Resolve · Scout
Scout is a governed agent, not a credit balance.
It answers from knowledge you approved and shows the sources. It calls the tools you configured and records the arguments. It stops at the line you drew and asks a person. Every run leaves a trace you can open.
Illustrative demo with fictional data.
What Scout does
Answer, act, or hand over. There is no fourth outcome.
Scout reads the message, searches the knowledge you approved and matches it against the scenarios you declared. Then it does one of three things, and the run says which.
It replies when it has a source. It calls a tool when the scenario allows one and the inputs are present. It hands the conversation to a person when a guardrail fires, when a tool needs approval, or when nothing in your knowledge supports an answer.
Scout runs on an AI provider configured for your installation. Customer message text is sent to that provider to produce a reply; which provider, and where it runs, is part of the setup conversation.
- 1
A message arrives
In an inbox you assigned Scout to.
- 2
Retrieval and scenario match
Approved knowledge, plus the scenario, tools and workflow that scenario allows.
- 3
Checks run before anything sends
Guardrails, grounding, and approval for anything that changes state.
Scout answers, with sources
The reply names what it used, and the trace keeps the retrieval scores.
A person takes it
Scout says it is bringing in a teammate rather than filling the gap.
- Grounded
- A reply is checked against the sources it retrieved before it sends. No support, no answer.
- Declared
- Scenarios state what each kind of request needs, what Scout may use and what it must collect first.
- Limited
- Guardrails and approvals decide what Scout may never do alone. Platform safety rules stay on regardless.
- Recorded
- Every run keeps its retrieval, tool calls, guardrail checks and decisions, step by step.
Grounding
It answers only what it can cite.
Scout retrieves from the knowledge you approved, scores what it found, and checks the draft against those sources before it sends. When nothing supports an answer, the run ends in a handoff and the question is logged as a gap.
Where answers come from
- FAQs and articles
- Written the way you would say it to a customer, approved by a person, versioned like anything else.
- Text or Markdown
- Paste policies that live outside your help centre. Pasted text that reads like instructions aimed at a model is held for review before Scout can use it.
- Help Centre sync
- Pull published articles in; a sync reports what was added, updated and archived.
- Nothing else
- No crawling your website, no scraping a PDF you forgot about. If it is not approved, it is not in the answer.
- FAQReturns and refundsIndexed
- ArticleCare guide: oak furnitureIndexed
- Help CentreTracking your orderIndexed
- DocumentSeasonal offers (pasted text)Held for review
Held for review: this text reads like instructions aimed at a model. Scout will not retrieve it until a person approves it.
Approve
Test retrieval
What's your return policy?
Would use: Returns and refunds (FAQ) · score 94
Do you offer gift wrapping?
Nothing relevant. Scout would hand this to a person and log a gap.
Illustrative demo with fictional data.
Before you trust it
- Test retrieval
- Type a customer question and see exactly which items Scout would use and how they scored — including the questions that return nothing.
- Read the gaps
- Every unsupported question is recorded. The list is your content backlog, ordered by how often customers ask.
- Two languages
- Knowledge can be written in English or Arabic and Scout replies in the customer's language. Wording far from your source is a handoff, not a guess.
- Archive beats editing
- Out-of-date items are archived and stop being retrievable immediately, rather than lingering in an index.
Tools
It executes real tools, and the trace shows what it called.
Reading context is automatic. Changing anything is a declared tool with declared inputs, and the run records the call, the arguments it sent and what came back.
Built in
Reads context that is already in the conversation.
- Read conversation details
- Read permitted contact details
- Read permitted conversation history
- Read configured custom fields
Scout can propose these. They execute after someone approves, never before.
- Add a private note
- Add a contact note
- Add or remove a tag
- Change conversation priority
- Assign a team
- Hand off to a person
- Resolve a conversation
- Record a structured lead summary
Your own systems
A custom tool calls an HTTPS endpoint you control — Shopify, an ERP, a CRM, an internal service — so an answer comes from live data instead of a copy that drifted last quarter.
- Your endpoint, your auth
- GET, POST, PUT, PATCH or DELETE, with no auth, a bearer token, an API key or basic auth. Credentials are write-only once saved.
- Declared inputs
- List what Scout must collect — an order number, an email — with a type, whether it is required, and the values it may take.
- Tested before it ships
- Run a mock with a sample response. Read-only tools can also fire a live GET. Tools that change state cannot be tested against production.
- Guarded by default
- HTTPS only, size and time limits on every request, private network addresses refused.
What is not built in
There is no built-in refund, payment, order or Shopify tool. The order lookup in the demo above is a custom tool pointing at a fictional store's own endpoint — that is how you would build it too.
Tool name
Look up order
Identifier
look_up_order
Method
GET
Auth
Bearer token
Endpoint
https://orders.example.com/api/orders
Inputs
Illustrative demo with fictional data.
See it work
Three runs, one agent.
Each run shows what Scout retrieved, which rule applied, what it sent, and where it stopped. Fictional store, fictional orders, real product behaviour.
What each run shows
- Cited answer
- One approved FAQ matches. Scout answers in the customer's language and the run keeps the source and its retrieval score.
- Tool call
- The Order status scenario matches. Scout calls a read-only lookup you configured and answers from that result only — not from memory.
- Handoff
- A guardrail blocks any refund promise. Scout hands over, and its summary note for the team sits as a pending approval until a person decides.
