AI task spotlight | Edition no. 07: Online Search

Every two weeks, we spotlight an AI task from Unit21's task library, something many compliance and fraud teams are configuring and running inside their workflows today.

‍

The last edition covered a live build: one program's specific definition of structuring, described in plain English and backtested before it went anywhere near a queue. This edition goes the other direction. Online Search is the task nobody should have to build, because every program already runs it and it looks roughly the same everywhere. It ships as a template, pre-written, and the whole configuration is one screen and a preview. 

‍

Why Online Search Is the Step Everyone Does by Hand

Online search is the most universal manual step in an investigation. An alert fires, an entity is flagged, and someone opens a browser. They take whatever identifiers the file gives them, a name, a user name, a business name, an email address, a phone number, and start searching. They read what comes back. They decide whether the person in the results is actually the person on the alert. They write down what they found, and they keep the links so the conclusion holds up later.

‍

None of that varies much by program. Unlike structuring, where a remittance company and a payroll company are genuinely looking for different patterns, online search means the same thing everywhere. Which is exactly why writing it from a blank page is wasted effort.

‍

The part that actually consumes time isn't running the query. It's the judgment that follows: three people share the entity's name, and only one of them, or none of them, is the subject. And then the documentation: a paragraph an examiner can read, with sources attached, that says what was found and what wasn't.

‍

Introducing Unit21's AI Task: Online Search

Online Search: What it does

‍

Online Search ships pre-built. On the Tasks tab of any Alert AI Agent, Create Task, Online Search, Use Template, and the task name, the search prompt, and the entity data fields to search on are already filled in. From there it's a single screen: confirm which identifiers to search, cap how many queries the agent may run, pick a real alert, and hit Preview Task. The agent runs a grounded search on every flagged entity in that alert and hands back a formatted write-up before the task is ever saved to a queue.

‍

The agent automatically reviews:

• Every flagged entity on the alert, searched on the identifiers you select: first, middle, and last name, user name, business name, doing-business-as, email addresses, and phone numbers
• Public-record and open-source results for each entity, including people-search aggregators, government sites, professional registries, Wikipedia and Wikidata, and genealogical records
• Whether a result that shares the entity's name actually corresponds to the subject, or to a different person who happens to share it
• Adverse media, sanctions, and criminal-history signals tied to the specific identity combination on file, not merely to the name
• Other records on the alert, optionally passed in as cross-reference context so the search is informed by what you already know

‍

Online Search: What the agent outputs

‍

• One section per flagged entity, headed with the entity's name, a single-paragraph narrative of what was found, and every source link collected at the end of that section
• Numbered, clickable citations behind each finding, so an investigator can open the underlying page and verify the research directly
• An explicit no-match statement when the search can't support a conclusion, including named distinctions where a same-name result is not the subject, instead of a forced or fabricated match
• A full reasoning trail, every query the agent generated, which ones it executed against your cap, and each grounded search step, visible end to end
• A Task Preview against a real alert before the task is ever saved, so you see the output on your own data first

‍

Online Search: Why this matters

‍

Previewed against a real alert, the task did the work and then did the harder thing. For one entity it generated seven candidate search queries, executed three against the cap that had been set, and reported that public databases surfaced a hockey player and a set of genealogical records under a similar name, and that these appeared to be distinct from the subject. No verifiable adverse media. No sanctions. No criminal history tied to that specific identity combination.

‍

That is a real investigative result. A clean, documented negative is what closes an alert defensibly, and it is considerably more useful than a confident answer assembled from the nearest matching name. The failure mode of automated online search has always been the opposite: a same-name hit treated as a match, and an investigator left to unpick it.

‍

The query cap matters for the same reason. You decide how much searching the agent is allowed to do per task, so the work stays scoped and predictable rather than expanding to fill whatever the model feels like exploring.

‍

Because it's built with Unit21's Agentic Task Builder, this runs inside your existing workflow, in plain English, with no engineering ticket, and the prompt stays editable if your program needs it narrower. And because every step of its reasoning is visible, with every source linked, investigators can act on its findings, challenge them, or escalate them with the backing to defend that decision under examination.

‍

The searching gets done. Your investigators make the call.

‍

About the AI task spotlight series

‍

The AI Task Spotlight runs every two weeks. Each edition covers one task from Unit21's library, covering what it does, how it works, and who it's for. If a task is solving a real problem for one team, it can probably solve the same problem for yours.

‍

Want to learn more? Sign up for a demo of our AI. Alternatively, stay informed of our AI by signing up for our next AI Task Spotlight.

‍