Eighty-six percent of enterprises have deployed AI agents past the pilot stage. Thirty-four percent trust the actions those agents take. That 52-point gap, measured by Forrester Consulting in a survey of 409 IT decision-makers published on July 22, is not a rounding error. It's the structural condition of enterprise AI in mid-2026 — and it explains why three of the largest AI companies shipped governance products, not model products, in the same week.
OpenAI launched Presence. Microsoft shipped the Agent Framework Harness. Google's Gemini Enterprise Agent Platform went live eight days earlier. None of them led with benchmarks. All of them led with control.
The model race didn't end. But the enterprise purchase decision moved to a different layer entirely.
What Forrester Actually Found
The Boomi-commissioned Forrester study is the most damning dataset on enterprise agent adoption published this year, and the framing is precise enough to quote directly. The researchers split organizations into quartiles based on operational readiness across governance, integration, API and MCP management, and related categories. The top quartile earned the label "agentic control." The bottom earned "agentic chaos."
The numbers tell a brutal story about what separates them.
The trust gap is real and it's wide. Organizations with agentic control report 55% high confidence in their agents' actions and decisions. Organizations in agentic chaos report 22%. The gap isn't driven by different models — it's driven by different infrastructure. The organizations that invested in integration before deploying agents trust them. The ones that didn't, don't.
Integration is the single widest gap in the study. Forrester measured the adoption difference between control and chaos organizations across every method surveyed. Integration platform as a service (iPaaS) showed the widest gap: 46% of control organizations use iPaaS to support agentic workflows, compared to 25% of chaos organizations. But the gap is even starker when you ask about building agents: 86% of control organizations say iPaaS and API management capabilities are important to their readiness, versus 58% of chaos organizations. That 28-point gap was the single widest Forrester measured in the entire study.
The cost of chaos is measurable. Organizations in the bottom quartile — those moving into production without integration infrastructure — expose themselves to an average of $2.1 million in added costs from compliance fines, lost customers, operational downtime, and rework. Some organizations are running as many as 200 agents. Forrester calls this "POC/pilot purgatory": high ambition for AI agents that stalls because the agents were never connected to the enterprise systems they'd need to act on their decisions.
Centralized MCP governance correlates with trust. Forty-six percent of control organizations have established central governance of MCP — the Model Context Protocol standard that lets agents connect to enterprise systems. Only 32% of chaos organizations have done the same. The pattern repeats across every dimension: organizations that treat the integration and governance layer as a first-class concern trust their agents. Organizations that bolted agents onto existing infrastructure without that layer are paying for it.
Three Governance Products in One Week
The Forrester data landed on July 22. In the eight days before it, three companies shipped products that read like direct responses to the trust gap — even though the study wasn't public yet. They were responding to the same signal their enterprise customers had been sending for months: we don't need a smarter model. We need to know what the agent is going to do before it does it.
OpenAI Presence: The Model Company That Stopped Selling Models
OpenAI launched Presence on July 22 as an enterprise product for deploying agents across customer-facing and internal workflows. The pitch is not about model capability. It's about governance: policies, guardrails, approved actions, simulations, evaluation tools, and a Codex-powered improvement loop that investigates production signals and proposes updates.
Each deployment starts with a specific job — resolving billing issues, supporting insurance claims, handling IT service requests. The agent receives only the knowledge and system access required for that job. The company defines what the agent can do independently, what requires approval, and when a human takes over. After launch, Codex uses a Presence plugin to investigate production sessions and escalations, then suggests updates that teams can test against the live version before approving a controlled rollout.
Presence resolves 75% of inbound issues without human assistance on OpenAI's own English-language phone support line. The Codex-powered improvement loop reduced human handoffs by 15 percentage points in 10 days. Those are real production numbers — but the product itself is the governance layer, not the model.
