> ## Documentation Index
> Fetch the complete documentation index at: https://docs.orderprotection.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Claim Automation
> How Order Protection's AI agent reviews claims, how automation rolls out, when it recommends versus acts on its own, and how to give it your brand voice.
Claim Automation is Order Protection's AI agent that reviews your eligible claims, gathers the relevant order, tracking, and policy context, and proposes a resolution the way a well-trained support agent would. Every suggestion is surfaced for review — and, once Order Protection has confirmed the agent is performing accurately for your store, low-risk claims can be resolved automatically.
This page walks through what claim automation is, how it rolls out, the settings you can view in **Settings → Automation**, how it shows up on your orders and claims, what the agent considers when it makes a decision, and how to give it a voice that sounds like your brand.
## What claim automation is (and why it matters)
When a claim comes in, the agent reads the context around it — the order, tracking, the customer's messages, and your standard operating procedures — and proposes the next step: approve, deny, message the customer, wait for more information, or escalate to a person.
Handling claims this way benefits both your team and your customers:
* **Faster resolutions.** The agent gathers context and drafts a resolution in about a minute, so straightforward claims don't sit in the queue waiting for a free agent.
* **Consistent policy application.** Every claim is evaluated against the same criteria and your own SOPs and brand voice, so customers get consistent outcomes no matter who — or what — handles the claim.
* **Human-in-the-loop by design.** Every store starts with the agent making suggestions a person reviews. Order Protection only moves a store to hands-off automation after a period of training, and monitors accuracy the entire time.
Automation gets smarter and more on-brand the more you invest in your [SOPs](/education-basecamp/settings/sops) and [Message Macros](/education-basecamp/settings/message-macros) — those are the raw material the agent learns your policies and voice from.
## How automation rolls out
Order Protection manages the rollout of automation for your store — you don't have to turn it on, tune it, or decide when it's ready.
Every store starts here. The agent reviews each eligible claim and recommends a resolution, but a person always makes the final call. Your team and Order Protection's own customer service team review those suggestions and give feedback.
That feedback trains the agent on your policies, voice, and edge cases. Order Protection watches accuracy and performance closely during this period.
Once the agent is consistently accurate for a given claim type, Order Protection graduates it to automatic execution, where the agent resolves qualifying claims on its own. Order Protection continues to monitor performance and will dial a claim type back to suggestion-only if accuracy dips.
Because Order Protection owns this rollout, automatic execution is enabled per store and per claim type on our side — not something you switch on yourself. The Automation settings (below) let you see exactly how your automation is currently configured.
## Automation settings
From the dashboard, go to **Settings → Automation**.
The Automation tab is organized into cards, each covering one part of how automation behaves. The rest of this section walks through them in order.
Most of these settings are **managed by Order Protection and shown here for transparency**, so you always know how your automation is configured. A few — like your agent's name and avatar — are yours to adjust, subject to your team-member permissions.
### Automation thresholds
This card is **read-only**. It's here to show you the evaluation criteria Order Protection uses when recommending or taking automated action, so you have full visibility into how your claims are handled.
At a high level, this card shows the criteria that gate automation: a claim-value range that determines which claims are eligible, and the share of claims eligible for automated handling. Once a store is on automatic execution, it also shows the share of eligible claims that can skip human approval and the minimum confidence a claim must clear to be auto-executed.
It also shows which **claim types** can be automated:
* Stolen, Damaged, Defective Item, Order Delivered Not Received, Wrong Item, Missing Item, Returned to Sender, Tracking Not Moving, and Wrong Address.
And the **resolution types** automation can use — Refund, Reship, and Store Credit (Refund and Reship are enabled by default; Store Credit is off by default).
A claim that spans multiple claim types is never auto-executed — it always goes to a person.
### Claim types, automatic execution, and notifications
For each claim type, this card shows how it's currently configured:
* **Enabled for automation** — the agent produces a suggestion for that type.
* **Auto-Execute** — the agent can resolve that type on its own, without waiting for a person.
* **Slack notification** — your team is pinged the moment a claim of that type is queued to auto-execute.
Order Protection manages these settings as part of the rollout described above: a claim type moves to auto-execution only once the agent is performing accurately for it, and reverts to suggestion-only if accuracy dips. When a type is on suggestion-only, the agent's recommendation still appears — it just waits for a person. You'll always see the current configuration here.
### Auto-execute delays (review window)
When a claim type is on automatic execution, a hold can be applied before the action actually fires. This gives your team a **review window** to step in and stop the action — or approve it early — before anything reaches the customer. A Slack alert is sent at the start of the window, so your team gets the full delay to intervene.
