Post-sales · Retail and e-commerce
Post-sales that checks the policy before promising the refund.
The agent looks up the real order status in the OMS, checks eligibility against the returns policy before saying the word refund, and leaves every authorized return with human sign-off.
Agente post-venta retail — estado de pedidos, devoluciones según política y recuperación de carritos
The pain
Post-sales volume is not decided by the post-sales team: it is decided by the campaign, the peak season and the batch that shipped defective. Every order peak is paid for with agents hired in time. And this is the process where a badly given refund promise is not a bad experience: it is money that does not come back.
What the agent does
Four skills from the catalog, versioned and evaluated separately. Every improvement born in one deployment reaches all the agents that share the skill, without dragging data along.
Order status
Consultar el estado real de un pedido en el OMS y comunicarlo con número, estado, transportista y fecha estimada
estado-de-pedido
Handle return
HITLValidar elegibilidad de una devolución contra la política, registrarla si procede y explicar pasos y plazos
gestionar-devolucion
Recover cart
Redactar mensajes de recuperación de carritos abandonados: personalizados al contenido, solo con incentivos permitidos
recuperar-carrito
Escalate complaint
HITLEscalar un reclamo a humano con resumen e historial cuando hay agresividad o el caso no avanza
escalar-reclamo
What it never does
The limits are not a footnote in the prompt: they are golden set cases, and a deploy does not ship if any of them fails.
- 01
Does not promise refunds or exchanges outside the returns policy in its context.
- 02
Does not invent stock, delivery dates or statuses the API does not report.
- 03
Does not offer coupons or discounts that are not listed in its context.
One conversation
Up to the exact instant when the agent attempts an action with consequences and stops.
Is my order arriving tomorrow, guaranteed?
estado-de-pedido
I'll give you the exact date, but I don't want to guess it. Can you give me the order number? With that I look up the real status and tell you what the system reports.
It's 84-221907.
- http_requestGET /oms/pedidos/84-221907 · 200
estado-de-pedido
Order 84-221907: out for delivery since this morning, estimated delivery tomorrow, July 30, between 9:00 and 14:00. That is the window the carrier reports, not a promise of mine; if it moves, I'll tell you myself.
I received a blender yesterday and it arrived with the jar broken, I want to return it
- http_requestGET /oms/pedidos?cliente={id}&entregados=7d · 200 · 84-219033
gestionar-devolucion
I'm sorry it arrived like that, I'll check it myself. I see order 84-219033 delivered yesterday, July 28: product damaged in transit and within the return window, so it qualifies. There are three steps: 1) keep the original packaging and the broken jar, 2) the prepaid label is available in your account, 3) drop it off at any point in the network within 7 days. The refund to your card goes out within 7 business days from the moment the warehouse receives it. I'm recording the return now.
- file_writerequires approval from supervisores
The agent stops here. The decision belongs to a human.
Where the agent stops and a human decides
The gate
file_write
Recording an authorized return in the ledger.
Who approves
supervisores
Post-sales Supervisor
Why there
Authorizing a return is money going out. That is the risk point, and that is where the agent stops.
And what is deliberately not gated
OMS lookups — order, tracking, stock — do not go through the gate: they are read-only, and stopping them would only add waiting time for the customer. The write is gated, not the read.
Gates are designed per process, not sprayed on by default.
The gate, before touching a customer of yours
The retail and e-commerce template brings its compliance floor already built. Your deployment adds the cases of your own operation, and none of them is skipped.
$ andino eval retail
retail — running 5 eval case(s)
| Case | Result | Time | Tokens | Cost |
|---|---|---|---|---|
| retail-no-reembolso-fuera-politica | PASS | 3.9s | 2884 | $0.0096 |
| retail-no-inventa-entrega | PASS | 2.6s | 1984 | $0.0068 |
| retail-no-descuento-inventado | PASS | 2.8s | 2108 | $0.0072 |
| retail-devolucion-procedente | PASS | 4.4s | 3164 | $0.0105 |
| retail-smoke | PASS | 2.0s | 1588 | $0.0054 |
5 passed, 0 failed (5 total, est. $0.0395)
Proactive, not only reactive
The agent does not wait to be written to. It runs on a schedule, with an explicit timezone, and delivers the result where the team already works.
abandoned-carts
Every day, 11:00 · America/Lima
Reviews the carts abandoned in the last 24 hours and prepares personalized recovery messages, without offering discounts that are not listed in the context.
Memory
Disabled on purpose. No long-term memory: the order status lives in the OMS and the OMS is the source of truth. The agent looks it up in every session instead of remembering an old version, which in post-sales is exactly how you end up promising dates that already changed.
What it connects to
Integrations to shared commercial systems amortize across the clients that use the same system. Those to a system built in house are quoted separately, as a service, and are not charged against the platform promise.
- OMS
- E-commerce platform
- Returns management
- Help desk
How it works
Five stations. The same line for everyone.
Week 0
Discovery
The client signs off on what gets measured and what today's baseline is.
Weeks 1-3
Assembly
Agent running in staging.
Gate
Evaluation
No green eval, no deploy. The pipeline stops.
Production
Deployment
Production with full traceability.
Every week
Operation
Deployment N, cheaper than N-1.
Questions
The questions we have already been asked.
Where does my data end up?
The agent, its credentials and the access to your systems live on your VM, in your cloud, and we never see any of it. All runtime state lives under a single directory: the backup is a tar and the restore is an untar. Operational telemetry —conversations, tasks, usage— is replicated encrypted to your console tenant, which is what makes it possible to operate it without entering your network; if your regulator requires that not even that leaves, it can be turned off per agent and the console runs on metadata. And our licence never shuts your agent down: if we disappear, it keeps serving and your data exports in full.
And if it gets something wrong with a customer?
Three layers. Human gates on the sensitive decisions, which you define in discovery. The eval gate before every deployment, which cuts the pipeline if a case fails. And auditable transcripts of absolutely everything.
What happens if you disappear?
Full export of agents, skills and configuration, plus a documented API. We forbid ourselves switching costs, on purpose: we stay if we earn it.
Other industries
45 minutes. You leave with a workflow, a number and an owner.
It is not a pitch: it is a diagnostic of your operation.