Preparing a suspected marketplace fraud alert for human investigation with AI agents
Agent Swarm gathers the records behind a suspected marketplace fraud alert and hands an authorized reviewer a source-linked investigation brief.
Agent Swarm can gather the records behind a suspected marketplace fraud alert, flag missing or conflicting evidence against checks the team defines, and hand an authorized reviewer a source-linked investigation brief. The reviewer decides every next step. What it takes off the team is the assembly: pulling the alert, the account, the transactions, the policy in force and the engineering logs out of systems four teams own, so one person can read them in one pass.
The scope is one alert that has already been raised, by an existing detection system or by a person. Whether the activity is fraud is not this workflow's question. A seller appealing a listing decision is a moderation appeal, a payout that does not match its transactions is a payout discrepancy, and a failed incoming payment, a payroll discrepancy or a support escalation is a separate handoff with its own owner.
When another tool fits better
A specialist fraud or risk platform. Detection, risk scoring, device and network signals, rules engines and case management at volume are what those products do, with signals a general agent workflow does not carry. When the question is which alerts should fire at all, that is the tool.
The payment provider's own dispute and risk tooling. Evidence inside a processor's network, along with chargeback handling and the liability decision, stays where the money moves. For provider-side signals, its own tooling answers faster than a collection script.
Manual triage at low volume. Two or three alerts a month and an analyst who can open the account and the transactions in three tabs: an agent would put a document between the analyst and the tabs.
Agent Swarm fits a different shape: the cross-team assembly step around one alert. The alert sits in a detection system, the account record in an admin, the transactions in a ledger or at the payment provider, the applicable policy in a document store, and the relevant logs in another team's observability stack. Operations reads the policy, engineering reads the logs, and the person who has to decide is often neither. When those records sit apart, the work is gathering evidence for the owner who decides. Choosing an AI operating model sets out that split.
An illustrative workflow
This is a fictional, custom workflow the team builds. It is not a built-in fraud feature, not a detector and not a customer result: Agent Swarm ships no scoring model and no case management, and every integration, permission and threshold below is configuration the team sets up and tests.
1. An alert arrives from the system that already exists. The trigger is a rule, a model score or a person on the risk or operations team. The workflow does not decide what counts as fraud, scores no account and opens no case on its own.
2. The fraud or operations owner sets the scope of the investigation. The time window, the timezone its boundaries use, and which systems are the approved source of truth for each record. Written down, that definition makes the brief reproducible; left implicit, two analysts retrieve two different sets of records for the same alert.
3. Configured connections gather the authorized records. They collect the alert and the reason it fired, the account record, the transactions inside the window, the policy version in force on the alert date, and the timestamp each record carries. Scripts reach those systems through script connections, and credentials come from the credential broker, so a raw secret stays out of script source.
4. A swarm-script node runs the defined checks and emits JSON. Retrieval, date arithmetic and the comparisons run as code, so the same inputs produce the same output and no model reconciles a timestamp or totals a column. Reports from multiple sources documents that split between deterministic collection and agent synthesis.
5. An agent assembles the investigation brief. The Researcher turns the JSON and the source records into one document: what the alert states, what the records show, which checks flagged and what could not be retrieved. Every line links its source record with a read time, and a missing record stays empty rather than showing an estimated value.
6. A human-in-the-loop gate waits for the authorized reviewer. A human-in-the-loop gate pauses the workflow, posts a short summary with a link to the full brief, and resumes only on a person's response. The reviewer decides: no action, return for more records, or escalate to the risk owner, a senior investigator, legal or a payment partner. The gate carries a timeout and a fallback, and nothing proceeds on silence.
Evidence checks
Each check below is illustrative and belongs to the team that runs the workflow, not to the product. No check concludes fraud. It flags what the reviewer needs to see.
- A transaction the alert references cannot be retrieved. The alert names an identifier the ledger does not hold, in the window or at all. The brief states the identifier, the systems searched and the read time.
- Timestamps disagree across systems or carry different timezones. The same event is recorded minutes or hours apart, or one record has no zone. The brief shows both raw values, the zone each was read in and the gap, and leaves the explanation to the reviewer.
- Account details differ between two records. A name, an address or a payout account differs between the admin and the ledger. Both versions appear side by side with links, and neither is chosen.
