integration. After the swivel chair came point-to-point interfaces, middleware and APIs. Now software does the integrating itself.
What a role record is
There is one object an orchestration agent can hold that no single core holds: how a role is performed. A role is a job as the people who hold it perform it. The object has two parts.
The role record is the history of decisions, corrections and approvals. A workflow log records what happened, and agent memory records what a user said. The role record also holds who corrected and who approved. It differs from a core's audit trail in two ways. It spans applications, and it carries calibration.
The calibration is what the agent learns from that record about how the people in the role decide and what they correct and approve. The role's authority is not the orchestration agent's to hold. The enterprise and the core grant it, and the agent exercises it by delegation.
Take an accounts payable lead, as an illustration. The role record holds which invoices the lead held back, who corrected the coding and who approved the exceptions. The calibration is the pattern in those corrections, such as which vendors get questioned and which get paid on sight. When the lead moves on, the successor inherits both.
The calibration is where the lock-in sits. Stated preferences export, and in March 2026 the vendors built import tools for exactly that.¹³ What months of corrections and approvals build is a different matter. No import tool I found claims to carry it. My claim is not that it cannot move. It is a different kind of record from a stated preference, so moving it is hard in practice.
It is the Salesforce move one level up. The account was the persistent entity across transactions, which let Salesforce make it a core of its own. The role is the persistent entity across applications.
Does the role record count as a core object?
The object test has three parts. It is not an attribute of any existing core: each core sees only its own part of how a role is performed. It can be held once: the record of decisions, corrections and approvals, with who made them. And process and rigor can be built around it: capture, review, calibration, and a handoff when the role changes hands.
The first two are possible now. As of October 2026 I have not found a product that holds a single authoritative role record.
The third decides. A consolidated record alone is an integration artifact: a data platform can collect events across applications without organizing a domain around them. The organization would also need to treat the record as authoritative enough to govern how the role is performed, not as one source of advice among many.
A pile of logs and memories is not a record anyone has built rigor around. The succession test shows the difference: a log tells a successor what happened, and a record built for the role tells the successor how it is done.
To hold the record is to be the party whose process and rigor make it authoritative. Custody of the data can stay with the employer.
Binding the agent to the role, not the person, answers what happens when the person leaves. The employer keeps the role's authority and role record, so a successor inherits and adjusts the calibration. The person keeps their own voice. What a successor needs is calibration, and it stays with the role. What follows the person is persona, the agent's voice and habits.
The next level of integration leaves a new rival to the bundle: an orchestration agent that works across many cores without belonging to any. Apply the object test from Application Equilibrium, which asks which object each option holds. An orchestration agent wins where it holds how the role is performed and the cores let it act. Where it holds neither, it is a replaceable conductor and the pull goes back to the cores.
Three fights
Three fights decide who holds the role record.
Ownership. The contracts have not caught up on where the employer's role record ends and the person's own voice begins.
Portability. Whether the record leaves when the agent is replaced, and in what form.
Privilege. A person with logins to three systems already combines what no single role was meant to see or do. An agent does it faster, at volume and without the person's judgment. No core's security model covers the combination, so the orchestration agent has to enforce the limit itself.
The evidence on who holds the role record is not in yet. In 2026 the cores opened to agents they could govern, through ServiceNow's layer and Workday's defined actions.⁶ They restricted the agents that act across systems on their own, through SAP's policy and ServiceTitan's Podium shutoff.⁷ ⁸ I read that as the cores treating the acting orchestration agent as the greater threat.
The strongest counter is that customers keep the authoritative copy in a data layer they control, or that an AI-native replacement core arrives. The first still leaves the core as the process that makes the record authoritative. The second is the third outcome.
Three outcomes and two dated observables
Three outcomes are possible for the contest over the role record, and what changes for the customer tells them apart:
Agents only ease access. Work begins in an agent, the cores still organize and control the work, and the cores win. If cores gate and price agent access and customers stay, this one arrived.
Agents orchestrate across the cores. An orchestration agent that works across the cores becomes valuable and the cores stay authoritative, which is a strategic shift but not a new domain. If customers pay that agent for the work and still sign against the core's record, this one arrived.
An agent holds the role record. It builds process and rigor around that record, which makes it a core itself. Customers changing a core system without migration pain signals this one only if the agent holds the role record and does more than translate between the old system and the new one.
Two dated observables are early indicators of the second and third outcomes, not proof that a new core has emerged. The dates and thresholds are set so that I can be wrong, and either can fail before 2030:
Agents orchestrate across the cores. By December 31, 2028, at least one vendor with no core object reports more than $100 million a year from an orchestration agent that works across several cores, while its customers still sign against the cores' records. The revenue counts only if the vendor reports it for the agent itself, not inside a larger assistant or suite subscription. $100 million is the size at which an orchestration agent is a business of its own, not a feature of someone's suite.
An agent holds the role record. By December 31, 2029, an employer treats an agent-built role record, not its systems of record, as the authoritative history of how a role is done, shown by a named employer's statement or an independent analyst report, not a vendor case study alone, and a successor performs the role from it.
