The constraint is not your agents

After an unforgettable Salesforce Dreamforce in 2026, many enterprises left with the wrong question on their mind.

Leaders are asking which agents to deploy, and how quickly, and they are interpreting the slow return on the ones they run today as a sign that the technology is not yet ready. It is a reasonable perspective, but it is the wrong one. The agents are not the reason the returns are thin. The reason is that capability was bought and partly deployed faster than the system beneath it was built, and no agent, however capable, can perform above the context it is given. The constraint has never been the agent. It is everything the agent depends on and does not yet have.

The UI is no longer the product
The most significant shift at Dreamforce was not a product announcement but the relocation of value beneath the agent. Since Agentforce launched in 2024, Salesforce has presented the agent as the breakthrough, intelligence that no longer waits in a chat window to answer questions but acts on the business directly. This year the emphasis moved to the data the agent reasons over, the permissions it inherits, the governance that contains it, and the operating discipline required to run a business through it rather than merely demonstrate one. The interface has come loose from the system of record, so work no longer has to begin in a single screen, and the screen is no longer the asset worth owning. What matters now is the layer beneath it, the connected context and controlled action that stay scarce once models and interfaces are everywhere, and Salesforce has built its future around that shift.

Salesforce has read this correctly and is building for it, with Data Cloud becoming Data 360 and absorbing the eight-billion-dollar Informatica acquisition, which brings cataloging, integration, quality, governance, and master data management into the single layer every agent draws on. MuleSoft’s Agent Fabric registers and governs agents running across Salesforce, Microsoft, AWS, and Google, so oversight no longer stops at one vendor’s boundary, and the Enterprise AI Harness treats permissions, audit, and policy as runtime infrastructure rather than documentation. Taken together this is less a feature release than a platform being rebuilt around the constraint this piece describes, and it gives an enterprise running Salesforce a coherent set of materials for the work that matters most.

Context sets a ceiling that capability cannot raise
An agent’s output is a direct function of the context it is given, and context is a richer thing than data, being resolved identity, current permissions, honored consent, and live process state, assembled into something the agent can reason over and act on. Most organizations hold data rather than context, scattered across systems, in incompatible shapes, and governed by entitlements no one has audited since the last reorganization, and that gap sets a limit best described as the context ceiling – the level of performance an enterprise’s data, permissions, and processes will support, which no additional agent can raise. Adding ten more agents beneath a low ceiling simply produces ten more things performing beneath it, because the ceiling rises only when the context under it is rebuilt, and rebuilding context, slow and unglamorous as it is, is where the returns are made. This is the difference between adding capability and building the system it runs on, since adding is fast and visible while building is neither until the point at which it becomes decisive.

Agents are running ahead of their data
Evidence that this is already happening is readily available, and IBM’s State of Salesforce study for 2026 to 2027 reports that 83% of organizations are now piloting, implementing, or operating agent orchestration while 9% have optimized it, a distance that measures the gap between adding capability and building the system it runs on, and one in which almost every enterprise currently stands. The same study a year earlier, for 2025 to 2026, found more than half of organizations naming poor data availability and quality as the leading barrier to agentic AI adoption, which is the context ceiling described in less technical terms. Gartner forecast in 2025 that more than 40% of agentic AI projects will be canceled by the end of 2027 on grounds of cost, unclear value, or weak controls, and a project canceled for unclear value is rarely a failure of the agent so much as capability that was added and never built into anything that could return.

Operating models were built for people, not agents
Capability strands because to date the operating model around it has not changed, and an agent that runs continuously, holds a goal, and returns to a person at set points is not software that a finance cycle, an attribution window, or a departmental plan was designed to hold. The organization usually feels this before it can articulate it, having licensed agents across service, sales, and supply chain while being unable to say with confidence which of them touched a given customer last week, under whose authority, or against which target, and in that position, it is paying for capability it struggles to know how to operate. The gap is an operating one rather than a technological one, and it is the one the platform cannot close on the enterprise’s behalf, because the questions it raises are questions of judgment rather than configuration: what a machine is permitted to decide, at what cost of error, under what oversight, and against what target. Where those questions are answered the tooling follows, and where they are left unanswered the agent pursues the wrong goal efficiently and at scale.

An earlier WPP Enterprise Solutions article this year reached the same conclusion from the content and experience side within the Adobe stack, which suggests the divide is a universal AI challenge rather than one unique to a specific platform.

The ceiling takes a different shape in every industry
The context ceiling takes a specific shape in each industry, and that shape tells the enterprise what it must build before capability can pay off.

In retail the ceiling is machine-readable product data, because agentic commerce now completes a purchase inside a search engine or an assistant rather than on the brand’s own site, so a merchant whose catalog an agent cannot read is absent from the transaction, and no agent of its own closes that gap.

In consumer goods the ceiling is the taxonomy and metadata beneath the content, since global-to-local adaptation generates thousands of variants a cycle, and without disciplined structure and clean feedback those variants erode the brand at the same speed they scale it, which makes the curation rather than the generation the work that pays.

In financial services the ceiling is consent and purpose limitation, because data gathered for compliance was never permissioned for growth, so an agent acting at speed on it raises a governance question before a productivity one, and the firms that can prove what an agent did, and why, will move faster than those that cannot.

In healthcare the ceiling is the compliance cycle, since medical, legal, and regulatory approval is the operating model rather than a process wrapped around it, and a process that runs to weeks compresses only when the approval, audit, and consent architecture beneath it is rebuilt, not when an agent is pointed at the old one.

In automotive the ceiling is the trust architecture at the boundary between manufacturer, national sales company, and dealer, where the question of who may act on whose customer has gone unresolved for twenty years and no platform setting resolves it, so an agent acting across that boundary sharpens the problem rather than settling it.

Across all five industries the mechanism is the same, with the ceiling set by the context the agent was never given and the work lying in building that context beneath the agent rather than buying more capability above it.

Advantage comes from building the context beneath the agent
The remedy lies not in a better agent but in treating the context beneath the agent as the thing being built and building it first. For an organization on Salesforce the platform now supplies much of what that requires, while what it cannot supply is the decision about what to build with it and the discipline to run it. Assess data readiness and agent readiness as a single question, scale in step with the context that can support the load rather than ahead of it, and put the governance, the permissions, and the operating model in place as the agent goes live rather than after it fails in front of a customer. Build the system once, connect it end to end, and operate it as a standing capability rather than a delivered project, because adoption, activation, and operation are the real work and the agent is only the last mile.

Two enterprises will run the same agents this year, one having built the context beneath them and gaining a little more from every action those agents take, the other having added capability to a ceiling it never raised and continuing to buy while wondering why the returns stay flat. The gap will not look large at first, because the return on context is quiet until it becomes decisive, but the advantage it creates is not bought so much as built beneath what was bought and then operated, which is why growth follows the system.

WPP Enterprise Solutions designs, builds, and operates the growth systems competitive businesses rely on, helping clients turn intent into execution worldwide. Our Salesforce practice brings more than 1,000 certified experts, over 2,500 implementations, and more than fifteen years of partnership across Sales, Service, Marketing, Data, Commerce, Loyalty, Agentforce, MuleSoft, and Einstein. We run Salesforce as one connected growth system, combining data foundations, governance, and operating-model discipline to turn agentic ambition into execution. The Dreamforce announcements were the easy part. The build, and the operation that follows, is where advantage lives.