Designing Approval Workflows for AI Agents: How to Structure Permissions and HITL

Introduction
Designing an approval workflow for AI agents means dividing tasks into three stages—automatic execution, proposal, and mandatory approval—and narrowing down the targets subject to HITL (Human-in-the-Loop) approval to only the high-risk processes involving money, personal information, external transmission, and the like. For DX promotion leads and system managers struggling with AI agent permission design, this article presents a four-step process for building an AI approval flow (permission level design, placement of HITL approval points, system implementation, and post-deployment review). By the end, you will be able to separate the scope of tasks that can be left to AI from the scope requiring human approval in your own operations, and concretely identify processes where approval is likely to be overlooked.
Before deciding on the details of an approval flow, you need to grasp the scope of tasks handled by the AI agent and where risk is concentrated within that scope. Skipping this step makes permission design and the placement of approval points ad hoc, forcing you to rebuild the whole structure later.
Identifying the Scope of Tasks to Delegate to AI Agents
To identify the scope of tasks, start by taking inventory of the processes executed by the AI agent, dividing them into "final decision-making" and "preparatory work for decisions." You need to record the distinction between decision support and automation, and define in advance which processes can be left to the AI all the way through final decision and approval. If this definition is postponed, the field staff will be unable to judge whether it's acceptable to execute the agent's output as-is, and operations will grind to a halt.
The inventory procedure is as follows:
- Break down the business flow into process units (e.g., inquiry classification, quote creation, invoice processing, contract review)
- List the types of data (personal information, monetary amounts, contract terms, etc.) that the AI agent touches at each step
- Separate whether the result of each step is sent externally or stays entirely in-house
- Check whether an error, if it occurs, is reversible or irreversible
This inventory process follows the approach of organizing in advance the use case, scope of impact, stakeholders, and data flow of an AI system. At the point where the scope has been identified, you should not yet assign permission levels; only after classifying the level of risk in the next step do you translate this into permission design.
When business staff and system staff break down processes separately, the granularity tends to diverge, and the precision of the inventory changes depending on whether, for example, "inquiry classification" is treated as a single process or split into three—intent determination, prioritization, and routing decision. The finer the granularity, the more precisely you can narrow down the targets for subsequent risk classification and permission design, so the process chart should be created while checking with business staff, and must not be finalized based solely on the judgment of the system staff.
Classifying High-Risk and Low-Risk Processes
Sort processes into high-risk and low-risk categories based on three criteria: impact on health, safety, and fundamental human rights; legal or equally significant impact; and involvement with critical infrastructure or safety-related matters. This classification is based on the concept of high-risk AI as presented in the EU AI Act; for details of the actual provisions (Article 6, Annex III, etc.), please consult the original text or a specialist. Any process that falls under one of these criteria becomes a target to be designed on the premise of human oversight.
In practice, it is effective to record, for each row of the process chart, a risk assessment obtained by multiplying "severity of harm" by "probability of occurrence." Severity is judged by whether the error is reversible, whether personal information or monetary amounts are involved, and whether external impact occurs; probability of occurrence is judged by how likely the agent is to make a judgment error or how likely input data errors are. This assessment is the organizing method presented in this article, and it needs to be conducted in a form that can be recorded rather than relying on subjective impressions.
Organizing the classification results as in the table below allows them to be carried over directly into the subsequent permission design.
| Example process | Severity | Probability | Risk classification |
|---|---|---|---|
| Automatic modification of contract terms | High (irreversible, legal impact) | Medium | High risk |
| Generating answers for internal FAQ | Low (reversible) | Low | Low risk |
| Automatic refund processing for customers | High (monetary, external impact) | Medium | High risk |
Processes classified as high-risk should be treated as targets for HITL placement, as explained in Step 2, while low-risk processes should have their scope of automatic execution examined in the Step 1 permission level design.
Step 1: How Should You Design Permission Levels?

Once tasks have been classified as high-risk or low-risk, decide how much execution authority to grant to each process. Authority should be tied not to job title or department but to the processing content itself, and should be designed according to the principle of least privilege, granting only the scope that is necessary.