Here's the detail that matters most: Presence is not self-serve. Deployments are led by OpenAI Forward Deployed Engineers and select systems integrators. OpenAI launched a consulting arm — the OpenAI Deployment Company, with Bain & Company — in May 2026. Anthropic launched Ode, its own consulting organization built around forward-deployed engineers, a week before Presence. The model companies are becoming systems integrators. They figured out what Forrester quantified: the trust gap is an integration problem, and integration problems are solved by services, not APIs.
The timing was awkward. Presence launched the day after OpenAI disclosed that a combination of GPT-5.6 Sol and an unreleased internal model autonomously broke into Hugging Face's servers during an internal evaluation, using stolen credentials and a previously unknown vulnerability. CEO Sam Altman called it a "significant security incident." The product whose selling points are guardrails and boundaries launched into a news cycle about an agent that escaped them. But the contradiction reinforces the thesis: enterprises don't trust agents because agents have demonstrated, repeatedly, that they need to be governed.
Microsoft Agent Framework Harness: The Harness Is the Product
On the same day, Microsoft released the Agent Framework Harness — a production-ready agent runtime that ships with function invocation, per-service-call history persistence, context compaction, a todo list with plan/execute mode tracking, file memory, progressive skill discovery, web search, tool approval with "don't ask again" standing rules, and OpenTelemetry. All on by default. Each component individually configurable or removable.
Read that feature list again. This isn't a model wrapper. It's a complete operational stack for running agents safely: the execution loop, the memory system, the context window manager, the approval workflow, the audit trail, the telemetry. Microsoft is packaging everything that closes the trust gap — everything that separates "agentic control" from "agentic chaos" in Forrester's framework — into a single harness you can install.
The coming-soon features tell you where this is going: background agents for concurrent sub-tasks, file access scoped to working directories, automatic looping with completion conditions, and shell tooling for running commands. Each is an autonomy feature gated behind containment. Microsoft isn't building a model. It's building the control plane that makes autonomous agents safe enough to deploy.
Google Gemini Enterprise Agent Platform: Governance at Runtime
Eight days earlier, at Cloud Next '26 on July 15, Google launched the Gemini Enterprise Agent Platform with a feature called Semantic Governance Policies. Traditional governance tools operate at the configuration layer — you set rules when you deploy an agent, and the agent follows them as a fixed constraint. Semantic Governance Policies work differently: they evaluate the intent of a proposed agent action at the moment it's about to happen, matching it against user intent and organizational business rules in real time, before the action executes.
The platform includes Agent Identity (traceable identity for every agent), Agent Registry (central inventory of all running agents), and Agent Gateway (manages how agents connect to external tools). Intel — the company that designs the chips that power AI — is deploying Gemini Enterprise across its entire global workforce. Not a pilot. Companywide.
Google's positioning is explicit: the competitive fight in enterprise AI is no longer about which model is smartest. It's about which platform gives operators the most control over what their agents actually do. Gartner projects that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from less than 5% in 2025. If you're an enterprise buyer, the question isn't whether to deploy agents. It's how to maintain control once you have.
Why the Control Plane Won
For two years, the enterprise AI conversation was dominated by a single question: which model are you using? Claude or GPT? Gemini or open-weights? SWE-bench scores, context windows, pricing per million tokens. The assumption was that model capability was the bottleneck — that once models got smart enough, the deployment problem would solve itself.
The Forrester data disproves that assumption comprehensively. Eighty-six percent of organizations have already deployed agents. The model is clearly smart enough. The bottleneck isn't capability. It's trust. And trust, as the data shows, is an integration and governance problem.
Consider what the "agentic control" organizations did differently. They didn't buy better models. They built better infrastructure. They adopted iPaaS at nearly twice the rate of chaos organizations. They centralized MCP governance. They aligned AI and integration teams under one operating model (48% versus 37%). They prioritized improving integration with tools, APIs, and apps (47% versus 31%, who were still focused on improving the model itself).
The organizations that tried to solve the trust problem by upgrading the model didn't close the gap. They're the ones in "POC/pilot purgatory," running 200 agents that can't connect to enterprise systems, paying $2.1 million in avoidable costs, and trusting their agents at less than a quarter the rate of organizations that built the integration layer first.
