Your Company Doesn’t Need More AI Agents

The $150,000-Agent Problem: Why More AI Agents Is Making Your Organization Dumber

The AI Agent Inventory That Will Keep You Up at Night

Here is a question I want you to answer right now, before you read another word.

How many AI agents is your organization currently running?

Not how many you approved. Not how many are in the IT ticket system. How many are actually running — across every team, in every function, touching your data and your systems and your customers?

If you paused before answering, you’re in good company.

Salesforce’s 2026 Connectivity Benchmark surveyed 1,050 IT leaders and found that the average enterprise is now running 12 AI agents — a number projected to hit 20 within two years. But here’s the part nobody mentions in the vendor keynotes: 50% of those agents operate in isolated silos with no coordination, no shared context, and no unified governance. And 27% of the APIs connecting those agents are completely ungoverned — no audit trail, no access controls, no compliance checks.

Gartner projects that by 2028, a typical Fortune 500 company will manage more than 150,000 AI agents. Only 13% of organizations believe they have the governance infrastructure to handle that.

From 15 agents in 2025 to 150,000 agents by 2028. Without governance.

That is not a technology roadmap. That is a slow-motion organizational crisis being dressed up as progress.

And the most dangerous part? Most of the agents making it worse were built by well-meaning people trying to solve real problems.

Photo by Marvin Meyer on Unsplash

How Every Company Ends Up With an Agent Sprawl Problem

Let me tell you how this happens. Not in theory. In practice.

A data analyst on the marketing team gets access to a low-code AI agent builder. She’s been manually pulling weekly performance reports for 18 months — three hours every Friday. She builds an agent in two afternoons that does the same job in four minutes. Genuine win. She tells two colleagues. They build their own agents.

Meanwhile, the sales ops team deploys an agent for lead scoring. Customer success builds one for churn prediction. Finance automates invoice processing. HR sets up an agent that screens resumes and sends initial outreach to candidates.

Every single one of those deployments was a legitimate business decision. Every single one solved a real problem. None of them were coordinated. None of them share a governance framework. Several are connecting to the same customer data through different API keys with different permission scopes. Two are producing conflicting outputs about the same customer — but nobody knows, because the outputs land in different dashboards that nobody has thought to compare.

According to the AI Incidents Database, reported AI-related incidents rose 21% from 2024 to 2025. That number almost certainly understates the actual exposure — because most organizations have no incident classification that captures an autonomous agent action as the root cause. The incident gets logged as a service restart, a latency event, a data discrepancy. The agent is invisible in the postmortem.

This is agent sprawl. And it’s not a technology failure. It’s a governance vacuum that technology rushed into.

The Two Wrong Lessons Most Companies Are Drawing Right Now

When executives hear about agent sprawl, the instinct usually splits into one of two directions.

Instinct A: Lock it down. Stop approving new agent deployments. Require IT sign-off on everything. Make the process harder.

Instinct B: Buy an orchestration platform. Find the right vendor, centralize everything on their stack, and let them solve the coordination problem.

Both are wrong. Not entirely — there are grains of truth in each. But they’re wrong about where the real problem lives.

Locking things down without providing a better alternative just drives deployment underground. This is exactly what happened with shadow IT in the 2010s. When companies told employees they couldn’t use Dropbox, employees used personal accounts and bypassed IT entirely. The risk didn’t go away — it became invisible. Agent sprawl is shadow IT’s more dangerous successor. A shadow SaaS tool is a passive risk. It stores data it shouldn’t store; the data sits there until someone reads it. A shadow agent is an active risk. It reads, decides, and writes. It doesn’t wait for a human to look at it before taking an action.

Buying an orchestration platform is the right answer to the wrong question. The right question isn’t “what tool can coordinate our agents?” It’s “what business outcomes are we trying to achieve with agents, and what’s the minimum agent architecture that achieves them reliably?” No orchestration platform answers that question for you. It can only help you manage the architecture you’ve already decided on.

The Wall Street Journal recently reported that companies including Lyft, DaVita, GitLab, and FICO are actively wrestling with how to control agent duplication, conflicting outputs, cybersecurity risk, and rising compute costs from proliferating agents. These aren’t laggard organizations. They’re among the most technically sophisticated companies in the world — dealing with the same problem everyone else is, because the problem isn’t a technical failure. It’s a strategy failure that technical decisions are now making visible.

Sreenivas Vemulapalli, Senior VP and Chief Architect of Enterprise AI at Bridgenext, stated it plainly: the strategic value for an enterprise lies not in building the agent’s brain or the plumbing that connects it, but in defining and standardizing the tools those agents use. The true competitive advantage belongs to enterprises that have meticulously documented, secured, and exposed their proprietary business logic as high-quality, agent-callable APIs.

Translation: the agents themselves are becoming a commodity. What isn’t a commodity is the decision about which workflows should be agentic at all.

Your Company Doesn’t Have an Agent Problem. It Has a Workflow Strategy Problem.

Here is the reframe I want you to carry out of this article.

Every company struggling with agent sprawl built their way into it the same way: bottom-up, use-case by use-case, tool by tool, without ever answering the upstream question.

