# How Should Enterprises Control AI Agents in 2026?

infonesia.fyi · September 26, 2026

> The Direct Answer Enterprise AI agent controls should operate as a permissioned control plane that connects each non-human identity to a human owner, a...

## The Direct Answer

Enterprise AI agent controls should operate as a permissioned control plane that connects each non-human identity to a human owner, a limited scope, approved tools, observable actions, and an emergency stop. The unit of control is not merely the model or chatbot; it is the complete path through which an agent receives context, selects a tool, changes data, sends communications, or spends money. In 2026, that matters because agents can pursue goals and act across software with some degree of autonomy, turning a bad instruction or excessive permission into an action at machine speed. The practical minimum is an identity for every agent, short-lived credentials, allowlisted destinations, action-specific authorization, tamper-resistant logs, and a tested mechanism to suspend the agent. Controls should begin with read-only tasks and low-risk internal actions, then expand only when evidence shows that the agent behaves as intended. This approach does not require a large platform purchase: a small team can start with an inventory, an owner, about 20 governed use cases, and weekly exception reviews.

**Also worth reading:** [How Can Indonesian Enterprises Reduce AI Costs Without Losing Control in 2026?](https://infonesia.fyi/knowledge/how_can_indonesian_enterprises_reduce_ai_costs_without_losing_control_in_2026.php) · [How Secure Are Indonesian AI Vendors, and What Should Enterprises Check Before Buying?](https://infonesia.fyi/knowledge/how_secure_are_indonesian_ai_vendors_and_what_should_enterprises_check_before_buying.php) · [Indonesia AI Market Data in 2026: What Should Enterprises and Investors Track?](https://infonesia.fyi/knowledge/indonesia_ai_market_data_in_2026_what_should_enterprises_and_investors_track.php)

## What Enterprise AI Agent Controls Actually Govern

An AI agent is a program that can interpret objectives, use tools, and take actions with some level of autonomy. That makes its risk profile different from a conventional SaaS login because a single compromised prompt can influence several downstream operations, including database queries, code changes, customer messages, or cloud configuration. Controls must therefore cover identity, context, tools, actions, data, and monitoring rather than just endpoint security. Examples include ContextFort, which focuses on visibility and controls for browser agents; Recursant, positioned as a mesh-based control plane; AGent Based Access Control, focused on agent identity and access; and ClawForge, which describes itself as mobile-device management for AI assistants. The categories are not identical, but together they show a market moving toward managed non-human identities and governed agent behavior.

The control boundary should distinguish at least four layers: the model, the agent’s identity, the tools it can call, and the target systems. A model may be hosted by a third party, while the agent and its credentials belong to the enterprise; the downstream database then applies its own authorization rules. This separation prevents the model vendor from becoming the effective policy authority. It also allows a team to replace one model or agent framework without rewriting every control. A useful policy states which data may enter the context, which actions the agent may attempt, which environments it may affect, and who can approve exceptions. If those conditions cannot be expressed in machine-enforceable rules, the use case is not ready for production autonomy.

## Why Traditional Identity and Security Controls Are Not Enough

Existing identity governance remains necessary, but granting a human or service account a broad role can become disproportionately powerful when an agent can execute that role continuously. Aembit’s support for Okta Cross App Access illustrates the direction toward extending enterprise identity controls to AI agents rather than treating them as ordinary users. Aembit’s approach reflects the same “least privilege” requirement found in broader access-management systems, with authorization tied to workload identity and contextual attributes. However, traditional IAM was generally designed around known users, applications, and static roles; agents add variable objectives, generated tool calls, prompt dependencies, and chains of delegated actions. A permission that is safe for a person performing one task at a time may be unsafe for an automated process repeating it hundreds of times.

A second limitation is that identity verification answers who is acting, but not whether a particular action makes sense. An authenticated agent may still receive poisoned instructions from a webpage, confused two customer records, or use a search tool to retrieve malicious content. Conversely, a legitimate agent may be blocked because a coarse access policy cannot distinguish a harmless draft from a published message. Effective controls combine conventional authentication with contextual authorization based on task, environment, data classification, destination, and transaction size. IBM’s guidance on governing third-party AI agents and Oracle’s work on database security controls beneath AI agents point toward this layered model: controls must sit beneath the agent, where the platform can independently enforce policy even if application logic is imperfect.