A delay is set independently for each action:
* **Resolve claim** — the refund, reship, or store credit itself.
* **Send message** — a reply to the customer.
* **Update shipping address** — a correction to where an order ships.
* **Validate offline claim** — verification of a claim filed for an order placed outside your store.
Delays range from immediate (zero) up to a 24-hour maximum.
A non-zero delay is helpful when a claim type is new to automatic execution — it gives your team time to watch the agent's actions before they fire immediately.
### AI model configuration
The AI model your store uses is managed by Order Protection and shown here for transparency — it's read-only, and you don't need to choose or tune it.
The one editable control is **Adaptive Thinking**, an on/off toggle that lets the agent reason through more complex decisions before acting. It trades slightly more processing for more careful reasoning on tricky claims, so it's best left on unless you have a specific reason to turn it off. It appears only on models that support it.
### Customize your AI agent
Give your agent an identity that customers and your team see attributed to automated actions:
* **AI Agent Name** — the display name shown on automated resolutions and messages.
* **AI Agent Avatar** — an optional image for the agent.
Attribution follows who did the work: when a person triggers an action, that person's name shows; when the agent completes an action fully on its own, your configured agent name shows instead. If you don't set a name, a default display name is used. Editing these requires the appropriate customization permission.
Give the agent an on-brand name so automated resolutions feel like part of your support team, not a faceless bot. This pairs naturally with the [brand voice](#brand-voice-and-agent-personality) you set below.
## How it works on your orders and claims
On any individual claim, automation moves through a predictable lifecycle: **analyzing → suggestion → review (or auto-execute countdown) → resolved.**
While the agent works, the claim shows an **"AI is analyzing this claim…"** banner (it can take up to a minute), with a **Stop** control for permitted users. The view updates live — no page refresh needed.
From there, one of two things happens, depending on where that claim type is in its [rollout](#how-automation-rolls-out):
* **Human approval.** The agent's suggestion waits as a recommendation until someone on your team acts on it. Every claim type starts here.
* **Automatic execution.** Once Order Protection has graduated a claim type to full automation, the agent resolves qualifying claims on its own, after any review-window delay.
A teammate monitoring activity can also watch what's queued to auto-execute (with a live countdown) and what recently fired in one place.
### The AI Suggestion card
When the agent finishes, its recommendation appears as an **AI Suggestion** card showing a recommended action, a confidence indicator, a timestamp, the Claim ID, and a plain-language reason for the recommendation. A **View Details** link is available to users with the right permission.
The recommended action is one you'd recognize from handling claims yourself — such as **Approve Claim**, **Deny Claim**, **Send Message**, **Wait**, or **Escalate** (a few other actions, like updating a shipping address or validating an offline claim, can also appear).
Your team has full control over what happens next:
* **Accept** runs the suggestion as-is.
* **Accept without Message** applies the resolution but skips the drafted customer message.
* **Reject** discards it and opens a short prompt — *"Why wasn't this suggestion helpful?"* — so your input improves future suggestions.
* **Regenerate** asks the agent to take another pass — on its own, or with your guidance (see [Course-correcting a suggestion](#course-correcting-a-suggestion) below).
Before accepting, the reviewer sees supporting detail: a **Resolution Summary** breaking down the proposed refund, reship, or store credit per item, and for message actions a **Suggested Message** showing the exact customer-facing text with real amounts filled in. For a **Wait** action, the card shows what the agent is waiting for — for example, a customer reply, a reply from your team, or a shipping update — and moves the claim to **On Hold**.
If a claim is set to auto-execute with a delay, the card shows an **"Auto-Executing"** banner with a live countdown. You can accept now to run it immediately, or reject to cancel before it fires.
### Course-correcting a suggestion
A suggestion won't always be right. When one is off, you don't have to accept a bad resolution or reject it and start over — you can **course-correct** the agent and have it try again with your input.
Add a short note telling the agent what to reconsider, then regenerate. The agent re-runs its analysis with your guidance factored in and returns an updated suggestion. For example:
* *"We need clearer photos of the damage before approving — ask the customer to resend."*
* *"This order is outside our coverage window, so deny instead of reshipping."*
* *"Offer store credit here rather than a refund."*
You can course-correct more than once, refining until the recommendation is right. Because a quick note usually gets you to the correct answer faster than starting from scratch, it's often worth course-correcting before rejecting a suggestion outright.
The guidance you give during course correction feeds back into the agent, helping it handle similar claims correctly on its own over time.