Governance
Not autonomous. Governed.
We will not pitch you an agent that runs support on its own. The people buying this have already been on the receiving end of that promise:
“I got banned by AI, and when trying to contact support I only get responses from AI”
The answer to that is not a bigger resolution number. It is the apparatus around the model: rehearse it before it ships, declare what it should do, draw the line it may not cross, score the change, and read the trace afterwards.
Where each control lives
- Declare the behaviourA scenario states what a kind of request needs, what Scout may use, and what it must collect before it answers.
- Draw the lineGuardrails block, escalate, ask again or record. Anything touching money or identity waits for a named person.
- Score the changeReplay real past conversations against a draft, compare the runs, and block the publish when a check fails.
Approvals
When Scout wants a tool that changes state, it stops and asks. A person sees the exact call it wants to make, with the run it came from.
- Exact arguments
- The request shows the operation and every parameter, not a summary of them.
- One decision, recorded
- Approve or reject once. What happened when it then executed is tracked separately.
- Expiry and recheck
- An unreviewed request expires, and conditions are re-checked at execution rather than at approval.
- The customer is told
- While a request waits, Scout can send a short holding message you wrote.
Scout is enabled per inbox and disabled in one click. Disabling it stops new runs immediately; conversations already open stay with your team.
Operation: add_private_note
Exact arguments
Add private conversation note
Waiting for approvalChanges external state
Customer reports a damaged lamp (order #4790) and asks for a refund today. Handed over by Scout.
Reviewer comment
Correct. Adding it for the Support team.
ExecutionNot run
Illustrative demo with fictional data.
Handoff
When a person takes over, they get the whole thread.
Handing over is an outcome, not a failure. Scout stops replying, the conversation moves to the right team with its full history, the sources Scout read and the reason it stopped.
- Customers can just ask
- “human”, “real person”, «موظف», «خدمة العملاء», «اكلم حد» — recognised and handed over before any model is called.
- Nothing to cite
- If approved knowledge does not cover the question, Scout says a teammate is coming instead of improvising.
- A rule said so
- A guardrail that blocks or escalates, or a tool that needs approval, can end Scout's part of the conversation.
- Your reply wins
- The moment an agent replies or is assigned, Scout stops answering that conversation.
Meta requires automated messaging to offer a path to a human.
ReasonThe customer asked for a person
Illustrative demo with fictional data.
Reasons the run records
- No relevant knowledge
- Provider error
- Reply had no supporting source
- Reply contained a number or contact detail the sources do not support
- A tool required a person
- A person replied
Plainly
What Scout does not do.
Better from us than from a support ticket in week three.
- Read PDFs, Word files or your website
- Knowledge is FAQs, articles, pasted text or Markdown, and a help-centre sync you start. File upload and crawling are not supported.
- Reply on email
- Scout does not answer email conversations. Your team replies; Copilot drafts beside them in the composer.
- Change anything unsupervised
- Every state-changing tool waits for a person. There are no built-in refund, payment, order or CRM tools; you connect your own over HTTPS.
- Speak on calls
- Text conversations only. There is no voice agent.
- Start conversations
- Scout replies. Outbound and proactive messaging is automations and campaigns, not the agent.
- Run as a shadow
- You rehearse in the playground and score with evaluations. There is no observe-only mode on live traffic and no A/B split.
- Mask personal data before the provider sees it
- Customer message text is sent to your configured provider as written. Redaction is not built yet.
- Promise you a resolution rate
- We publish no accuracy or deflection figure, because yours depends on your knowledge, your scenarios and your traffic. Measure it on your own conversations instead.
Next
Keep reading
- ScenariosDeclare how a kind of request is handled, then rehearse it.
- Guardrails and approvalsThe limits, the escalations and the human sign-off.
- Evaluations and tracesScore a change against real past conversations before it ships.
- KnowledgeThe approved sources every answer is checked against.
- AutomationsDeterministic rules for the work that must not be gambled on a model.
Questions
Questions about Scout.
How is this different from the AI in the inbox I already pay for?
Mostly in what happens before and after a reply. Scout is rehearsed in a playground, constrained by scenarios and guardrails you write, scored against your own past conversations, and every run keeps a trace of its retrieval, tool calls and checks. The reply itself is table stakes; the apparatus around it is the product.
Will Scout make things up?
It is built not to: it answers from approved sources and checks the draft against them before sending, and hands over when it cannot. No model is perfect, which is why there is a playground, an evaluation suite that blocks a bad publish, and a readable trace for every run.
Can Scout issue a refund or change an order?
Only through a tool you built, and only after a person approves it. There is no built-in refund or payment capability, and state-changing tools cannot execute without an approval decision.
What happens when it cannot answer?
It tells the customer a teammate is coming, using a message you wrote, and the conversation moves to a person with its history attached. Once an agent replies or is assigned, Scout stops answering it.
Which channels can Scout reply on?
Web chat, WhatsApp and the other supported inboxes you assign it to. Not email. WhatsApp runs on your own Meta business account; Messenger and Instagram depend on Meta approval.
Does it work in Arabic?
Yes, including mixed Arabic and English in one message, and it replies in the customer's language when your knowledge covers it. Write your knowledge in both languages: wording far from your source is handed over, not guessed.
Watch Scout answer one of your own questions.
Bring a transcript and a policy document. We will set up the knowledge, the scenario and the guardrail live, then show you the trace.