Order & delivery care

Resolve “where is my order?” before it becomes a ticket

Order status questions are the most predictable contacts you receive and the least valuable use of a support agent. Worse, they surge exactly when your team is already stretched. This blueprint resolves the routine order journey from live system data, handles approved changes and returns inside value thresholds, and pulls a person in when the money or the relationship warrants it.

AI WORKERSORCHESTRATED PROCESSHUMAN AUTHORITY3 roles, trust-scoped6 stages, each with an owner1 decision ownerOrder workerResolution workerException worker1RecogniseOrder worker2Read live stateOrder worker3ResolveResolution worker4Handle exceptionException worker5RecoverRetention special…6Close the loopResolution workerRetention special…CONTROLS AROUND EVERY STAGE4 guardrails · 4 decision rules · 5 channels · tamper-evident audit
The blueprint at a glance — worker roles feed the orchestrated stages, and the dashed paths are the escalations into human authority.
The team
3 AI worker roles escalating into 1 human decision owner.
The process
6 orchestrated stages, each with a named decision owner.
The channels
Web chat, SMS, WhatsApp, Email, Voice on one shared context.

The shift

What breaks today, and what changes

The challenge

  • Status and delivery questions crowd out the contacts where a human actually changes the outcome.
  • Peak season volume arrives faster than seasonal hiring can absorb it, and quality drops as the queue grows.
  • Customers repeat their order details on every channel switch because context does not travel with them.

What changes

  • Routine order questions resolve in the conversation, using live order and carrier data rather than a static FAQ.
  • Approved changes, returns, and reorders complete inside configured value thresholds without a queue.
  • High-value orders, repeat contacts, and negative sentiment reach a person with the full history attached.

The team

Every seat has one job and a defined level of autonomy

This is what gets configured in Team Canvas: role-scoped AI workers, the humans they escalate into, and a trust level on each connection that sets what a worker may do on its own. Trust starts tight and widens on evidence.

AI workers

  • Order worker

    Looks up live order and shipment state and explains the next real milestone.

    Owns
    Order visibility
    Trust level
    Acts within policy — read-only, after identity check
  • Resolution worker

    Executes approved changes: address edits, returns, replacements, and reorders.

    Owns
    Routine resolution inside thresholds
    Trust level
    Acts within policy up to value threshold, then approval required
  • Exception worker

    Detects lost, damaged, and stuck shipments and opens the right carrier or warehouse case.

    Owns
    Exception classification
    Trust level
    Acts with approval on anything with a cost attached

Human decision owners

  • Retention specialist

    Takes high-value, repeat-contact, and unhappy customers with the full journey visible.

    Owns: Goodwill and service recovery

The process

The conversation starts the work. The workflow finishes it.

Each stage records the decision that was made and the role accountable for it — which is what makes the run reviewable afterwards rather than a black box. Where the platform executes a stage rather than a person or agent performing it, that is named too, and accountability still sits with the role.

  1. 1

    Recognise

    Is identity confidence sufficient to expose this order?

    Order worker

    Run by integration

  2. 2

    Read live state

    Is this order on its normal journey, or is it an exception?

    Order worker

  3. 3

    Resolve

    Is the requested action inside the configured value and policy threshold?

    Resolution worker

  4. 4

    Handle exception

    Carrier case, warehouse case, replacement, or refund?

    Exception worker

  5. 5

    Recover

    Does order value, sentiment, or contact history warrant a human now?

    Retention specialist

  6. 6

    Close the loop

    Are the helpdesk, order system, and customer all updated and consistent?

    Resolution worker

    Run by workflow

Decision rules

The non-negotiables encoded in the workflow, not left to a prompt.

  • Identity is verified before any order detail is disclosed, on every channel.
  • Order status claims come only from live order and carrier data — never from an inferred estimate.
  • Refunds, credits, and goodwill above the configured threshold require human approval before they are offered.
  • Negative sentiment, third contact on one issue, or high order value routes to a person with the full history.

Controls & guardrails

What makes this safe to run in production, and provable afterwards.

  • Identity checks before disclosure and value thresholds before action.
  • Live system data as the only source for any status claim.
  • Sentiment and repeat-contact triggers that force a human handoff.
  • Auditable writeback so the helpdesk record matches the conversation.

Platform capabilities

What this blueprint runs on

Nothing here is bespoke. Each blueprint is a configuration of the same platform, which is why the second one you launch reuses the governance you already reviewed.

Channels

  • Web chat
  • SMS
  • WhatsApp
  • Email
  • Voice

Systems it reads and writes

  • Ecommerce & order platform
  • Carrier tracking
  • Helpdesk
  • Payments & refunds

Measurement

What we instrument from day one

These are the measurements the pilot puts in place, not benchmark results. You set the targets against your own baseline — and the same instrumentation is what decides whether a worker's trust level widens.

Containment rate

Resolved without a human

Tracks the share of order contacts fully resolved in conversation, split by contact reason so you know what is actually being absorbed.
Repeat-contact rate

Same issue, second touch

The honest counter-metric to containment: a contact that returns was not resolved, however it was closed.
Threshold discipline

Approvals vs auto-actions

Shows how many refunds and goodwill actions ran automatically versus went to approval, and whether the thresholds are set correctly.

Rollout

How this one goes live

A deliberately narrow start, supervised, with autonomy widened per role once the record supports it.

  1. Weeks 1–2

    Status and tracking on one channel

    Answer order status and delivery questions from live data on web chat. Read-only, so the grounding and identity checks are proven before anything writes.

  2. Weeks 3–4

    Approved actions inside thresholds

    Enable address changes, returns, and reorders up to a deliberately low value threshold, with everything above it going to approval.

  3. Week 5+

    All channels, peak-ready

    Extend across messaging, email, and voice with shared context, and raise thresholds where the audit record supports it before peak season.

Questions we get asked

Order & delivery care, in practice

How do you stop the AI inventing a delivery date?
Status claims are sourced from live order and carrier systems through the integration layer, and the workers have no authority to estimate. If the system of record has no date, the worker says what is actually known and, where useful, opens a carrier case rather than guessing.
Can it issue refunds?
Within thresholds you configure. Below the threshold the resolution worker completes the refund and writes it back; above it, the action requires human approval before it is offered to the customer. Every automatic and approved action is recorded with the reasoning.
How does this handle peak season?
Capacity is not staffed by headcount, so the same blueprint absorbs surge volume across every channel at once. Containment on routine order questions is what frees your team for the exceptions that actually need judgement during peak.

What changes in Retail & ecommerce

Peak volume is absorbed by containment on routine order questions, with refunds and goodwill gated by value thresholds.

Read the full Retail & ecommerce analysis

Related blueprints

Next step

Build this team around one measurable win.

We map your version of this process, compose the worker roles, connect your systems, and agree the escalation points with the people who own them — then run it supervised.