### Automation details
The **Automation Details** view is the deeper, read-only record of a single automation run — useful for auditing why the agent did what it did. In plain terms, it shows:
* **The action taken** and the suggested resolution.
* **AI Decision vs. Agent Decision** — the *AI Decision* is what the agent recommended (approve, deny, resolve externally); the *Agent Decision* is what happened to that recommendation (Accepted by a person, Rejected by a person, or Auto-Executed).
* **A confidence score** and a plain-language **reason** for the decision.
* **A compliance report** describing how the decision lines up with your policies and SOPs.
It also includes a **Confidence Scorecard**: an overall quality score for the decision that reflects a blend of signals — such as how much supporting information was available — along with a fraud-risk level (Low, Medium, High, or Unknown) and a clear **Requires Human Review** flag when a safety check routes a claim to a person instead of resolving it automatically.
## What the system analyzes
This section is intentionally high-level. Order Protection shares the **categories** of information the agent considers, but not the exact criteria, weights, or thresholds — that's a deliberate choice that keeps the system fair and resistant to manipulation.
When the agent evaluates a claim, it weighs many kinds of information together, including:
* **The claim itself** — its type, the items involved, and the claim value.
* **The order** — fulfillment, shipping, and tracking status.
* **Customer history** with your store, including repeat or duplicate filings.
* **The conversation** — the claim's message history and any internal notes.
* **Fraud and risk signals.**
* **Evidence** the customer uploaded, such as receipts or damage photos.
* **The strength and completeness** of the available information.
* **Your policies, SOPs, and precedent** — how similar claims have been handled before.
* **The model's own confidence** in its decision.
Each decision carries a confidence score, and independent safety checks (guardrails) can route a claim to a human regardless of that score. These checks are intentionally conservative and always favor sending a claim to a person when in doubt — for example, when a claim is unusually high-value or the available evidence is thin. The full set of checks isn't published in detail.
## Brand voice and agent personality
Two per-store documents shape how the agent sounds, shown read-only in the **AI Agent Voice and Personality** card:
* A **Brand Voice** document — how the agent should sound.
* A **Store-Specific Instructions** layer — the specific rules, preferences, and exceptions you want it to follow.
You can view and copy both. They're generated from your customization inputs — specifically your active [SOPs](/education-basecamp/settings/sops) and your customized [Message Macros](/education-basecamp/settings/message-macros). The more you invest there, the more on-brand and precise the agent becomes.
If your Brand Voice currently shows something like **"Overrides — N/A"**, that's normal for a new setup, not an error. It simply means no custom deviations are on record yet, so the agent falls back to sensible defaults for your store.
A well-defined brand voice turns a generic bot into something that sounds like *your* support team and knows your boundaries — consistent tone, clear do's and don'ts, and explicit escalation triggers.
The example below is a generic starting point you can adapt — not a required format, and not Order Protection's internal template. Fill in the bracketed placeholders with your own details.
```markdown theme={null}
# Brand voice notes — [Your Store Name]
## Personality & tone
- Warm, upbeat, and concise — write like a helpful human on our support
team, not a corporate bot.
- Use the customer's first name when it's available. Keep replies to
2–4 short sentences.
- Match our brand energy: [e.g., "friendly and playful" or
"premium and understated"].
## Do's
- Acknowledge the customer's frustration before explaining the next step.
- Use plain language; skip logistics jargon like "carrier scan exception."
- Offer one clear next action in every message.
## Don'ts
- Don't promise refund or delivery timing we can't guarantee.
- Don't use ALL CAPS, excessive exclamation points, or more than one emoji.
- Don't blame the carrier or the customer.
## Words we use
- Refer to customers as "[members / guests / friends]."
- Call our products "[Collection name]," never "SKUs" or "units."
# Store-specific rules & overrides — [Your Store Name]
## Resolution preferences
- Prefer a reship over a refund for lower-value orders; offer store
credit as the first alternative.
- For fragile items, request a photo before resolving.
## Overrides
- Extend the "no customer response" window for international orders.
- Never auto-approve claims for [specific high-value collection] —
route them to a person.
## Escalation preferences
- Escalate to a human when the claim is high-value, the customer
mentions a chargeback or legal action, or the same customer has filed
several claims in a short window.
- Where to send it: assign to the "[queue or team name]" queue.
```
Want a custom Brand Voice built out for your store? Your Customer Success Manager can help you shape it from your SOPs and macros.
## Questions?
If you have questions about where your store is in its automation rollout, what the settings mean, or building out a custom brand voice, reach out to your Customer Success Manager or your shared Slack channel.