How Can You Trust an AI-Driven Review? Six Months of ARGOS Omni ①
How Can You Trust an AI-Driven Review? Six Months of ARGOS Omni ①
Previously, we shared the latest updates to ARGOS ID check, our eKYC solution, and how the service has evolved over time.
This time, we're turning our attention to ARGOS Omni, our AI Agent for document review and KYB operations.
Since its alpha launch in April, Omni has continued to evolve through real-world use. It has expanded beyond analyzing documents and delivering review results to help teams understand why a decision was made, identify where issues occurred, and manage AI-driven reviews more effectively in day-to-day operations.
So, what has changed in Omni over the past six months?
In this article, we'll walk through the key updates that have helped Omni evolve into a more transparent and reliable AI Agent for real-world KYB operations.
What Changes When AI Can Review Documents and Deliver KYB Results?
ARGOS Omni is an AI Agent designed to automate repetitive review tasks traditionally handled by human operators, using documents and information submitted by businesses.
Over the six months since its alpha launch in April, one of the most important questions we encountered wasn't simply:
“Can AI perform the review?”
There was a more important question:
“Can reviewers trust the result enough to use it in real operations?”
For critical processes such as KYB, simply providing a result isn't enough. Reviewers need to understand what information led to a decision and be able to trace what happened when the outcome differs from what they expected.
Otherwise, even if AI performs the initial review, a human reviewer still has to go through everything again from the beginning. At that point, much of the value of automation disappears.
That's why, over the past six months, ARGOS Omni has focused not only on automating more tasks, but also on creating an environment where people can understand, verify, and validate the decisions made by AI Agents.
In this first part of our Omni update, we'll look at the features designed to record analysis results, explain why those results were produced, and help teams identify which analyses require additional human review.
Manage Review Results as PDF Reports
Once an analysis is complete, users can select Download PDF from the results page to export the analysis as a report.
The report includes the analysis ID, completion time, profile ID, Analysis Score, summary, and key decision variables. Reports are available in both Korean and English.
From Viewing Results to Keeping a Reliable Record
If review results are only available on screen, that may not seem like a major limitation during everyday operations.
But the situation changes when teams need to retrieve and submit historical decision records for internal audits or regulatory reviews.
For example, an internal audit team may request the decision rationale for a sample of rejected cases from the previous quarter. A reviewer may also need to retrieve the complete review history for a specific business.
If the only way to respond is to capture individual screens and manually compile them into a separate document, another layer of manual work is created on top of an otherwise automated review process.
Omni's PDF reports are designed to preserve what information was reviewed and what result was produced at the time of the analysis in a single record.
Because the analysis ID and completion time are included, teams can identify exactly when each analysis was performed. The report also preserves the original result even if the product interface changes later.
Reports can also be retrieved via API, making it possible to automatically download and store them in an organization's internal archive as soon as an analysis is completed.
Instead of manually assembling documentation whenever it is requested, teams can build an operational process where auditability and recordkeeping are incorporated into the review process from the beginning.
Report generation has also been separated onto a dedicated server. An analysis is marked as complete once its report is ready, reducing the need for users to wait again after opening a completed analysis.
See Not Only the Result, but Why It Happened
In automated reviews, the result itself is only part of what matters.
Teams also need to know:
“Why did this result occur?”
In Omni's Report tab, users can review the rationale associated with individual fields alongside the analysis results.
When a single step extracts or verifies multiple pieces of information, Omni doesn't apply one explanation to the entire step. Instead, each field can display its own reason or status.
If a value isn't successfully populated, Omni doesn't simply leave the field blank. It indicates why the value is missing.
Reason | What It Means |
|---|---|
Parent item unavailable | The parent object was not generated, so the child field could not be generated |
Value is empty | The field exists, but no value is present |
Not included in final output | The value was extracted but was not included in the final output |
Engine call failed | The external engine assigned to the step failed to execute |
The Same Empty Field Can Require a Very Different Response
Suppose the representative's name is missing from an analysis result.
Previously, a reviewer might have had to investigate several possibilities one by one: Was there a problem with the submitted document? Did extraction fail? Was there an issue with the policy configuration?
In some cases, the reviewer might even ask the customer to resubmit the document, only to discover that the actual issue was related to the system or output configuration.
When the reason is visible immediately, the next action becomes much clearer.
If the status is Engine call failed, there's no reason to ask the customer to submit the document again. If it says Not included in final output, the reviewer can inspect the output structure rather than starting with the extraction process.
In other words, the workflow shifts from guessing “Why didn't this work?” to identifying “Where should I look?”
This not only reduces the time reviewers spend rechecking results from the beginning, but can also help prevent unnecessary document requests and repeated troubleshooting.
For analyses completed before this feature was introduced, decision reasons will remain empty. Reasons are not generated retroactively for historical analyses.
Use Analysis Score to Prioritize Reviews
If every analysis still requires the same level of manual review, the benefits of automation remain limited.
This becomes particularly important for organizations processing hundreds of reviews each day. Teams need a way to determine which analyses actually require closer human attention.
Since July, Omni has provided an Analysis Score at the top of the analysis results page.
Analysis Score is a 0–100 score that indicates how completely an analysis was executed according to the configured workflow.
The score is calculated using three main components:
Component | Weight | What It Measures |
|---|---|---|
Step Execution | 40% | Percentage of defined steps that were actually executed |
Step Quality | 20% | Status of executed steps and whether engine calls were completed |
Output Completeness | 40% | Percentage of final output fields that contain values |
Results are categorized into three levels: HIGH (80–100), MEDIUM (65–79), and LOW (0–64).
For example, analyses with a HIGH score can be reviewed quickly, while those categorized as LOW can be prioritized for closer inspection.
Rather than reviewing every case with the same level of attention, teams can use the score as a reference to identify where human review is most valuable and focus their resources accordingly.
However, Analysis Score should not be interpreted as an approval or rejection score.
An analysis may pass all verification steps but still receive a lower score if the final output schema isn't sufficiently populated. Approval decisions should therefore be made by considering the individual verification results alongside the Analysis Score.
Making AI Decisions Verifiable by People
One of Omni's key areas of focus over the past six months has been ensuring that AI-driven decisions don't remain a black box.
Analysis results can now be preserved as PDF reports, the reasoning behind individual fields can be reviewed, and Analysis Score can help teams identify analyses that may require additional attention.
Ultimately, effective KYB automation isn't only about how many results an AI Agent can produce.
For AI-driven reviews to become part of real operations, people need to understand the results, verify the reasoning when necessary, and have enough context to trust the process.
So what comes next when teams want to go beyond reviewing AI-generated results and start adjusting and managing how the AI Agent performs the work itself?
In Part 2, we'll cover the next set of Omni updates, including step-level workflow editing, new verification engines for Korean businesses, credit usage tracking, and expanded audit records all designed to give teams greater control over how Omni operates in real-world environments.