## A Practical Implementation Sequence in Seven Stages

First, create an inventory of every agent, including assistants embedded in productivity tools, coding systems, customer-service platforms, and internal automations. Assign a named human owner, business purpose, model provider, data sources, tool list, environments, and risk tier. A sensible starting portfolio is 10 to 20 agents or workflows, because that is large enough to expose repeated policy needs but small enough for accountable review. Agents that can write code, execute commands, move money, alter permissions, or contact customers should initially be treated as high risk. The inventory should also record dormant and forgotten agents, which are often overlooked because they no longer appear in active projects.

Second, issue a separate identity for each agent or bounded role rather than sharing a general employee account. Use short-lived credentials, rotate secrets automatically, and prohibit embedded passwords in prompts or source code. Third, replace broad “admin” access with tool-level permissions: a market-intelligence agent might search approved sources and write a draft, while publication requires human approval or a narrower publishing identity. Fourth, place outbound controls around browser navigation, API calls, email recipients, file downloads, and code execution. Every exception should have an owner, reason, expiration date, and review date. Fifth, log prompts where policy permits, retrieved context, tool calls, authorization decisions, outputs, and final actions, while applying retention and privacy limits. Sixth, test normal, ambiguous, adversarial, and failure scenarios before deployment. Seventh, define kill thresholds and rehearse suspension at least twice per year, including revocation of credentials and interruption of downstream jobs.

A useful go-live threshold requires 100% of production agents to have an owner and inventory record, 100% of credentials to be non-shared, and at least 95% of tool calls to be covered by explicit allow or deny rules. Teams should set stricter thresholds for destructive actions: zero unapproved production writes and zero standing production-admin credentials are reasonable targets. Human approval may be required for external publication, financial transactions above a fixed amount, privilege changes, security-control changes, and access to regulated data. These are operating thresholds, not universal legal standards, so organizations should adjust them to their control environment and applicable Indonesian, Singaporean, or other SEA obligations.

## Control Options and Platform Comparisons

There is no single product category called an enterprise agent control platform. Buyers may assemble controls from IAM, browser security, API gateways, data governance, model gateways, and agent observability tools, or adopt a platform that manages part of that stack. The comparison below describes architectural options rather than endorsing particular vendors or claiming equivalent feature coverage.

| Feature | Central agent control plane | Existing IAM, DLP, and API controls | Manual operating model |
| --- | --- | --- | --- |
| Agent identity | Purpose-built workload or agent identities are common | Often requires custom extensions and role design | Shared or personal accounts are common |
| Tool and action policy | Can connect task, agent, tool, target, and approval | Strong for data, users, APIs, and endpoints | Depends on spreadsheets and approvals |
| Browser-agent visibility | Purpose-built browser sessions and tool-call records | May require proxies, extensions, and log stitching | Usually limited to screenshots or chat transcripts |
| Deployment effort | Highest initial integration effort | Moderate because many controls already exist | Lowest setup cost, highest operational burden |
| Suitability | Regulated or scaled agent fleets | Smaller deployments with known systems | Pilots, low-risk internal experiments |
| Principal weakness | Can duplicate IAM and data-security functions | May not understand goals, context, or multi-step actions | Slow, inconsistent, and difficult to audit |

Central control planes offer the most coherent policy experience, but they can create another vendor dependency or reproduce capabilities already available in the enterprise stack. Existing IAM, DLP, API management, and database controls are usually easier to procure incrementally, yet they may struggle to connect a prompt to a later action across several systems. A manual model is acceptable for a tightly bounded pilot with fewer than approximately five workflows, but it does not scale well once agents create hundreds of daily tool calls. Mature organizations often need a hybrid model: retain IAM as the credential authority, use data platforms to enforce classification and transaction policy, and add agent-specific visibility for context, intent, and behavior. Databricks and Snowflake discussions about secure AI workflows and an agentic control plane reflect this direction, although their exact products and deployment models should be evaluated independently.

