The seven tensions of an agentic enterprise and Agentforce (Salesforce)

Over the past few months, Ross and I have spoken with enterprise technology leaders across two large consumer goods firms, a healthcare payer, and a leading global asset management firm. These conversations converged on three themes: Software-as-a-Service (SaaS) consolidation, the path to building an agentic enterprise, and the economics of running agents at scale. They also surfaced a consistent set of unresolved questions, or tensions, that enterprises are increasingly grappling with.

It is in this context that we evaluated Salesforce’s recent AgentExchange announcement and examined how major enterprise platforms are navigating these tensions.

The enterprise software market has always competed on features, pricing, and integration depth. Agentic Artificial Intelligence (AI) introduces a fundamentally different dimension. Salesforce, Microsoft, Google, ServiceNow, and Amazon Web Services (AWS) are all positioning themselves to become the operational core around which an enterprise’s agent strategy is built. Where that core ultimately sits will shape how agents access data, how they are governed, how costs scale, and how easily they interoperate across systems.

Understanding what is at stake requires looking beyond vendor claims. The agentic AI landscape is defined by a set of tensions that no platform has fully resolved and that every enterprise will need to navigate, regardless of which platform it chooses.

Seven tensions defining the agentic enterprise

1. Platform consolidation vs. vendor lock-in

Enterprises want to consolidate AI onto fewer platforms, preferably their systems of record, because those platforms already contain the context agents need to act accurately: customer history, workflow logic, compliance rules, and transactional data accumulated over years. Running agents on a platform that already holds this business context also reduces inference overhead, making inference both cheaper and more reliable on native data sources. However, this consolidation also deepens vendor lock-in in favor of systems of record.

2. Speed vs. trust in agentic deployments

Certifying an agent for enterprise use requires validating its behavior across dynamic real-world conditions, a problem for which no established framework yet exists. Enterprises deploying agents before such standards mature are accepting risks they cannot fully measure. Enterprises waiting for these standards, however, risk ceding competitive advantage to those moving faster.

3. Agent sprawl vs. governance

Agent deployment decisions are often made at the team level, and the speed and distribution of adoption can quickly outpace centralized governance models. By the time oversight catches up, the agent estate is already too large and fragmented to govern retroactively.

4. Interoperability vs. proprietary orchestration

Enterprises want the freedom to build agent workflows that span multiple vendor ecosystems without being locked in. Open standards such as Model Context Protocol (MCP) and Agent-to-Agent (A2A) were supposed to enable this interoperability. However, every major vendor (Salesforce, Microsoft, ServiceNow, and Google, is simultaneously racing to become the system that governs how enterprise agents are deployed, what they can do, and how they hand off tasks to one another. The vendor who wins this orchestration layer effectively controls the enterprise’s broader agent operations. The tension is this: open standards have made agents interoperable, but they have also intensified the competition to govern those agents.

5. The context advantage vs. open ecosystem pressure

A system of record-native agent inherits years of customer history, workflow logic, and relationship data without needing to reconstruct any of it. However, the more capable an agent becomes, the more its value depends on operating across systems outside that platform, including Enterprise Resource Planning (ERP), document management systems, and finance. The tension is this: while the context advantage is deepest within a platform’s own boundaries, an agent’s value compounds when it can operate effectively beyond them.

6. Shift in procurement economics

Marketplaces are attempting to make agents transactable in the same way software products are. However, agents are not discrete products with clearly defined scope, negotiated pricing, or well-defined ownership boundaries. Choosing an agentic AI vendor is therefore not a traditional procurement decision. The vendor becomes a strategic partner whose governance model and long-term ambitions directly influence the reliability of the enterprise’s most critical business processes. The commercial frameworks that govern how software is bought, implemented, and measured were not built for this kind of relationship. Procurement teams are being asked to operate within a distribution model that still lacks established rules for what agents actually are.