If neither resolves as stated, the first outcome stands or the thresholds were wrong. The headline itself would be wrong if, by December 31, 2030, customers in a domain retire their core and keep the authoritative record in an agent or in a layer they control.
What a Buyer Should Ask at Renewal, and What to Watch
In 2032 the agent carries the cost of connecting one product to another. I expect that to shrink the suite's oldest advantage, one vendor to integrate. It leaves the 2000 question horizontal software has always posed: integrated, the bundled version from the vendor you already pay, or best of breed, the tool that does one thing deeper.
The object test still applies: ask each vendor which object it holds and where its innovation lands. Buy for 2032 and ask in 2026. So far AI is speeding up whoever already holds the object. Ask about existing, proven integrations, not roadmap promises. Put the six contract questions beside the object test.
Six contract questions
RealPage, a property management software vendor, paused new activations of EliseAI, a leasing AI company, on October 7, 2026. Seven Domains covers it. RealPage cites EliseAI's obligations under the RealPage Exchange agreement, along with an agent never submitted for review and tenant data protection.¹⁴
The vendors are already rewriting their side. SAP and ServiceTitan changed terms in April 2026. Salesforce changed Slack's API terms in 2025 to bar bulk export, persistent copies and use of the data in large language models. In Hunton's account, customers without negotiated limits on unilateral amendments had to redesign their programs.¹⁵
Six questions belong in any renewal that gives an agent access to your data. They are for counsel, not legal advice:
Agent access. Who may connect an agent to my data, which processes may it start and not only which data may it read, what review do they need, and does the vendor's own agent face the same review?
Pricing. Does agent traffic count toward my fee, and on what metric?
Change control. How much notice comes before access terms change, and can they change without my consent?
Continuity. If the vendor pauses an integration, what happens to the ones already live?
Exit. What comes back to me if I leave, in what form, and can an agent read it?
Identity. Does the agent act under a user's role, as itself, or as the role with no person in it? Who answers for it then, does the audit trail show which, and what terms govern an agent the employee brought?
What to watch. The two observables above and the vendor terms, which change every few months.
The access stays with the cores, and the role is open. Whichever way the contest goes, a few things stay hard to supply. What Stays Scarce in a World of Agents covers them and who carries them.
The cores keep the access. The contest is over the role.
There are four seats in the name. The fifth one is yours.
¹ Salesforce Headless 360, introduced April 15, 2026 at TrailblazerDX: Beri. Salesforce, April 21, 2026: 6 New Innovations Redefining Salesforce Development, which says every capability Salesforce has built over 25 years "is now available via API, MCP, or CLI." That post does not use the name Headless 360. Marc Benioff's April 17, 2026 post, "No browser required. Our API is the UI," is quoted by SaaStr.
² Salesforce's SOAP API, 2000: Nordic APIs. eBay's API, November 20, 2000, and Amazon Web Services, July 16, 2002: API Evangelist. The sequence of surfaces is set out, with its earlier dates, in the first piece of the series.
³ Salesforce was founded March 8, 1999: Wikipedia on Salesforce.
⁴ Resume Now survey of 1,020 workers, May 2026, reported June 24, 2026: Allwork.Space, which gives the 76% figure, and Orange Leader, which gives the sample size and survey month. The figure counts AI tools, not agents with enterprise access.
⁵ Sierra, Outcome-based pricing for AI agents, December 10, 2024: Sierra says it gets paid only when it completes a task, with no reliance on seat-based pricing. It gives no dollar figure. Intercom, pricing, September 24, 2026: seat plans of $29, $85 and $132 per seat per month billed annually, and Fin AI Agent at $0.99 per outcome on every plan. Intercom's help page on Fin outcomes gives the same $0.99 per resolution and carries no date.
⁶ ServiceNow's release, May 5, 2026: ServiceNow. Workday's Agent Ready Tools: Workday data sheet.
⁷ SAP's API policy, last updated April 27, 2026: The Register.
⁹ SAP's Joule agents exposed over A2A: SAP Architecture Center. The page says two-way communication with third-party agents through the gateway is not yet supported. A2A is hosted by the Linux Foundation: Linux Foundation, April 9, 2026.
¹⁰ Ashu Garg and Jaya Gupta, Foundation Capital, December 3, 2024: How Systems of Agents will collapse the enterprise stack. Satya Nadella on the B2G podcast, reported December 23, 2024: CX Today.
¹¹ Bessemer Venture Partners, May 29, 2025: Roadmap: AI systems of action (Droesch and Patel). It says AI is degrading the traditional moats of systems of record and calls it a working thesis that systems of record will be completely reimagined within five to 10 years.
¹² Microsoft Entra Agent ID, owners, sponsors and managers, page last updated July 21, 2026.
¹³ Anthropic's memory import, March 2, 2026: SiliconANGLE. Google's Gemini import, March 26, 2026: 9to5Google.
¹⁴ RealPage CEO Dirk Wakeham, October 12, 2026: Multifamily Dive. The October 7 pause was first reported by Thesis Driven. The Thesis Driven letter is partly paywalled, and the CEO's statement was read in Multifamily Dive.
Comments