## Design Principles That Reduce Real Risk

The most important principle is to constrain capability rather than attempt to infer safe intent from every prompt. An agent should be unable to access an entire database when its task only requires a filtered view. It should not be able to publish directly when a draft-and-approve step is sufficient. Transaction and action limits should be enforced outside the model, because the model can suggest a policy but cannot reliably police itself. Data tools should return only fields required for the task, and sensitive values should be masked or tokenized when possible. This approach reduces consequences even when the underlying model is wrong, manipulated, or unexpectedly capable.

Context deserves equal treatment. Browser agents can encounter untrusted instructions, prompt injection, manipulated search results, and malicious files, so retrieved content must be treated as data rather than as trusted policy. Retrieval systems should document source permissions, refresh intervals, and citation provenance, while agents should be prevented from converting external instructions into new permissions. A human owner should approve templates that materially affect external users, not approve every harmless token operation. Logging should capture enough evidence to reconstruct an event, but organizations should avoid retaining every prompt indefinitely, particularly when prompts contain personal or commercially sensitive information. A balanced retention period should follow purpose and law rather than a universal number.

The control design should also account for non-determinism. Testing should measure failure rates over repeated runs rather than relying on a single successful demonstration. A reasonable pre-production test set is 100 representative tasks plus adversarial variants, with a 95% or higher task-completion target for low-risk workflows and a stricter ceiling on unauthorized side effects. High-impact actions should have a target of zero material policy violations during testing. Teams should evaluate false approvals and false blocks separately, because a system that blocks safe work may appear secure while forcing users to bypass it. Controls should therefore preserve enough usability to prevent teams from creating shadow agents outside the governed environment.

## Common Mistakes and Cost Trade-Offs

A common mistake is confusing an LLM gateway with a complete agent-control solution. A gateway can mediate model requests, apply content rules, and track token use, but it may not govern credentials, browser sessions, retrieved context, tool arguments, or downstream writes. Another mistake is buying a platform before defining ownership and policy. “IAM for agents” may provide non-human identities and conditional access, while browser visibility, data controls, and action approval still need to be integrated. Organizations also make the mistake of treating model providers as the sole control layer, even though responsibilities are often distributed among the model vendor, agent builder, enterprise integrator, and destination system.

Pricing varies too much for a defensible generic claim. Agent-control platforms frequently use custom enterprise pricing based on agents, identities, tool calls, sessions, protected users, data volume, deployment type, and support requirements; public list prices are therefore uncommon. Existing IAM and API products may already be covered by enterprise subscriptions, while a pilot can sometimes be assembled with existing seats and a modest cloud or observability budget. The relevant comparison is total operating cost, including integration, policy administration, security engineering, model consumption, audit storage, and the labor required to review exceptions. A nominal low license price can become expensive if every agent requires custom connectors or if incident investigation consumes scarce staff time.

Startups such as Recursant, ContextFort, AGent Based Access Control, and ClawForge indicate active product experimentation, but maturity, regional support, and production references should be checked directly. The 27 September 2026 date does not justify assuming that every advertised capability has been independently validated at scale. Ask vendors for architecture diagrams, identity ownership, audit-log portability, data residency, failure behavior, breach history, support response times, and a test exit plan. A credible control should let a customer disable the vendor without losing all evidence or leaving credentials behind. Contract terms should also state whether logs and prompts remain available after termination and how model or data usage is governed.

## When to Act, Pilot, or Pause Deployment

Enterprises should act now if they already have multiple production agents, especially when those agents can access customer data, production code, financial systems, or external communications. Regulatory and customer expectations are increasing, while the supplied research context notes that agents have been expanding faster than confidence and control. Waiting for a single universal standard can create unmanaged growth, but buying before a use case is defined can produce unused governance. A sound trigger is repeated use by more than one team, deployment beyond a sandbox, or any request for standing write access. At that point, inventory and least-privilege controls should precede broader rollout.

