Systems of record stay. The tools around them consolidate.

A mid-market company runs a handful of systems of record: the CRM, the ERP, HR and payroll, finance, and usually a warehouse that holds history from all of them. Each owns an authoritative copy of something. The customer record lives in the CRM. The ledger lives in finance. These systems carry the data model, the approval rules, the audit history, and the compliance settings that make the records trustworthy. IDC makes the same point in its rebuttal to the sell-off: a system of record does not disappear when the interface on top of it changes.

Around them sits a second layer that is easy to overlook because each piece is small. A form collects vendor details and someone keys them into the ERP. A workflow tool copies new customers from the CRM to the ERP. A relay catches the storefront’s webhooks and lands them for analysis. A parser reads supplier invoices out of a shared inbox. A BI subscription emails the Monday revenue pack. A portal shares monthly results with a client. A catalog explains what the fields mean.

Lay those tools against the systems they touch and a pattern shows. Every one of them reads from one place and writes to another. None of them owns a record of its own. They copy, reshape, route, and present data that a system of record already holds.

Satya Nadella described business applications in late 2024 as databases with business logic on top, and predicted that the logic moves to an agent tier working across many databases at once. The databases stay in that picture. What moves is the logic between them, which is exactly what the second layer is made of.

Klarna is the case people cite, usually wrongly. In August 2024 the company said it had shut down Salesforce, and the coverage read it as AI replacing a CRM. In March 2025 its chief executive corrected the record: Klarna did not replace its SaaS with a language model. It found its knowledge scattered across many SaaS products, consolidated that data into an internal stack built on a graph database, and built lightweight tools on top. Unifying the data came first, and the tools became redundant once it was reachable.

The economics explain why. A tool in the second layer was never expensive because its logic was hard to write. It was expensive because of everything around the logic: credentials for each system it touches, a runtime that fires on time, permissions on who may change it, a record of what it did, and somebody who hears when it breaks. An agent can now write the logic in minutes. The rest is still needed, and it is the part the subscription was really paying for.

If authoring is nearly free, a rebuilt workload lands wherever those other needs are already met. That rules out the obvious candidate. Gartner expects 40 percent of enterprise applications to include task-specific agents by the end of 2026, up from under 5 percent in 2025. Each of those agents lives inside one application and sees one system. A workload in the second layer spans at least two systems by definition, so it cannot live inside any one vendor’s agent.

It lands with whoever already holds governed access across the systems: the connections and their credentials, roles that decide who reaches what, an audit trail, a record of what the data means, and a runtime. That is the position Plexara is built for. Your agent writes the workload as an automation that runs on demand, on a schedule, once for each event, form submission, or approved email, or once per request from another system. It reads and writes through the connections the platform already holds. The person it runs for approves what it does in plain language, and any later change in its behavior is approved the same way. Every call is audited under the same roles that govern every agent session.

The components cover most of the layer. Webhooks and a mailbox source take events and email in as queryable tables. Forms collect intake into a table and start a run per submission. Reports and dashboards refresh themselves and post to Slack, Mattermost, or email through notification channels. The catalog describes every connection. Clients and partners see exactly what is shared with them through your identity provider, with no access to the agent.

Gartner also expects more than 40 percent of agentic AI projects to be cancelled by the end of 2027, citing cost, unclear value, and inadequate risk controls. The governance in that list is the same governance a rebuilt workload needs. A workload that runs under audited roles, with its behavior approved and its failures reported, is the kind that survives that review.

Nobody should cancel a working tool mid-contract on the strength of a slide. The practical unit of change is one workload, timed to that tool’s renewal. Zylo’s 2026 index found 36 percent of SaaS licenses unused and 81 percent of SaaS spend controlled by business units rather than IT, which is how an estate accumulates three tools that each move the same data. Renewal is the one moment each of those purchases gets looked at again. IDC gives buyers the same advice from the other side of the table.

A move has five steps. The agent rebuilds the workload against the same systems. The rebuild runs beside the old tool on the same inputs while the old tool still does the real work. Outputs are compared row by row and column by column until they match or every difference is explained. The person the work runs for approves the automation in plain language. Then it takes the real work, and the old contract lapses when it comes up.

The money involved is real and modest. Zylo puts median SaaS spend at $9,455 per employee, about $1.9 million a year for a 200-person company. The categories in the second layer are a small share of that, on the order of 3 to 20 percent depending on how many of them a company buys. The large line items are the systems of record, and they stay. The stronger case is operational: one place to see what runs, one set of credentials, one audit trail, and one alert when something stops.

A media organization retired every flow in its legacy integration platform this way. Each flow was rebuilt as an automation, run beside the original, and checked against the old output column by column before it took over. One customer sync now makes a single call to the CRM where the old flow made five. Client staff now write and own automations of their own.

A retail organization is partway through the same process. A customer sync between its CRM and its ERP has run every two minutes since the old pollers were switched off. Invoice, warehouse, location, and card-settlement flows have been rebuilt and tied out column by column, and one of them now reports the failures the old flow dropped without a trace. A library of legacy reports was rebuilt as automations, and the comparison turned up a double count the old report had been carrying.

Neither company changed its CRM or its ERP. The work between those systems moved onto the platform, one flow at a time.

The systems of record stay, and Plexara works through them rather than instead of them. Several categories that sit near the second layer stay too, because their core job is something an agent-written automation does not do: replicating hundreds of sources with change capture, self-service exploration for hundreds of BI users, branching survey forms with payments, enterprise integration with EDI and on-premises agents, and extracting fields from scanned documents.

Inside that line, the second layer consolidates one workload at a time onto the platform that already holds the access, the meaning, and the record. Replacing the five-vendor data stack covers the same pattern one level down, in the data infrastructure itself.