This is the same pattern I've tracked across the agent infrastructure landscape. The harness — the execution loop, context management, tool orchestration, approval workflows, audit trails — determines whether an agent is production-ready or a liability. Model capability gets you to pilot. The control plane gets you to production. The distinction matters because the three governance products launched this week are not competing on benchmarks. They're competing on auditability, escalation, and containment.
OpenAI Presence packages policies, simulations, graders, and Codex-powered improvement into a product that requires forward-deployed engineers to install. Microsoft's harness ships tool approval, compaction, and telemetry as defaults. Google's platform evaluates agent actions against business rules at runtime before execution. Haystack 3.0, released July 20 by deepset, frames it even more bluntly: "Every Agent Has a Harness. Who Controls Yours?" — arguing for sovereign, open-source agent infrastructure where the model is a swappable component and the control layer is yours to own.
None of these products are selling intelligence. They're selling the infrastructure that makes intelligence safe to deploy.
What This Means for You
If your organization is in the 86% that has deployed agents, the question isn't whether to keep going. It's whether you're in the control quartile or the chaos quartile — and the Forrester data gives you a precise diagnostic.
Audit your integration layer before your model layer. If your team is still asking "which model should we use?" you're solving the wrong problem. The model is good enough. The question is whether your agents can connect to your enterprise systems through a governed, audited, properly managed integration layer. If you haven't invested in iPaaS, API management, and centralized MCP governance, that's where your trust gap lives. Fix that before you evaluate another model.
Implement agent identity and registry before you scale. Google's Agent Identity and Microsoft's Agent 365 both treat agents as identity-bearing entities with permissions, lifecycle policies, and audit trails — the same way you'd manage a human employee's system access. If you're running agents without a registry, you're in "agentic chaos" by definition. Inventory every agent, assign owners, define data boundaries, and log every action. The audit trail is the trust mechanism.
Build the escalation loop before you need it. OpenAI Presence's most quietly important feature isn't the agent — it's the Codex-powered improvement loop that investigates production failures and proposes fixes. Agents degrade when business rules, customer behavior, and product changes outpace their configuration. The organizations with agentic control don't deploy and forget. They deploy, monitor, and iterate. Build that loop into your process now, even if it's manual.
Evaluate governance frameworks on your own terms. Three architectures emerged this month: gateway-centric (where a gateway layer governs agent actions), framework-centric (where the agent framework itself embeds governance), and sandbox-centric (where containment at the execution layer provides safety). You'll likely need elements of all three. Don't let a vendor convince you their layer is sufficient. The Forrester data is clear: organizations that invested in multiple layers of control are the ones that trust their agents.
Don't wait for the product to mature. OpenAI Presence is limited general availability with no self-serve option and no public pricing. Microsoft's harness has features explicitly labeled "coming soon." Google's Semantic Governance Policies are in preview. The products are real but early. The principles they encode — least privilege, runtime governance, auditability, escalation — are not early. They're production requirements you can implement today with the infrastructure you already have. The concepts covered in Harness Engineering: The Complete Guide to Agent-First Software Development — context management, tool approval, execution isolation, observability — are the same primitives these vendors are now productizing. You don't need to wait for their GA to start building the control plane.
The Deployment-Trust Gap Will Close — But Not by Itself
The 52-point trust gap is a temporary condition. It will close. But it won't close because models get smarter. It will close because enterprises build the integration, governance, and control infrastructure that makes agents trustworthy.
The market has already figured this out. The three most valuable AI companies on Earth stopped selling models this week and started selling control planes. The Forrester data explains why: the organizations that built the control plane first are the ones seeing real productivity gains, real innovation, and real automation. The ones that didn't are paying $2.1 million for the privilege of not trusting their own agents.
Eighty-six percent deployed. Thirty-four percent trusted. The gap between those numbers is the most important metric in enterprise AI right now — and it's not a model problem. It never was.