Resolve
Limits Scout cannot talk its way past.
A guardrail is a hard constraint evaluated after Scout decides and before anything executes. Which tools exist at all, how much they may move, what must never be said, and who signs off. When a check fails, the conversation goes to a person — with the reason attached.
Illustrative demo with fictional data.
Four kinds of limit
Written once, for the whole workspace.
There are far fewer guardrails than scenarios, and they are owned by whoever is accountable for money and claims — not by whoever wrote the case.
- Tool allowlist
- A tool Scout has not been given does not exist. Read-only lookups can be on while anything that writes — refunds, cancellations, address changes — stays off, or stays on with a ceiling.
- Value ceilings
- A number above which a person decides. Refunds over a limit, discounts over a percentage, orders over a value. Below the line it executes; above it, it waits.
- Content rules
- Sentences that never leave, regardless of what a scenario says or how a customer phrases the request. No guaranteed delivery date the tool did not return. No price without a catalogue match. No legal or medical advice.
- Escalation triggers
- Conditions that end the automated path immediately: a complaint, a repeated question, a request for a person, a detected language your team has not approved Scout for.
Approvals
Approval is a state the action sits in, not a notification someone misses.
When Scout proposes something above a ceiling, it does not send and apologise later. The action is held, the customer gets a reply that is true, and a named person decides.
- The request carries its reason
- The proposed call, its arguments, the guardrail that stopped it, and what Scout read before proposing it. The approver is not guessing.
- Approve, edit or decline
- Editing is normal: approve the refund but add a return pickup. The executed action is the edited one, and the edit is in the record.
- It executes under a person's name
- The audit line names who approved it and when. “The AI did it” is never an available answer.
- Declines are data
- A guardrail that fires constantly is either set wrong or protecting you from a scenario that needs rewriting. Both show up in the run history.
refunds.create
order #4790 · 1,240 EGP · reason: damaged on arrival
Proposed by Scout 11:04 · edited and approved by Nour Hassan 11:06 · executed 11:06
Escalation
A human path out of automation is required, not optional.
This is not a BlendDesk opinion. Meta's own messaging policy requires that a business using automation gives people a way to reach a human. Build it in and the compliance question stops being a question.
Meta requires a human escalation path from automation.
The customer service window is 24 hours. Outside it, only approved templates can be sent.
That makes escalation a timing problem as well as a policy one. A conversation that sits in a queue past the window cannot be answered with a free-form reply at all. BlendDesk surfaces the remaining window on the thread, so a handoff happens while a normal reply is still possible.
- The whole thread goes with it
- The person who picks it up sees what the customer said, what Scout read, which tools ran and what was blocked. Nobody asks the customer to repeat themselves.
- Escalation is instant and logged
- A trigger fires, the conversation is assigned, and the reason is written into the run. There is no state where a customer is stuck with an agent that cannot help.
- Reachability is checked
- If nobody is on shift, Scout says so honestly and the conversation is queued for the next person, rather than being held open behind a cheerful holding message.
Off switches
Turn it down, turn it off, put it back in draft.
The audience for AI in customer messaging has been burned by systems with no visible controls. So the controls are the product, not a settings page nobody can find.
- Per channel
- Let Scout answer on web chat while WhatsApp stays human-only. Channels are switched independently.
- Per scenario
- Order status on, refunds off. Unpublishing a scenario takes it out of matching immediately.
- Suggest instead of send
- Scout drafts into the composer and a person presses send. Same retrieval, same citations, no autonomous reply.
- Workspace stop
- One switch takes Scout out of every conversation. In-flight approvals stay in the queue for a person; nothing executes on its own.
The record
Every block, every approval, every override is written down.
A guardrail that cannot be audited is a promise. The run trace lists the checks that ran, which one failed, what was proposed, who decided and what finally executed.
- Scenario matchedMatchedDamaged product complaints · 88%
- Knowledge search2 sourcesReturns and refunds · approved 12 Oct
- Tool proposedHeldrefunds.create · 1,240 EGP · order #4790
- Refund ceiling 500 EGPBlockedAbove the limit — blocked before execution
- Escalated to SupportHanded overNour Hassan · 11:04
- Approved with an editApprovedRefund plus return pickup · Nour Hassan 11:06
- Reply sentSentCustomer told the refund is on its way
Illustrative demo with fictional data.
Questions
About guardrails and approvals
What happens to the customer while an action waits for approval?
They get a real reply immediately — one that says a teammate is picking it up — and the clock on the 24-hour service window is visible to your team. The customer is never left on an open question while an action sits in a queue.
Can Scout be talked out of a guardrail?
No, because the guardrail is not part of the conversation. It runs after Scout has decided and before anything executes, against the proposed action and the drafted text. Persuading the model does not change what the check does.
Who can approve?
Whoever you give the permission to. Approvals are assigned to people and teams, the approver's name is on the executed action, and the audit line is part of the conversation record.
Does Meta actually require a human escalation path?
Yes. The WhatsApp Business Messaging Policy, updated 23 September 2026, requires businesses using automation to provide a way for people to reach a human. BlendDesk ships that path as a default rather than as something you remember to add.
What if we want Scout to never send anything on its own?
Run it in suggest mode. Scout does the retrieval, the tool lookups and the drafting, and your team presses send. Everything else on this page still applies — the same guardrails run on the draft.
Bring your customer conversations together.
Tell us a little about your team. We'll show you how BlendDesk could fit the way you already talk to customers.