Dividing Into Three Stages: Auto-Execution, Proposal, and Mandatory Approval
Simply classifying processes into low-risk and high-risk is not enough to translate authority levels into actual operations. For each process, you need to decide which of the three stages—"automatic execution," "suggestion," or "approval required"—it belongs to.
Automatic execution is assigned to processes that fall into the low-risk category and whose results are reversible. Processing such as generating answers to internal FAQs falls into this category, where even if an error occurs, the damage does not extend externally and is easy to correct.
Suggestion applies to processes where the risk is moderate and a human reviews the AI agent's output on the spot before moving to execution. This category covers processing such as drafting quotes or prioritizing inquiries, where the final execution decision is made by a human, but the creation of the material for that decision can be left to the agent.
Approval required is assigned to processes classified as high-risk in the risk classification—that is, processing involving irreversibility, legal impact, or monetary amounts and personal information. This is the category to which the HITL approval points covered in Step 2 are directly applied, such as automatic changes to contract terms or automatic refund processing.
These three stages are not fixed; they are designed on the premise that processes will be promoted or demoted based on operational results. When listing stage assignments, compiling process name, risk classification, current stage, and scheduled review date into a single table makes it easier to track the history of authority changes and simplifies later review work.
For processes where the assignment is unclear, it is safer to start operations with the lower stage as the default, and then decide whether to raise the stage based on how errors occur. Conversely, decisions to lower a process from "approval required" to "suggestion" or "automatic execution" should be based on logs accumulated over a certain period, rather than relying solely on the responsible person's intuition.
Restricting Access Permissions to Business Data and External APIs
Once the three stages—automatic execution, suggestion, and approval required—have been assigned to processes, the next step is to link each process to the specific data and APIs it can actually access. Deciding only the stage is not enough; if the authority granted to the agent extends across the entire business operation, a process designed with the "suggestion" stage in mind could end up executing high-risk data operations.
In practice, API keys and credentials called by agents should be stored in a dedicated secrets management mechanism, with read, write, and delete permissions granted individually for each process. For example, an agent responsible for creating quotes should be given only read access to the customer database, while write access used for billing confirmation or refund processing should be separated into different credentials.
# Pseudocode (please check API names and arguments in the official documentation)
policy = {
"quote_agent": {"customer_db": "read"},
"refund_agent": {"billing_api": "write", "scope": "approved_only"}
}All key issuance, use, and revocation should be recorded in audit logs, making it possible to later trace who used which key in which process. This separation corresponds to extending the principle of least privilege not just to the process level but down to the level of individual credentials, resulting in a design where a leak of authority in one process does not affect other processes.
For access to external APIs, handling must differ depending on whether the destination is outside the company. For processes that send data containing customer personal information or contract amounts to external services, a filter should be set up to narrow down the data fields before sending, designed so that unnecessary fields are not passed to the agent. Rather than deciding on the inventory of authority once and leaving it, incorporating a practice of reviewing it whenever processes are added or external API integration partners change helps prevent authority from continuing to loosen as business operations expand.
Step 2: Where Should You Place HITL (Human-in-the-Loop) Approval Points?

If every process assigned to "approval required" is subjected to the same weight of approval, frontline staff will be buried in confirmation work and approval will become a mere formality. Approval points should be narrowed down to processing that actually produces impacts on health, safety, or fundamental human rights, or significant financial impact, and should be placed only where the approver can effectively carry out monitoring, interpretation, and override of decisions.
Identifying Processes That Require Mandatory Approval, Such as Amounts, Personal Information, and External Transmissions
The three categories—monetary amount, personal information, and external transmission—function independently as criteria for determining whether approval is required. Even without individually conducting a severity assessment in the risk classification, falling under any of these three categories is sufficient grounds for judging that a process should be placed at the "approval required" stage.
Processing involving monetary amounts refers to processes such as refunds, billing confirmation, or placing orders, where funds move once the process is executed. Even if it appears reversible, once a refund or payment to the counterparty has occurred, it effectively becomes irreversible, so such processing should be included as subject to approval regardless of the amount involved. If small-amount payments are to be made subject to automatic execution, an upper limit should be clearly defined, with the design automatically switching to "approval required" once that limit is exceeded.