- The policy version in force at the alert time is missing or ambiguous. The alert date falls before the first version on file, or inside a period two versions claim. The brief shows the dates as recorded and does not pick which rule applied.
- A record falls outside the requested window. A transaction, a sign-in or a log line sits outside the window the owner set, where it may still matter. The brief flags it with its date rather than dropping it silently.
Privacy, access and incomplete evidence
Account and transaction data are among the most sensitive records a marketplace holds, and an alert adds an allegation. Four rules belong in the setup, not in one analyst's habits.
Least-privilege read access. Each connection holds the narrowest scope that answers the approved questions: read on the tables and endpoints the owner named, and nothing that writes. Script connections scope a connection to an agent or a repo, so the sources behind an alert are reachable by the workflow that investigates it and not by every agent in the swarm.
No workarounds. When a system refuses access, the brief names the record and the owner who can grant it. Nobody pastes a screen of transactions into a chat or drops an export somewhere convenient so the workflow can continue. A named gap can be closed. A gap filled informally becomes a copy of personal data with no owner.
Named readers and roles. Who may read account and transaction data, who may open a brief, and which named owner is accountable for this workflow belong in the governance record. That guide also states the limit of the mechanism: RBAC gates user-token requests, not operator-key or agent calls, so the credentials a collection step holds are what bound an agent.
Retention. A brief holds personal data and an allegation. How long it is kept, where it lives and who may delete it are the team's policy, worth writing before the first run. Keeping every brief and every run log indefinitely is a liability, not an archive.
One section of the brief states what could not be retrieved, what came back empty and which checks could not run. A brief that resolves a gap by inference is worse than one that names it, because months later nobody can tell the two apart. Missing evidence limits any conclusion, and the brief says so where the gap is.
What stays human-owned
Five decisions stay with people in every version of this workflow.
- Whether it is fraud. Judgement about intent, pattern and consequence is a person's, made with the brief in hand. The workflow has no path to a finding.
- Any account action. Restricting an account, holding a payout or suspending a seller belongs to the risk owner, in the system that owns the action.
- Customer and partner contact. What a buyer or a seller is told, and by whom, belongs to the person who owns that relationship.
- Reporting to payment partners and authorities. Chargeback representment, law-enforcement requests and regulatory reporting follow the team's own process and, where it applies, counsel's.
- Policy and threshold changes. Loosening a rule or changing what counts as an alert is a business decision, not an output of a workflow.
Nothing here is a compliance control, a detection guarantee or a regulator-ready audit trail. It assembles evidence for the people who make those calls, and it contains no automatic enforcement.
Building blocks
Each piece below ships in agent-swarm and is documented. Pointing them at a suspected fraud alert is the illustrative part.
- Reports from multiple sources: deterministic
swarm-scriptcollection that emits JSON, then a Researcher agent that writes the brief from it. - HITL gates: the blocking
human-in-the-loopnode, its approve and reject branches, the timeout and the fallback. - Script connections and scripts credential broker: how a script reaches a detection system, an admin, a ledger and a log store while the secret stays in the broker.
- AI operations governance: RBAC roles, one named owner per workflow, and the rules for who may read account and transaction data.
- Team collaboration: one shared context for the operations, engineering and risk people on the same alert.
- Choosing an AI operating model: when a specialist risk platform or the payment provider's own tooling is the better answer.
Operator checks
A short list to settle before the first run.
- The alert source and the owner who receives the brief are both named.
- The approved systems of record are listed, with read-only scopes confirmed for each.
- The checks are written down and tested on past, non-sensitive cases, including one with a missing record.
- The brief template marks missing evidence and refuses to estimate a value.
- The gate has a timeout, a fallback path and a named backup reviewer.
- Retention and access for briefs are agreed before the first brief exists.
- The specialist platform, not this workflow, remains the detector.
Preparing marketplace moderation appeals for human review with AI agents
Agent Swarm gathers the permitted records, the dated policy and the original decision for a listing appeal, and a person decides.
Reviewing customer billing disputes with AI agents
Agent Swarm gathers the records behind a disputed charge, runs deterministic comparisons and hands finance a source-linked review packet. People decide.