The Operations and Finance Case for Running AI Agents on Pre-Approved Business Surfaces
Approved tools and data limits keep agent costs, actions, and audits traceable.

Rolling out AI agents to thousands of non-technical employees requires pre-approved business surfaces. Those surfaces are the tools, data sources, and workflows that agents can use under access rules the system enforces. They make broad access financially accountable by keeping permissions, actions, and approvals visible, while avoiding the remediation, audit gaps, and stalled pilots that make ungoverned deployment expensive.
Why pre-approved surfaces are a structural requirement, not a governance preference
The Cloud Security Alliance’s 2026 research note describes a stark gap between ambition and readiness. A large majority of enterprises plan to deploy agentic AI within two years, yet only a small fraction have a mature AI-agent governance model. For operations and finance leaders, that mismatch means the commitment to automate is arriving before the ability to account for what automation does.
The Model Context Protocol, or MCP, makes that gap harder to contain within individual projects [1][2][3]. MCP was donated to the Agentic AI Foundation, a directed fund under the Linux Foundation, in December 2025, with backing from OpenAI, Google, Microsoft, AWS, Bloomberg, and Cloudflare. It is cross-vendor infrastructure. The protocol connects agents to file systems, databases, APIs, code repositories, and enterprise SaaS, so every added server expands the surface available to agents.
That expansion changes the deployment decision. Approval must establish that a model is suitable or that a pilot works, as well as which business capabilities the agent can reach and under whose authority. Without that boundary, leaders approve an automation project without a clear limit on its operational reach.
What agents actually do on a surface and why that changes the risk calculus
An agent acts as an MCP client, requesting capabilities from servers that expose tools through structured, discoverable interfaces. Those tools can query databases, call APIs, and trigger workflows. The agent can combine them within one interaction. Unlike a person following a fixed sequence of screens, it composes a sequence of actions as the task unfolds.
A request to “prepare a quarterly report” may lead an agent to read a CRM, access financial databases, fetch external data, generate visualizations, and email the result. Each step has a separate compliance implication, and the model selects those steps at runtime. The request sounds narrow. The execution may not be. That is why the surface is a first-order financial variable.
Prompting can discourage a bad path, but it cannot prohibit one. Strict boundaries need enforcement where the agent accesses tools and data, because conventional security controls were not designed to judge the intent behind an agent’s tool call. The EchoLeak attack, CVE-2025-32711, used a specially crafted email to make Microsoft 365 Copilot silently exfiltrate sensitive user data to an attacker-controlled endpoint without a user click.
The specific MCP attack vectors that make ungoverned surfaces financially dangerous
Tool poisoning begins in a place most users never see. Attackers put malicious instructions inside tool descriptions, which enter the agent’s context and can be treated as authoritative instructions. Invariant Labs demonstrated this in Cursor IDE, where a hidden instruction caused an agent to exfiltrate a user’s SSH private key without any visible indication.
Cross-server tool shadowing extends that problem beyond the malicious tool. Invariant Labs demonstrated how one malicious MCP server could alter an agent’s understanding of adjacent trusted servers. Servers need to be approved with these links in mind, because an untrusted server can change how the agent uses trusted ones.
Credential aggregation creates a different route to broad exposure. MCP servers may hold credentials for multiple enterprise systems, making one compromised server a path into everything it touches. The Drift/Salesforce incident in August 2025 illustrated the credential-theft pattern, although no MCP server was involved. Attackers breached Salesloft’s GitHub account, stole OAuth tokens from Drift’s AWS environment, and used them to access customer Salesforce instances through the API.
Why authentication gaps are the financial exposure point operations leaders underestimate
The MCP authorization specification defines an OAuth 2.1 framework but makes authorization optional. A server can lack client authentication because it was designed that way, as well as because someone set up authentication badly. Trend Micro found 492 MCP servers that were open to the internet and had zero client authentication or traffic encryption.
The Cloud Security Alliance’s research note reports that only a small fraction of MCP implementations use OAuth, leaving identity binding absent from the vast majority of connections. Earlier versions of the authorization specification also used anonymous Dynamic Client Registration, which let clients register without saying who they were. The 2026-07-28 specification deprecated that approach in favor of Client ID Metadata Documents, while retaining it for backward compatibility. Protocol evolution does not substitute for an enforced identity boundary in a deployment.
For finance, identity is the basis of attribution. When an agent cannot be tied to a named user, its actions cannot be reliably assigned to that user’s authority. Excess permissions deepen the problem because agents can operate continuously and autonomously rather than being limited by human attention and working hours. The resulting audit gap is not just missing security detail. It weakens the organization’s ability to explain who authorized an action and reconstruct what happened.
The operational cost of ungoverned deployments that finance leaders rarely see in advance
Governance overhead is a known, bounded cost that can be planned before deployment. Ungoverned incident costs arrive later, without a comparable boundary. The work may include remediation, incident response, rebuilding the record for an audit, and rebuilding a pilot that cannot pass security review. A low-cost launch can become an expensive operating model.
