Atlacis

One AI assistant. Three workflows.

A single AI handles returns triage, abuse detection, and evidence assembly — staying inside its boundaries with human approval on every high-risk decision.

What the AI does in each workflow

Returns Triage

Eligibility + routing

Inputs

  • Order data (items, value, date)
  • Customer history (return frequency, lifetime value)
  • Channel metadata (Shopify, Amazon, eBay, Walmart)

Boundaries

  • Operates within version-controlled rules
  • Can't approve cases above risk threshold
  • All scoring factors are visible and logged

Outputs

  • Risk score with contributing factors
  • Rule match (version + conditions met)
  • Routing: auto-resolve or escalate

Human Review

  • Cases above threshold go to human review queue
  • Override requires actor ID + documented reason

Abuse Detection

Pattern detection + flagging

Inputs

  • Cross-channel return patterns (frequency, timing, value)
  • Customer behavior signals (serial returns, item swaps)
  • Historical abuse indicators from resolved cases

Boundaries

  • Flags patterns — doesn't deny or restrict customers
  • Can't modify customer accounts or block transactions
  • All flags include explanation and confidence level

Outputs

  • Abuse risk classification (low / medium / high)
  • Specific signals that triggered the flag
  • Recommended action with rationale

Human Review

  • High-risk flags require human review before action
  • Rule changes based on patterns need admin approval

Evidence Assembly

Chargeback evidence + formatting

Inputs

  • Order confirmation and invoice data
  • Shipping proof and carrier tracking
  • Customer correspondence (support tickets, emails)

Boundaries

  • Can't submit evidence to processor without sign-off
  • Can't modify original source documents
  • Formats to processor specs (Stripe, Adyen, Visa TC40)

Outputs

  • Processor-formatted evidence pack
  • Completeness score with missing-item checklist
  • Draft response narrative for reviewer

Human Review

  • Evidence pack requires reviewer approval before submission
  • Reviewer can edit, append, or reject any document

Four layers of control

Rules, approvals, transparency, and logging — from first action to final resolution.

Rules Engine

Every AI action is bounded by configurable, version-controlled rules. The assistant can't exceed its scope.

Approval Flows

High-risk actions need human approval. The AI surfaces recommendations — people make the call.

Explainable Signals

Every risk score and flag includes the specific signals behind it. No opaque decisions.

Activity Log

Every AI action, human override, and system event recorded with actor, timestamp, and reason. Fully exportable.

AI + human, in sequence

01

Ingest

New case arrives via API, webhook, or sync. The AI extracts structured data.

02

Evaluate

The AI computes risk score and matches to versioned rules. All factors logged.

03

Route

Low-risk cases auto-resolve. High-risk cases enter the human review queue.

04

Assemble

The AI drafts evidence packs, comms, and recommended actions.

05

Approve

Human reviewer evaluates the recommendation. Approve, reject, or escalate.

06

Record

Every decision, override, and outcome written to the activity log.

See how the AI handles your workflows.

Book a demo and we'll walk through AI setup, rules, and approval flows for your use cases.