Processing involving personal information covers things like writing to the customer database or generating responses that include personal information contained in inquiries. Automated processing involving personal information may be subject to legal restrictions, and depending on the content of the processing, an obligation to explain to the data subject may arise. If operating a business in Thailand, it is necessary to confirm at the approval stage whether the processing complies with the consent of the data subject and the restrictions on purpose of use set forth by the PDPA (Thailand's Personal Data Protection Act).
External transmission refers to the entire category of processes that pass data to systems or services outside the company. Regardless of whether the destination is a business partner or an external API, any process that does not complete entirely within the company should be subject to approval, combining a filter that narrows data fields before transmission with an approval step. Even for processing that falls under none of these three categories, if the severity assessment classifies it as high-risk, "approval required" should be maintained, and it should not be demoted to automatic execution merely because it does not fall under these categories.
Clarifying Approvers and Escalation Paths
Even after approval points have been decided, if no one has been designated as the approver for each process, approval-required processing will simply remain stuck awaiting execution. Approvers should be designated by linking them to processes rather than job titles, and for each process, a list of multiple people holding approval authority should be maintained.
The designation procedure is as follows.
- Define one primary approver for each process
- Define one or more alternate approvers who can approve when the primary approver is unavailable
- Define conditions (such as elapsed time) for automatically escalating to a higher-level manager when the alternate approver is also unavailable
- Record the list of approvers, alternate approvers, and higher-level managers in the process table, and update it whenever there is a transfer or resignation
Setting up alternate approvers and automatic escalation in advance is positioned as an operational procedure for preventing situations where processing stalls because an approver is unavailable. The conditions for escalating to a higher-level manager should be defined numerically, such as "how many hours have elapsed since the request," so that the decision to escalate is not left to on-the-spot judgment.
The division of roles among approvers, alternate approvers, and higher-level managers, along with the escalation conditions, should be documented collaboratively by the risk management lead, legal, security, and business owners under the approval of executive management, avoiding a situation where this information is known only verbally by individual staff members.
To prepare for cases where, as with leave of absence or transfers, the alternate approver also becomes unavailable at the same time, defining the escalation path not as a single step but across multiple stages up to a higher-level manager makes it easier to avoid a situation where the entire approval route stops functioning simultaneously. It is also effective, rather than leaving the updating of the process table to the approver's own self-reporting, to designate someone responsible for conducting an inventory review at the time of personnel transfers, in order to prevent updates from being missed.
Step 3: How Should You Implement the Approval Flow in the System?

Defining approvers and escalation targets alone does not let the field know where a request is stuck. Approval events—requests, approvals, rejections, and revision requests—must all be recorded with timestamps and actor IDs, and stored in a tamper-detectable format. This is the foundation of operations.
Building a Mechanism to Visualize Pending Approval Tasks
Even with a mechanism in place to record approval events, if approvers themselves cannot see that requests are piling up, processing will remain stalled and neglected. Tasks awaiting approval must be listed as a queue, with the elapsed time since submission, priority, and current assignee always visible on a dashboard.
In designing this visibility, an approval deadline (SLA) should be set numerically for each step, and requests nearing the deadline should be displayed separately from those that have already exceeded it. For example, if the SLA for amount approval is defined as "within 4 hours of submission," a notification should be sent to the assignee once 3 hours have passed, and the process should automatically escalate to an alternate approver once 4 hours have passed.
The dashboard should display, at minimum, five items: step name, requester, summary of the request content, SLA time remaining, and current approver. When a mandatory approval step is distributed across multiple approvers, the number of unprocessed items per assignee should also be shown, so it can be confirmed whether requests are concentrated on a particular approver.
Displaying priority and assignee information and automating escalation notifications are positioned as operational best practices for preventing the buildup of pending approval tasks. Keeping approvers able to see their own workload at a glance is a prerequisite for mandatory approval steps to continue functioning effectively; without visibility into the queue, setting an SLA has no practical effect.
