Agentic AI Sprawl: Why 94% of Enterprises Now See Their Own Agents as a Risk

Mainstream Adoption, Uncontrolled Growth

In just eighteen months, agentic AI moved from proof-of-concept to production at near-universal scale. The OutSystems 2026 State of AI Development report, based on responses from nearly 1,900 global IT leaders surveyed between December 2025 and January 2026, reports that 97% of enterprises are exploring system-wide agentic AI strategies and 96% are already using AI agents in some capacity. Adoption is not the question anymore. Governance is.

The same survey found that 94% of organizations report concern that AI sprawl is increasing complexity, technical debt, and security risk. That is not a fringe worry — it is the near-unanimous view of the IT leaders running the deployments. When an emerging technology is adopted faster than the controls around it, the risk surface grows on every axis at once: data access, identity, auditability, cost, and model behavior.

What Sprawl Actually Looks Like

The sprawl problem has a specific shape. According to SD Times’ coverage of the report, 38% of organizations globally report mixing custom-built and pre-built agents, creating AI stacks that are difficult to standardize and secure. Different business units select different platforms; a marketing team builds agents on one vendor, finance on another, engineering spins up its own framework. Each stack has its own authentication model, logging pipeline, and approval flow — or none at all.

The result is a proliferating attack surface. Every agent with credentials to a production system is a potential path for privilege escalation, prompt injection, or data exfiltration. Every unmonitored model call is a cost line item that no one owns. And because agents increasingly call other agents, a single compromised identity can cascade through multi-step workflows that no single person architected.

The Governance Gap

The gap between adoption and control is the clearest number in the report: only 12% of enterprises have implemented a centralized platform to manage sprawl. Most are still experimenting with governance approaches that vary by team and region. The remaining 88% are running agentic AI with ad-hoc oversight — some combination of spreadsheet inventories, team-level policies, and retroactive security reviews.

That asymmetry explains why the same survey shows high deployment confidence alongside high sprawl anxiety. IT leaders believe in the technology’s value; they are losing visibility into how it is actually being used. In coverage in PR Newswire’s release, OutSystems researchers describe the current state as adoption running years ahead of observability.

Why Standard Cloud Governance Doesn’t Fit

Enterprises have been through platform sprawl before — with SaaS apps, with shadow IT, with microservices. But agentic AI has properties that make traditional governance harder to apply:

• Non-determinism. The same input can produce different outputs. Testing and audit frameworks built for deterministic systems don’t cleanly capture agent behavior.
• Compound identities. An agent can act on behalf of a user, then spawn sub-agents that act on behalf of it. Identity-and-access-management systems were not designed for these delegation chains.
• Cost elasticity. A single malformed prompt can trigger thousands of API calls. Traditional FinOps dashboards report the bill weeks after the damage.
• Rapid vendor churn. The model, framework, and orchestration layers are all moving every quarter. Governance policy written for last quarter’s stack is already out of date.

Standard cloud governance — IAM, network policy, budget alerts — catches some of this, but not enough. The 12% who have built centralized platforms are investing in agent-specific controls: model gateways, policy engines that enforce rules on agent invocations, and observability stacks designed to trace multi-agent call chains end to end.

What “Getting Ahead of Sprawl” Looks Like

The IT leaders in the 12% with centralized platforms share a few common moves:

1. One registry, not many. A single inventory of every agent running in production, who owns it, what data it touches, and what actions it can take.
2. Policy at the gateway. Agent calls flow through a policy layer that enforces rate limits, permitted tools, and data-classification rules — before the agent reaches the model.
3. Unified observability. Logs, traces, and cost metrics for every agent call land in the same system, queryable by engineering, security, and finance.
4. Lifecycle discipline. Every agent has an owner, a review date, and an explicit retirement path. Unused agents are decommissioned, not left running indefinitely.

None of these are technically novel — enterprises have done this for microservices and SaaS. The novelty is treating agents as a first-class governance object, not a feature of whichever platform they happen to run on.

The Bigger Signal

Agentic AI has crossed the adoption threshold. What’s left is the harder work: bringing the control plane up to the scale of the deployment plane. The 94% concerned about sprawl are telling CIOs where to invest next. Enterprises that build centralized agent governance in 2026 will have a structural advantage in 2027 — both in speed of deployment and in audit posture. Those that don’t will spend the next two years unwinding accumulated complexity they could have prevented with a clearer platform decision today.

Sources & Further Reading

• Agentic AI Goes Mainstream, but 94% Raise Concern About Sprawl — Business Wire / OutSystems
• Agentic AI Goes Mainstream, but 94% Raise Concern About Sprawl — SD Times
• OutSystems Release — PR Newswire
• OutSystems 2026 Agentic AI Report Coverage — MarTech Series