A stalled pilot also costs more than its development budget. Delayed automation postpones the expected return, while re-piloting creates organizational friction and consumes another round of effort. A peer organization that reaches production first can widen the competitive gap during that delay. Pre-approval belongs in the investment calculation for this reason. It affects both the path to production and the expense of protecting production afterward.
Incident reconstruction adds another layer of uncertainty. Because agents compose actions across systems, investigators may need to trace tool calls through every MCP server involved in a session. Without tamper-proof logs, that reconstruction may be impossible. Shadow MCP connections create the same problem before an incident occurs, because security teams cannot audit connections they have never mapped, scoped, or approved.
What the ROI picture looks like when agents operate on governed, bounded surfaces
There is a stronger case for investing in high-volume work with clear limits than in an autonomous employee with an open-ended role. Customer-service resolution, finance back-office automation, and software-engineering assistance offer tasks with verifiable success signals and human review for consequential decisions. Success can be checked against a defined task. That structure makes both performance and mistakes easier to account for.
Agents that resolve contained customer-service tickets can cost a fraction of what it costs for people to handle them, while code-review agents that complete routine pull requests can cost a fraction of the cost of a senior engineer’s time. Those comparisons depend on defined scope and attributable actions. Otherwise, what looks like savings can leave out the work needed to find and fix errors after they spread across systems.
Human review is part of that return model. It provides a place to catch consequential mistakes while allowing routine work to proceed within approved boundaries. The objection that pre-approval slows deployment also needs a production comparison. An ungoverned pilot that stalls in security review consumes more time than a governed one that reaches production on its first attempt.
The regulatory layer that is turning governance from optional to mandatory
The EU AI Act adds legal exposure to the operational costs of deferral. The EU Commission’s enforcement framework states that enforcement powers for certain provisions apply from 2 August 2026. Maximum fines for the most severe violations are set at the higher of a fixed amount or a percentage of global turnover. “Govern later” is no longer just an internal scheduling choice.
From 2 August 2026, traceability and logging obligations require covered systems to record actions automatically. For an agent working across MCP servers, reconstructing behavior requires tool-call-level logs. The tool-calling layer is part of the compliance surface because auditors need the underlying action record to reconstruct agents’ actions.
AIUC-1 Q2 2026, effective April 15, 2026, added dedicated controls for MCP and A2A authentication, transport security, message integrity, and runtime containment. NIST’s AI Agent Standards Initiative, launched in February 2026, is expected to publish an interoperability profile in Q4 2026. These standards efforts, distinct from legal obligations, reinforce the same direction by making protocol-level governance an explicit requirement.
The four governance controls that define a pre-approved surface in practice
For operations and platform teams, the practical response is four controls that make broad deployment possible. The first is a per-agent tool allowlist, declared centrally and enforced. Without it, each MCP server can apply its own authorization logic, allowing scope-filtering rules to diverge across the environment.
The second is OAuth 2.1 identity binding, through which agents inherit scoped user permissions. Every action then has an accountable user behind it. This enables attribution and audit without treating an anonymous agent session as sufficient evidence of authority. It also connects the agent’s permitted work to the permissions of the person it serves.
The third is a centralized MCP gateway that gives security teams one inventory of servers, skills, and agents. With that view, security teams can more easily assess and control growth. The gateway also provides the enforcement point for the allowlist and identity binding. Without a common view, shadow usage can accumulate beyond the approved deployment.
The fourth is mandatory human approval before irreversible actions, including database deletions, financial transactions, and outbound data transfers. That review protects the point where a mistake becomes costly to reverse or cannot be reversed at all. It also permits wider autonomy below that threshold. When consequential actions require sign-off, non-destructive work can have a broader allowlist without leaving every decision to unchecked model judgment.
How can IT enable AI agents for thousands of non-technical employees?
Non-technical employees need an approved path to useful work. Giving an employee a blank MCP configuration or an unapproved server registry leaves them to solve a technical governance problem. It makes access depend on choices that most employees are not equipped to evaluate.
The lesson for platform selection is to judge the ability to connect tools as part of the governed deployment path. The relevant criteria are enforced per-agent access, scoped user identity, centralized visibility, and human approval for irreversible actions. These criteria distinguish a platform that enables business users from one that merely gives them configuration options. The right way to use an agent should also be the easiest way.
Security and IT then become the teams that establish usable access rather than the teams asked to block unsafe access after deployment. Approval becomes part of the platform instead of a separate obstacle for every employee to navigate. KPMG found that a large majority of enterprise leaders cite security, compliance, and auditability as the most critical requirements for agent deployment. When the platform enforces them, non-technical employees can use approved capabilities without each becoming an expert in the risks underneath.