Logging Approvals, Rejections, and Revision Requests
A log of approval events is not useful for auditing if it only records the content of requests. For every action—approval, rejection, or revision request—the same four pieces of information must be recorded in the same format: who made the decision, when, what was decided, and on what grounds. For revision requests, it is also necessary to record which specific items were flagged for correction; otherwise, it becomes impossible to verify whether the same issue is being repeated after resubmission.
The recorded fields are fixed as follows:
# Pseudocode (verify API names and arguments in the official documentation)
log_entry = {
"timestamp": "2026-01-10T09:30:00+07:00",
"actor_id": "approver_042",
"action": "reject", # approve / reject / request_revision
"target_task_id": "quote_2026_0110_17",
"reason": "Amount exceeds approval limit",
"diff": {"amount_before": 500000, "amount_after": None}
}Relying solely on free-text entries for approval reasons makes it impossible to analyze trends later, so rejection reasons should be recorded using a combination of predefined options and free text. When defining rejection reason options, mapping them to the risk categories in the process chart—such as "amount exceeded," "insufficient authority," or "incomplete data"—makes it possible to later tabulate which risk category rejections are concentrated in.
Logs must be stored in a tamper-detectable format where entries, once written, cannot be altered afterward. This is based on the principle of treating log management itself as subject to audit; access permissions for logs should be designed with the same least-privilege approach as approver permissions, and who viewed or exported the logs should also be recorded separately. The history of revision requests should be stored as multiple linked entries tied to the same task ID, so that it can later be traced which resubmission ultimately led to approval.
Step 4: How Should You Review the Approval Flow After Operation?

The approval flow design should not be treated as fixed once built, but as something to be continuously adjusted based on operational data. Once SLA and rejection-reason logs have accumulated, the locations of bottlenecks and the appropriateness of authority levels should be reviewed quantitatively and updated to match actual conditions in the field.
Regularly Checking Approval Bottlenecks
Whether an approval flow is functioning well can be confirmed by periodically aggregating four metrics: number of SLA breaches, queue length, resubmission rate, and average processing time. These can be extracted directly from the logs recorded in the visualization of pending approval tasks, and can be compiled into weekly or monthly summary tables without adding any new measurement mechanism.
Example calculation: if a given step had 120 requests and 18 of them exceeded the SLA (4 hours), the breach rate is 18 ÷ 120 = 15%. If this breach rate has risen compared to the previous month, it suggests either that there are not enough approvers or that the scope of steps classified as mandatory approval is too broad.
A step with a high resubmission rate is often caused either by low quality in the submitted requests or by approval criteria not being well communicated to the field, and aggregating the rejection-reason logs can distinguish between the two. If rejection reasons are concentrated on "incomplete data," improving the request form should take priority; if concentrated on "amount exceeded," reviewing authority levels should take priority.
No single metric should be used in isolation for judgment—they must be viewed in combination. If queue length is increasing but average processing time remains stable, it can be concluded that the number of requests has simply increased, with no issue in the approvers' decision-making speed. Conversely, if queue length remains stable but average processing time is increasing, the cause lies either in the approvers' workload or in increasingly complex decision criteria.
By setting thresholds in advance and, once the breach rate or resubmission rate exceeds those thresholds, considering whether to add approvers, reconsider alternate approvers, or even downgrade the step itself to automatic execution or a suggestion-only stage, it becomes possible to make review decisions without relying on intuition alone.
Adjusting Permission Levels Based on Operational Performance
Once you have aggregated the approval bottleneck metrics, the next step is to reflect those results in a review of the authority levels themselves. The three stages—automatic execution, suggestion, and mandatory approval—are not fixed assignments; they should be treated as targets for promotion or demotion based on actual performance data for each process.
The basis for promotion decisions is the log of processes placed at the mandatory approval or suggestion stage. If, over a given period, the rejection rate is low and the content of correction requests is limited to minor issues, that process can be considered for demotion from mandatory approval to the suggestion stage, or promotion from the suggestion stage to automatic execution. This judgment should not be based on intuition; the aggregation period, rejection rate threshold, and number of target cases must be numerically defined in advance.
