Make the existing guide easier to find
Useful when a merchant does not know what a valid photo should look like. The guide still cannot tell whether the uploaded image matches the business being registered.
Lower effort, but limited coverage.
Farhan Fauzan JamaludinI designed an AI-assisted pre-check in AstraPay’s merchant KYB web dashboard, translating model findings into actionable, non-blocking warnings. After stable rollout in September 2026, targeted mismatch submissions declined from 54.6% to 35.0% compared with May–June, a 35.9% relative reduction. Among decided AI-checked submissions, clean cases were approved 75.9% of the time versus 39.4% for flagged cases. The findings are observational and human verification remains the final decision-maker.
I designed how Llama findings become actionable warnings inside AstraPay’s existing KYB web dashboard, without turning an uncertain model output into an approval decision.
Four mapped caution groups appear together, with clear choices to fix the submission or continue.
Field-level messages and replace-file actions make corrections visible where the merchant can act.
All displayed dashboard screens are the original designs. Select a stage or follow its controls to inspect how the warning and correction work.

The grouping artifact documents how detailed finding conditions become stable user-facing copy: business-place photo, product photo, business-information mismatch, and non-direct or manipulated photos.
Inspect the caution-mapping artifact ↗
Possible mismatch, not an automatic rejection
Fix the flagged data or continue
Normal human-led KYB verification
From May–June to the stable September rollout, targeted KYB mismatches fell from 54.6% to 35.0% of submissions, a 35.9% relative reduction. Approval among decided cases rose from 35.8% to 56.6%. These observed changes are not an isolated causal estimate of the AI pre-check.
Original AstraPay KYB dashboard screenshots and original caution-group mapping supplied by the designer. UI copy remains in its original Indonesian.
KYB approval was falling while submission volume was also weakening. I did not want to improve approval by adding another mandatory step that made the flow harder to finish.
I started with rejection data and quick interviews with several merchants who had been rejected. The three most common problems were mismatches between business information and the storefront photo, product photo, or business name.
Some cases came from guidance that was missed or misunderstood. Others genuinely needed checking because the uploaded image or business name did not clearly represent the business being registered.
The goal was to catch likely mismatches earlier without letting AI decide whether a merchant should be approved or rejected.
Original research and decision-mapping artifacts: recurring rejection reasons, approval trend, and the mapping from model findings to product decisions. None of these images is a simulated mobile interface.
The decline mattered more because submission volume was weakening during the same period.

The interviews showed that guidance needed work. The rejection data also showed cases where the actual submission needed to be checked before it was sent.
Useful when a merchant does not know what a valid photo should look like. The guide still cannot tell whether the uploaded image matches the business being registered.
Lower effort, but limited coverage.Use the business information and photos already available in Step 1 to show a warning before the application reaches verification.
AstraPay already had Llama available through Vertex AI, which made this a more practical option to ship quickly.Try the pre-check flowThis flow separates the responsibilities of the merchant, product layer, Vertex AI, and verification.
Business name, category, storefront photo, and product photo already exist in the KYB flow.
The product layer sends the relevant inputs before the merchant submits.
The model looks for likely inconsistencies across the business information and photos.
Technical output becomes stable conditions and merchant-facing feedback.
A warning only appears when a defined condition needs the merchant's attention.
The merchant can correct the flagged input or continue when they believe the submission is already right.
The application enters the normal verification process.
The AI pre-check does not replace verification.
Flow connections
AI only provides a finding before submission. Verification still decides whether the merchant is approved or rejected.
I worked with the developer on the prompt, condition definitions, and output structure. The response needed to be predictable enough for the product layer to map it to the right message and next action.
Prompt detail only helped when it improved a decision the product could use. Output that was too open-ended made the mapping harder to maintain and made incorrect findings harder to evaluate.
The model produces a finding. The product still decides how that finding changes the merchant experience.
The combined rate for the three rejection reasons this pre-check was designed to catch declined in the clean before-and-after comparison.
Measured on all submitted KYBs, not only cases checked by AI.
The approval rate also rose from 35.8% to 56.6% among decided applications. This is a directional observation, not a measured causal effect of the AI feature.
129 of 163 September submissions had an AI pre-check (79.1% coverage). July and August were a pilot with uneven usage. October 1–21 was a partial follow-up and is excluded here.
Among 125 AI-checked September submissions with a final decision, cases with no stored score below 70 were much more likely to be approved. Cases with at least one score below 70 were more likely to receive a targeted mismatch remark.
of AI-flagged cases were still approved by human reviewers (28 of 71).
Within the 71 AI-flagged decided cases, merchants who ran validation again had directionally better approval and targeted-mismatch results than those who checked only once.
A score below 70 was a caution signal, not an automatic rejection. The merchant could fix the submission or continue, and verification retained final authority. Human approval is not proof the AI made a false positive.
A recheck means more than one AI validation before submission. It does not confirm that the merchant edited their data. The groups are small, self-selected, and these differences are not conclusive at conventional statistical thresholds.
Source: AstraPay KYB operational tables and reviewer remark analysis provided by the designer. PRE covers May–June 2026. July–August was an uneven pilot. POST is September 2026. Targeted remarks refer only to the three mapped BUSINESS_DATA rejection reasons. Rates are descriptive and do not establish causation.
After production usage became consistent in September 2026, the combined rate of the three targeted rejection reasons was 35.0% of submitted cases, compared with 54.6% in May–June. That is a 35.9% relative reduction in the issues the pre-check was specifically designed to surface.
Approval among cases with a final decision was 35.8% PRE and 56.6% POST. This is useful downstream context, but the groups differed across time and not all September submissions were AI-checked. Other changes in merchant mix and verification can contribute to the movement.
Within September’s 125 decided, AI-checked submissions, clean cases were approved 75.9% of the time versus 39.4% for cases with a stored score below 70. Among those flagged, cases with another pre-check showed directionally better outcomes. Running another check does not establish that an edit happened or that the check caused a better decision.
The strongest evidence is alignment between the pre-check signal and human review. AI still remained advisory because 39.4% of flagged cases were approved by human reviewers.
I would connect each warning shown, action selected, actual input change, and subsequent AI result to the final reviewer remark. The current attempt count only proves another validation, not that the merchant corrected the flagged evidence.
A carefully labeled reviewer evaluation set would help measure false flags and missed mismatches by document type. Human approval remains a useful outcome, but it is not a perfect ground-truth label for every AI finding.