The Unbundling of the Seat

Salesforce just told the market what an agentic transaction costs. They have not told it who decides whether the transaction should execute, or who is responsible if it shouldn't have. Three meters are separating where one used to live. The procurement contract has not caught up.

On February 25, 2026, Salesforce reported Q4 FY26 earnings. Agentforce annual recurring revenue reached $800 million, up 169% year-over-year. The company introduced a new unit of measurement called the Agentic Work Unit, with 2.4 billion AWUs delivered to date across Agentforce and Slack, growing 57% quarter-over-quarter. Salesforce has processed 19 trillion tokens, up 5x year-over-year. In November 2025, three months before the earnings call, the company closed its acquisition of Informatica, bringing master data management, the data catalog, governance, quality, and metadata services into the platform. Read those four facts in sequence and the strategy is unmistakable. Salesforce has decided that the unit of value is no longer the seat. It is the work an agent does on behalf of the seat. And it has begun acquiring the substrate that makes that work governable.

This essay is about what the rest of the industry is going to learn, harder and slower, from the move Salesforce made first. The seat — the single line item that has anchored enterprise software pricing for thirty years — is coming apart. Where one bundle used to hold the commercial meter, the access right, and the accountability locus together, three separate planes are now emerging. Each has its own buyer. Each has its own architecture. None of them is yet bound to the others by contract.

The historical frame

The seat was a brilliant simplification. For three decades it served as the universal bundling unit of enterprise software, holding together three distinct things that the industry treated, correctly, as one. A seat priced the work — the vendor charged per human in the system, and that scaled with the value the human created. A seat permitted the work — the access controls were attached to the seat, and revoking the seat revoked the access. A seat located the accountability for the work — if a transaction went wrong, an auditor could trace it to a human who signed in, and that human was inside an organization, and that organization could answer for the action.

The bundle worked because the executor was always human. The pricing meter and the access right and the accountability locus could share one anchor because the anchor existed in the world — a person, with a manager, inside a company, subject to a contract. The seat was not a technical artifact. It was a legal one, wrapped in a technical interface. The industry priced humans because humans were what executed the work.

Agents are not humans. Agents do not have managers. Agents do not sign employment contracts. Agents can execute thousands of transactions per hour across a dozen vendor surfaces, none of which the agent is "logged into" in any sense the seat model anticipated. The bundle has nothing to anchor on. The three things the seat used to hold together are now floating free of each other, and each is being re-anchored — by different vendors, in different contracts, at different speeds.

What changed in the last six months

Three separations have arrived together. Each is being pushed forward by a different actor with a different incentive, which is why they look like unrelated trends and are not.

The first is commercial. The vendors have moved from per-seat pricing to per-work-unit pricing because that is where the value moved. Salesforce calls the unit an Agentic Work Unit. Microsoft calls it a Copilot Credit. ServiceNow describes operational actions through its Action Fabric. The semantics differ. The structural move is identical. Each vendor has installed a second meter on top of the seat, and each is calibrating it to capture the value that used to flow through the seat. The 19 trillion tokens Salesforce processed in FY26 represent the volume. The 2.4 billion AWUs represent the value. The ratio between them — roughly 8,000 tokens per AWU — is the first time the unit economics of agentic execution has been a public number.

The second separation is operational. The vendors have also begun claiming that they govern the agent's actions, not merely meter them. Salesforce ships at least four MCP servers now. Databricks renamed its AI Gateway to Unity AI Gateway in April 2026 and described it, in their own materials, as "the enterprise control plane that governs access and monitors activity across MCP servers and LLM endpoints." Each vendor is asserting that the operational decision — whether the agent's action should proceed — is governed inside its own walls. This claim is also true within each surface and false across surfaces. The buyer who deploys agents across Salesforce, Databricks, and Workday now has three vendor-walled operational meters and no cross-vendor adjudicator. The decision layer has unbundled from the pricing layer, and no single vendor is positioned to put it back together because no single vendor is the one the enterprise can trust to govern its competitors' surfaces.

The third separation is the one nobody on the vendor side is in a hurry to discuss. The accountability locus has nowhere to land. When an agent executes a transaction that turns out to be wrong — a PII write to the wrong record, a refund issued against the wrong account, a contract clause altered without authority — the auditor's question is who approved it. The vendor will say its surface honored the access controls attached to the principal. The IdP will say it authenticated the principal correctly. The policy engine will say the agent was within policy. Each will be technically correct. None will be sufficient. The thing that used to answer the auditor's question was the seat, because the seat located a human inside an organization with a contract. The seat is the thing that has just been removed from the chain.