7. Inference cost volatility as a strategic risk

Model commoditization was supposed to make enterprise AI more economical. In practice, it has had the opposite effect. As inference costs declined, the business case for deploying agents across more workflows became easier to justify, which expanded overall consumption faster than unit costs declined. Enterprises that moved aggressively on agentic AI are now managing cost structures far beyond what their initial pilot economics anticipated.

How are vendors addressing the tensions

The tensions outlined above are already shaping how vendors structure their agent platforms and marketplaces. The following table summarizes how major vendors are approaching each challenge.

Tension	Software vendors’ approach
Platform consolidation vs. vendor lock-in	Discovery, procurement, activation, and governance are increasingly being pulled consolidated into the same control layer. Microsoft does this through M365 and Azure integration, ServiceNow through its workflow stack, and AWS and Google through runtime and infrastructure layers.
Speed vs. trust	Trust controls are increasingly embedded at deployment and runtime rather than established upfront. Microsoft enforces identity and compliance through Entra and Purview, ServiceNow audits through AI Control Tower, and AWS isolates execution environments.
Agent sprawl vs. governance	Governance is being introduced at entry points such as marketplaces, registries, and builders. Microsoft and Google are building centralized registries, ServiceNow extends governance through workflows, and AWS ties registry controls to infrastructure governance.
Interoperability vs. proprietary orchestration	Integration is opening through protocols such as MCP and A2A, while execution remains anchored within each platform. Google is advancing A2A, ServiceNow supports both A2A and MCP through Agent Fabric, and AWS remains largely model-agnostic.
Context advantage vs. open ecosystem pressure	Vendors are embedding agents directly into the systems where enterprise data and workflows already reside. Salesforce integrates agents into Customer Relationship Management (CRM) and Slack, Microsoft into M365 and Dynamics, and ServiceNow into workflows and Configuration Management Database (CMDB). AWS and Google rely more heavily on Application Programming Interfaces (APIs) and data layers to access enterprise context.
Shift in procurement economics	Marketplaces increasingly combine discovery, contracting, and activation within a single workflow. Microsoft ties agents to marketplace contracts and enterprise agreements, AWS links marketplace activity to infrastructure billing, and Google integrates marketplace capabilities into its broader agent platform.
Inference cost volatility	Cost controls are being addressed primarily at the infrastructure and packaging layers rather than at the agent layer itself. AWS and Google focus on compute pricing and efficiency optimization, while Microsoft and Salesforce bundle use through credits and subscriptions.

AgentExchange and its implications for the agentic enterprise

Against this backdrop, Salesforce’s launch of AgentExchange provides a concrete example of how these tensions are being operationalized.

AgentExchange consolidates AppExchange, Slack Marketplace, and the broader Agentforce ecosystem into a single marketplace where agents, tools, and integrations can be discovered, transacted, and deployed directly within enterprise workflows. The result is a unified platform that brings together over 10,000 Salesforce apps, 2,600 Slack apps, and more than 1,000 pre-built agents, tools, and MCP servers.

A TLDR interpretation of this announcement could be: “hey, this is just a distribution move, a marketplace expansion.” In reality, it is an attempt to pull three critical layers, procurement, activation, and runtime proximity, inside Salesforce’s boundary. If successful, this shifts control away from fragmented, team-level adoption toward a Salesforce-mediated operating model in which agents are discovered, approved, and embedded within its ecosystem by default.

As a system of record, Salesforce faced a strategic risk: if agents are discovered, orchestrated, and executed outside its ecosystem, for example through Microsoft, AWS, or independent frameworks, Salesforce risks becoming merely another system of record that agents call into, rather than the environment in which agents live. AgentExchange is designed to prevent that outcome by making Salesforce the default surface where agents are sourced, activated, and embedded within workflows.

AgentExchange and the seven tensions

Lets see how AgentExchange fairs against the seven tensions we described earlier. Before walking through them, I have reordered the tensions to match how AgentExchange is approaching them.