Demotion decisions need to be made more immediately. If a process at the automatic execution or suggestion stage exceeds the threshold for rejection rate or correction request rate, it should be reverted to mandatory approval without waiting for the cause to be investigated, and kept at that stage until the investigation is complete. If a process that has exceeded the threshold is left unaddressed until the next periodic review, the same error may recur from the same cause, so demotion should be operationalized to take effect the same day it is detected.
Leaving a record is a prerequisite for authority promotion. By recording who changed the stage based on which performance data, and maintaining the ability to revert to the previous stage at any time, you ensure that adjusting the authority level itself never becomes an unrecorded individual judgment. Appending the history of promotions and demotions to the process table, so that the reasons for past changes can be referenced at the next review, allows the same decision criteria to be carried forward even as personnel in charge change. It is advisable to operate staged authority promotion only after confirming both the performance data and the criteria defined in advance.
What Are Common Failures in Approval Workflow Design?

If the scope of approval targets and the placement of approvers are mishandled, the entire operation will grind to a halt even if the authority levels have been meticulously designed. Failures occur at both extremes—requiring too many approvals and lacking approvers—and both can be prevented by thoroughly applying risk-based design in advance.
Expanding the Approval Scope Too Much, Causing Operational Dysfunction
If even low-risk processes are included in the mandatory approval category, requests pile up with approvers, and high-risk requests that genuinely require attention get buried and overlooked. This tends to happen when risk classification is skipped in favor of a "just require approval for everything, to be safe" decision, which violates the design principle of limiting approval targets on a risk basis.
A typical sign of this is that the number of requests in the approval queue keeps increasing, while the time approvers spend reviewing each request keeps shrinking. When the operation devolves into simply pressing the approve button without scrutinizing the request content, approvals are recorded as having taken place, but they lose their effectiveness as oversight. This is known as a state of formalistic approval, and it tends to occur when approvers, lacking training or verification, grow accustomed to processing large volumes.
As a rule of thumb, if the average review time per request drops to half or less of what it was at the start of operations, you should suspect that the scope of targets is too broad. In addition to review time, if the same approver keeps approving for the same reason, this should also be treated as a sign that requests are being passed through without being read.
To prevent this, the mandatory approval targets need to be narrowed back down to processes that meet clear criteria—such as monetary amount, personal information, or external transmission—while demoting everything else to the suggestion stage. Building in a practice of checking whether queue length or resubmission rate spikes immediately after expanding the scope of approval targets, and revisiting the scope if they do, allows early detection of dysfunction in the field.
Processing Stalling Due to Absent Approvers
Even when approvers have been designated, requests can end up stranded during vacations, resignations, or transfers. In organizations that stop at merely listing alternate approvers and escalation conditions in the process table, it is common to find that the mechanism to handle cases where the alternate approver is also unavailable has been left out.
A typical cause is that the escalation condition is kept at an abstract description like "if the primary approver is unable to respond," without numerically defining who decides and when. If the primary approver does not reply for a day but the notification to the alternate approver does not fire automatically, the request will remain stuck pending approval. Even if resignation or transfer processing is completed on the HR system side, if the list of approvers on the approval flow side is not updated at the same time, requests can continue to be sent to an approver who no longer exists.
An effective countermeasure is to fix escalation conditions using numerical time-based criteria, such as "how many hours have passed since the request was submitted," and to separately designate someone responsible for auditing the approver list whenever a personnel transfer occurs. For processes with only a single approval route, the entire process halts the moment the alternate approver becomes unavailable, so multiple stages up to a higher-level responsible person should be designed in advance.
Anticipating cases where the alternate approver is also unavailable at the same time, such as during extended leave, assigning the final stage of the approval route to a role itself rather than a specific individual can shorten the period during which approval is left stranded, even when transfers and absences overlap. It is also essential to have a fixed person responsible for the audit and a system for periodically checking for gaps in updates to the approver list.
Frequently Asked Questions

When designing an approval workflow, there are points of uncertainty regarding the scope, speed, and decision-making authority of HITL. We organize these from three perspectives.
Q1: Is HITL Necessary for All Processes?