The seat was not a technical artifact. It was a legal one, wrapped in a technical interface.

The principle: three meters, one liability

What is emerging is a three-plane architecture where one plane used to suffice. Each plane has its own meter, its own buyer, and its own staleness tolerance. None of them is currently bound to the others by a contract that survives a regulator's scrutiny.

The first plane has emerged because the vendors needed it to emerge. Agentic work units, copilot credits, action-fabric meters — these exist because Salesforce, Microsoft, and ServiceNow have a commercial reason to define them and the platform power to enforce the definition. The CFO can read the new pricing. It is on the invoice.

The second plane is emerging more slowly because the vendors who claim it have an incentive to define it narrowly. A control plane that governs only Salesforce surfaces serves Salesforce. A control plane that governs only Databricks surfaces serves Databricks. A cross-vendor operational control plane serves the enterprise and serves no vendor — which is exactly why no vendor is racing to build it.

The third plane has not yet emerged at all. The accountability locus is the open question, and it is the one the EU AI Act, state-level US legislation, and sectoral regulators in finance and healthcare are about to answer for the industry whether the industry is ready or not. The default answer, in the absence of an alternative, is that the enterprise carries the liability — because the enterprise is the named party in the regulation. The vendor is not. The agent is not. The locus defaults to the buyer because regulators are not in the business of suing software.

What this means at procurement

The procurement conversation has not caught up to the unbundling, and it is creating exposure that boards have not yet priced. Three questions need to be on the table for every agentic-capable vendor renewal in the next twelve months, and each one belongs to a different officer.

The CFO's question is the visible one. What is the meter, is the unit defined transparently, can failed work be excluded, are the rate cards fixed for the term, can usage be capped per agent or per workflow. If the agent reduces human seats elsewhere in the organization, does the commercial model adjust accordingly, or does the enterprise pay twice for the same work. These questions are answerable but they are not always being asked. The temptation in 2026 procurement is to treat agentic credits as a small line item to be revisited later. The next renewal is when "later" arrives, with usage already embedded and no leverage to renegotiate.

The CDO's question is the harder one. Who decides whether the agent's action proceeds. If the answer is the vendor — and for any single-surface action, it is — then the CDO is accepting that the operational meter is vendor-walled. That is acceptable for a Salesforce-only workflow. It is not acceptable for a workflow that touches three vendors, because no single vendor can govern across the other two. The CDO who signs a procurement contract without addressing the cross-vendor case is signing an architectural commitment by silence. The agent will eventually act across surfaces. There will eventually be a wrong action. The architecture will not have an answer.

The General Counsel's question is the one nobody is asking out loud. When the wrong action happens, what is the evidence the enterprise will produce. The seat used to produce that evidence by default. The agent does not. In the absence of a confidence-scored, multi-source-witnessed, per-transaction audit log emitted by a control plane the enterprise owns, the GC is going to discover that the regulator's question — who approved this — has no satisfying answer. The vendor logs are partial. The IdP logs are partial. The application logs are partial. None of them is the audit trail a regulator will find convincing because none of them adjudicated the decision. They each witnessed a fragment.

The closing observation

The Informatica acquisition is the part of the Salesforce story most easily overlooked, and it is the one that most decisively confirms the thesis. Salesforce did not buy a CRM company. Salesforce bought MDM, data catalog, governance, quality, and metadata management — every category the industry has spent three decades treating as "the answer" to enterprise data control. The largest pure-play SaaS company in the world has decided that the substrate above the source systems is now the strategic ground. The acquisition is an admission, from inside the wedge, that vendor-walled governance is not sufficient and that the platform that controls the cross-vendor substrate controls the next decade.

Salesforce will attempt to own that substrate inside its own walls. So will Databricks, with Unity AI Gateway. So will Microsoft, with the agentic surface of Microsoft 365 and Azure. Each will build a vertical control plane that is excellent within its surface and structurally incapable of governing across surfaces. The enterprise will be left holding three meters, three operational planes, and one accountability locus — and the locus will be the enterprise's own, by default, because the regulators have no one else to ask.

The control plane the agentic era requires is the one that binds the three planes back together. Not the vendor's pricing meter — the enterprise's pricing visibility. Not the vendor's operational gate — the enterprise's cross-vendor adjudicator. Not the vendor's audit log — the enterprise's confidence-scored evidence trail. The seat was a bundle. The agent unbundled it. The work of the next decade is to bundle it back together at the layer above the vendors, owned by the buyer, defensible to the regulator. That layer is what governance was always supposed to be.

Three meters. Three buyers. One liability. The seat used to bundle them. The control plane has to.