Shift in procurement economics

This is the tension AgentExchange appears most explicitly designed to address. By unifying procurement, activation, and billing into a single surface, Salesforce imposes structure on a model that currently lacks it. However, the underlying commercial unit being bought is still not well defined: agents are not discrete products with defined scope, clear ownership boundaries, or predictable cost structures. While AgentExchange creates a distribution mechanism, it does not resolve what agents fundamentally are as a commercial entity.

Agent sprawl vs. governance

AgentExchange lowers deployment friction, accelerating the rate at which agents enter the enterprise. The mandatory security review for marketplace listings introduces an important control layer at the point of entry. What is interesting is that agents sourced through AgentExchange enter through a governed channel. However, it remains unclear whether agents built directly on Agentforce or sourced externally can be governed through the same mechanisms. At present, the governance boundary appears tied more closely to the marketplace than to the enterprise’s full agent estate.

Interoperability vs. proprietary orchestration

AgentExchange supports third-party agents, tools, and MCP servers from partners across the ecosystem. However, discovery, activation, and deployment all still route through Salesforce. Open standards operate at the connectivity layer, while control remains concentrated at the orchestration layer. In effect, AgentExchange expands Salesforce’s control over a broader set of third-party capabilities.

Context advantage vs. open ecosystem pressure

AgentExchange reinforces Salesforce-native deployment by pulling more capabilities within its ecosystem boundary. This deepens the native context advantage for enterprises already operating heavily within Salesforce. While it enriches the Salesforce-native environment, it ultimately reinforces platform boundaries rather than lowering them.

Platform consolidation vs. vendor lock-in

By pulling procurement, activation, and runtime proximity within Salesforce’s boundary, AgentExchange simultaneously strengthens consolidation and deepens lock-in. The question for Chief Information Officers (CIOs) is whether the consolidation it enables is worth the dependency it creates.

Speed vs. trust in agentic deployments

AgentExchange’s security review is a listing-level vetting process. It does not change the absence of a deployment-level validation standard. Enterprises activating agents through AgentExchange are still accepting the same unquantified behavioral risk as those activating agents elsewhere.

Inference cost volatility

AgentExchange is a distribution and activation mechanism. It has no bearing on the cost structure of running agents at scale. Flex Credits improve cost visibility but do not address the structural problem: consumption expands faster than unit costs decline at production scale. This challenge sits outside the scope of what AgentExchange is designed to address.

What this means for Salesforce customers

AgentExchange is a well-constructed move. Salesforce has correctly identified a major enterprise pain point (fragmented discovery, procurement, and activation), and built a credible mechanism to address it within its ecosystem. The key question for CIOs is whether the consolidation it enables is worth the dependency it creates.

That same question is being asked simultaneously across every major enterprise platform:

Microsoft is consolidating its marketplace into a single surface through the unified Microsoft Marketplace, while Agent 365, generally available as of May 1, 2026, provides governance and identity layers for agents at scale
Google rebranded Vertex AI as the Gemini Enterprise Agent Platform at Cloud Next 2026, absorbing Agentspace into a unified platform spanning model selection, agent development, governance, and distribution from a single console
ServiceNow has embedded AI, governance, and its Context Engine into every product it sells, effectively ending the era of AI as an add-on capability
AWS launched Agent Registry through AgentCore, providing enterprises with a cross-cloud catalog for discovering and governing agents regardless of where they are built or hosted

Every one of these moves reflects the same underlying competitive logic: the platform that controls how agents are discovered, activated, and governed becomes the operational core of the enterprise’s broader agentic strategy. While the structural moves are similar, the terrain on which each vendor is competing is different.

For Salesforce customers, AgentExchange directly accelerates the path toward agentic AI deployment. The context, governance, and workflow infrastructure agents need already exist within the platform. The decision is how to build on it, with full awareness that every agent sourced through AgentExchange deepens a dependency that compounds over time.

If you enjoyed this blog, check out, Enterprise AI is only as good as its data: Salesforce just proved it – Everest Group Research Portal , which delves deeper into another topic relating to Salesforce.

If you have any questions, please reach out to Abhishek Singh at [email protected] or Ross Tisnovsky at [email protected].