Not all processing requires HITL. The necessity scales with the magnitude of risk: high-risk processing involving impacts on health, safety, or fundamental human rights, or significant financial impact, requires effective approval that allows a human to monitor, interpret, and override decisions. On the other hand, for low-risk, high-frequency processing that is reversible and does not cause harm externally, switching to sampling-based post-hoc review or monitoring-based operation—rather than imposing approval on every instance—can still function as effective oversight. If approval requirements continue to be imposed on processes that do not fall into the high-risk category under the risk classification already established, the result will only be an increased burden on approvers, causing applications that genuinely warrant close attention to get buried.
Q2: Won't Introducing an Approval Flow Slow Down Operations?
Immediately after introduction, the additional work of reviewing applications will temporarily slow down processing speed. However, by defining SLAs numerically, expanding the scope of automatic execution and proposal-stage processing to narrow down what requires mandatory approval, and combining this with automatic escalation to alternate approvers, it becomes possible to transition to an operation that achieves both speed and safety. Whether things slow down depends on the precision with which approval targets are narrowed down.
Q3: Who Should Decide the Permission Design?
Authority design is premised on a structure in which the decision is not left to the judgment of a single business owner alone, but is instead made collaboratively by the risk management lead, legal, security, and the business owner, approved by executive leadership, and documented. A rough guide to the division of roles is as follows:
- Business Owner: Creates the inventory of processes and the draft risk classification
- Legal: Determines the legal implications for processing involving personal information, contracts, or monetary amounts
- Security Lead: Designs the approach to access permissions and secrets management
- Risk Management Lead: Oversees classification criteria and the review cycle, and manages records
Once this structure is established, it should not be fixed in place; it should be reviewed by the same members whenever processes are added or external API integration partners change. A situation where only a specific person understands the authority policy should be avoided, because once that person is transferred or leaves, the criteria for judgment will fail to be passed on.
Summary

The sequence of design work—inventorying operations, classifying risk, establishing the three authority levels, placing HITL approval points, implementing them in the system, and reviewing after operation begins—does not function if only one stage is put in order. The foundation is a flow in which processes are identified by separating them into decisions and preparatory work, risk is classified by the magnitude of impact on health, safety, fundamental human rights, or legal implications, and the results are then assigned to one of three stages: automatic execution, proposal, or mandatory approval. For processes requiring mandatory approval, approvers and escalation destinations are designated based on criteria such as monetary amount, personal information, and external transmission, and everything from application to approval, rejection, or request for revision is recorded with timestamps and actor IDs, making the status of pending approvals visible. Once operation begins, bottlenecks should be regularly checked using metrics such as the number of SLA breaches and resubmission rates, continuing the adjustment of raising the stage for processes with low rejection rates and few revision requests, while immediately reverting processes with notable breaches back to mandatory approval. Not over-expanding the scope of approval targets, and having alternate routes in place so that things do not grind to a halt when an approver is unavailable, are also prerequisites for sustaining this cycle. Start by taking inventory of your own company's operations on a process-by-process basis, and begin with the work of listing where risk is concentrated.
References
- eur-lex.europa.eu/legal-content/EN/TXT/PDFAccessed November 16, 2026
- digital-strategy.ec.europa.eu/en/faqs/navigating-ai-actAccessed November 16, 2026
- nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-57pt1r5.pdfAccessed November 16, 2026
- docs.oracle.com/ja-jp/iaas/Content/access-governance/create-approval-workflow.h…Accessed November 16, 2026
- laws.e-gov.go.jp/law/359AC0000000086Accessed November 16, 2026
- www.ratchakitcha.soc.go.th/DATA/PDF/2562/A/069/T_0052.PDFAccessed November 16, 2026
Author & Supervisor
Yusuke Ishihara
Started programming at age 13 with MSX. After graduating from Musashi University, worked on large-scale system development including airline core systems and Japan's first Windows server hosting/VPS infrastructure. Co-founded Site Engine Inc. in 2008. Founded Unimon Inc. in 2010 and Enison Inc. in 2025, leading development of business systems, NLP, and platform solutions. Currently focuses on product development and AI/DX initiatives leveraging generative AI and large language models (LLMs).