A pilot is appropriate for internal research, drafting, summarization, and code suggestions when actions are reversible. The pilot should have a 30- to 90-day period, named owner, approved tools, clean or synthetic data where possible, and explicit success criteria. Production deployment is reasonable when the workflow has stable inputs, measurable business value, tested authorization rules, rollback capability, and accountable monitoring. Pause or restrict a workflow when audit coverage is incomplete, an agent can take irreversible actions without approval, source permissions are unclear, or control performance cannot be measured. Temporary shutdown should be routine rather than exceptional: teams should practice it before an incident to verify that credentials, scheduled jobs, browser sessions, and downstream approvals are actually stopped.

For Indonesian and SEA teams, the operating model should account for data residency, cross-border processing, local sector rules, cloud regions, and customer contracts. A global platform may be technically sound but commercially or legally awkward if local support, Bahasa Indonesia capabilities, or data-location commitments are absent. A regional control layer should not weaken centralized policy, however; it can standardize evidence while allowing country-specific retention, approval, and residency configurations. The best near-term objective is not “zero risk,” which is unrealistic for probabilistic software, but bounded impact, rapid detection, and reliable recovery. Progress should be reviewed with a small set of numbers: percentage of agents inventoried, percentage using non-shared identities, unauthorized-action rate, mean time to revoke, percentage of high-risk actions requiring approval, and incident recurrence.

## The 2026 Control Baseline

By 27 September 2026, an enterprise should be able to answer four questions for every production agent: who owns it, what can it access, what actions can it take, and how would it be stopped? Those answers should be available in an inventory and enforced by systems, not only in a policy document. The baseline includes dedicated identities, short-lived credentials, least-privilege roles, allowlisted tools, contextual authorization, data classification, prompt and action logging, exception expiry, human approval for high-impact operations, and tested revocation. It also includes a clear division of responsibility between model, agent platform, identity provider, data system, and business owner. This division is especially important because no individual product or model can guarantee safe autonomous behavior on its own.

The next step is not necessarily a large procurement. Inventory existing agents, select approximately 10 to 20 representative workflows, classify their risks, and pilot the control model in one business unit. Measure unauthorized actions, approval delays, false blocks, incident investigation time, and time to revoke. Expand only when controls improve without materially suppressing legitimate work. Agent adoption will continue, but enterprises do not need to choose between unrestricted autonomy and a total ban; they can use graduated autonomy based on evidence, reversibility, and business value. That is the more defensible basis for enterprise AI agent controls in 2026.

## Quick answers

### What is the difference between AI agent governance and an LLM gateway?

An LLM gateway usually mediates model requests, model choice, content policy, token usage, and sometimes logging. Agent governance extends to identities, retrieved context, tools, browser sessions, downstream writes, approvals, and revocation. A gateway can be one component of a control plane, but it does not automatically govern actions taken after the model responds.

### How many AI agents should an enterprise inventory first?

A practical initial target is 10 to 20 agents or bounded workflows, including forgotten and low-use deployments. The exact number matters less than achieving complete ownership, identity separation, and risk classification before expanding. Larger organizations should still begin with a representative slice and scale the process through a maintained inventory.

### What actions should require human approval?

Human approval is usually appropriate for external publication, financial transactions, production writes, privilege changes, security-control changes, and access to highly sensitive data. Thresholds should reflect the business value and reversibility of the action. Approval should be attached to the actual action and target rather than to an entire agent session.

### Do AI agents need separate identities from employees?

Separate, non-shared identities are generally safer because agents operate continuously, use tools, and may need different permissions from the people who supervise them. Short-lived workload credentials and agent-specific roles reduce the effect of leaked secrets. Some low-risk pilots can use controlled sandboxes, but production agents should not rely on shared personal accounts.

### How much does enterprise agent governance usually cost?

There is no reliable universal price because many products use custom enterprise pricing based on agents, identities, sessions, tool calls, data volume, deployment, and support. Existing IAM, DLP, and API platforms may reduce integration costs, while a focused pilot can use current enterprise capabilities. Buyers should compare total cost, including policy administration, monitoring, storage, integrations, and incident response.

Canonical: https://infonesia.fyi/knowledge/how_should_enterprises_control_ai_agents_in_2026.php
Markdown: https://infonesia.fyi/knowledge/how_should_enterprises_control_ai_agents_in_2026.php/index.md