When GPT-4 arrived in March 2023, it became evident to those paying attention that large language models could touch almost every workflow. At Envisso, we quickly realised that LLMs could and should change how merchant risk is discovered, measured, and acted on.
Our first realisation: the web already knows
Before LLMs, risk estimation leaned on two levers: financial patterning and manual reviews by analysts. These remain useful today, but they are neither exhaustive nor scalable.
Our first breakthrough was simple: the richest source of risk signal is the open web. Rather than focus on news about listed companies, we turned to where even the smallest merchants live, Google Maps, Trustpilot, and other review platforms. LLMs let us use this crowdsourced knowledge at scale and with nuance.
A star rating alone is blunt. What matters is why customers complain. With LLMs, we surface patterns like “no returns honoured,” “non-delivery after payment,” and “item not as described.” Reviews become labelled risk factors rather than sentiment scores. This was our first proof that LLMs can turn diffuse text into decision-grade structure.
Our second realisation: websites are dossiers
Next we asked: can we flag risky merchants by studying their own websites? Experienced analysts already do this. Scaling that judgement was a matter of formulating the task.
We built agents that read a site like an investigator, covering policies, product pages, merchant identity, and more. LLMs extract and cross-check:
- Does the merchant have policies that are genuine, well structured, and appropriate for the jurisdiction?
- Do site details corroborate the information provided to the PSP?
- Does the site use basic security measures?
- Are product categories consistent with the declared MCC, or is there mismatch risk?
The output is structured risk evidence, not a vague “this looks bad.”
Our third realisation: fewer charts, sharper decisions
As our Risk Console filled with high-signal data, we hit a human limit. People do not want more widgets; they want the one thing to do next. So we pushed LLMs from extraction into synthesis.
We now generate brief summaries for each category and an overall view. Our agents read and rank. Your analysts review and decide.
What changed for our partners
- Coverage at depth: We read the internet trail of each merchant.
- Speed with evidence: Minutes to a defensible view that cites sources and policies.
- Specificity: Not just “bad reviews,” but the type and recency of complaints mapped to risk categories.
- Traceability: Every finding links back to the exact text and page location, enabling audit and appeal.
This is flipping risk into a growth strategy: recognise good merchants faster, focus checks where they matter, act before disputes pile up.
The impact
Our partners are onboarding legitimate merchants in minutes instead of days, catching risks that traditional methods miss, and turning compliance from a cost centre into a competitive advantage. We’ve proven that the web’s noise can become a signal that drives growth. This is the compounding advantage our partners are experiencing today.