The upstream question is not “what can we automate with agents?” It’s “what is our theory of how AI agents should create competitive advantage in this specific organization?”

That sounds abstract. Let me make it concrete.

A company with a clear workflow strategy for agentic AI can answer three questions with specificity:

Which workflows are worth making agentic? Not every task that can be automated should be. The workflows worth making agentic are high-frequency, rule-bound enough for agents to execute reliably, consequential enough to justify the governance overhead, and strategic enough that faster execution actually moves a business outcome. A workflow that takes two hours and happens once a month probably doesn’t meet that bar. A workflow that takes thirty minutes and runs four hundred times a day across the sales organization probably does.

What is the governance model before the first agent deploys? Who can create agents? What data sources can they access? What actions can they take without human approval — and what actions require it regardless of confidence score? Who is accountable when an agent produces a wrong output that gets acted on? These aren’t IT questions. They’re business questions that IT needs to enforce technically.

What does the agent inventory look like, and who owns it? Every agent in production should have an owner, a documented purpose, a defined scope, an audit trail, and a retirement date or review cadence. Exactly like employees. An agent with no owner is not a productivity tool. It’s a liability.

The companies doing agentic AI well in 2026 — the ones seeing the 5.8x ROI that McKinsey identified in top performers — are not the ones with the most agents. They’re the ones with the fewest agents necessary to achieve the outcome: each one governed, each one auditable, each one tied to a measurable business result.

More agents is not progress. More governed outcomes per agent is progress.

The GOSA Test: The Four Questions to Ask Before Your Next Agent Deployment

Before your organization approves another agent, run this audit. I call it the GOSA framework — Governance, Outcome, Scope, Accountability.

G — Governance. Does this agent have a defined owner? Can it be paused or rolled back immediately if something goes wrong? Is every action it takes logged with enough context to reconstruct what happened and why? If any answer is no, the agent is not ready to deploy.

O — Outcome. What specific, measurable business outcome does this agent improve? Not “it saves time.” What is the baseline, and what improvement are you targeting? If you cannot name a number, you are deploying for activity, not results.

S — Scope. What is the explicit list of things this agent is permitted to do? What is the explicit list of things it is not permitted to do? What actions require human approval before execution? Scope should be written down, reviewed by someone outside the deployment team, and enforced technically — not just assumed.

A — Accountability. When this agent makes a mistake — and it will make mistakes — who is responsible? What is the remediation path? Is there a kill switch, and does someone other than the agent’s builder know how to use it?

An agent that passes all four tests is worth deploying. An agent that fails any one of them is a liability dressed as a productivity tool.

Most organizations would fail this audit on the majority of their currently running agents. That’s not an accusation. It’s an invitation to run the inventory before a governance failure forces the conversation.

What Good Agent Architecture Actually Looks Like

Let me give you the positive case, because I don’t want to leave you only with the warning.

The organizations building agentic AI well in 2026 follow a recognizable pattern. They started by mapping their highest-value, highest-frequency workflows — not by cataloging everything that could theoretically be automated, but by asking where agent-speed execution creates the clearest competitive advantage.

They built governance infrastructure first: a centralized agent registry, a standardized permission model, a clear escalation protocol for actions that exceed an agent’s authorized scope. They defined agent-callable APIs — documented, secured interfaces into their core business logic that any agent can access, rather than each agent building its own connection to the same underlying data.

Then they deployed agents minimally. Not maximally. The goal was never to have the most agents. It was to have the right agents — fully governed, producing auditable outcomes against defined baselines.

The results are real. Gartner found that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from fewer than 5% in 2025. But the organizations capturing value from that shift are not the ones deploying fastest. They’re the ones deploying most deliberately.

MIT Technology Review’s analysis framed it precisely: autonomous agents will soon run thousands of enterprise workflows, and only organizations with unified, trusted, context-rich data infrastructure will prevent chaos and unlock reliable value at scale. Unified, trusted, context-rich. That describes a data strategy and a governance architecture — not an agent count.

My Prediction: The Agent Consolidation Wave Is 18 Months Away

By the end of 2027, the companies currently racing to deploy the most agents will be running agent consolidation programs — auditing, decommissioning, and restructuring their agent portfolios the same way enterprises ran app consolidation programs after the SaaS sprawl of the 2010s.

The consolidation will be triggered by one or more of three events: a governance failure that creates a compliance exposure, a cost reckoning when inference bills come due at scale, or a competitive realization that a smaller organization with better-governed agents is executing faster with less overhead.

86% of IT leaders surveyed by Salesforce already say they worry that agents add more complexity than value due to integration failures. That worry becomes a budget conversation within 18 months.

The companies that position now — that run the agent inventory, build the governance layer, and deploy deliberately rather than prolifically — will not need a consolidation program. They’ll already be on the other side of the problem their competitors are about to discover.

Gartner projects 150,000 agents per Fortune 500 company by 2028. Only 13% of organizations are prepared to govern that.

The question is not whether your organization will have more agents two years from now. You will. The question is whether you’ll know what they’re doing — and whether anyone will be accountable when they get it wrong.

Build the governance layer first. Then deploy the agents.