Quick-commerce customer support that reads the order before it answers.
Where is my order, a short-picked bag, expired curd, a delivery marked done that never arrived, a UPI debit with no order. Settl's AI agent reads the packing record, the rider's position, the payment gateway and the customer's refund history, then answers, credits or refunds within the limits you set — and sends anything bigger to a person with one click to approve.
The live ETA, then the credit your policy promises.
Most of a quick-commerce queue is one question. Settl reads the order's status and the rider's live position and answers with the real ETA, on whichever channel the customer used. When the order is past the window your policy defines, the agent applies the delay credit itself — a ₹50 wallet credit after 30 minutes, if that is your rule — and tells the customer. Nobody has to ask, and nobody has to approve a credit the policy already allows.
Each item claim is decided from what the order says happened.
The agent does not take a claim on trust or refuse it on suspicion. It reads the packing record, the proof of delivery and the customer's own history, then applies your policy the same way every time.
- Short-picked — 15 of 20 packs scanned at the store: refund the five, automatically if under your cap.
- Damaged — 6 of 12 eggs broken, photo attached: partial refund for six, automatically.
- Expired — a curd batch past its date: refund the item, and open an expiry-audit task for that store.
- Wrong item — mustard oil instead of sunflower: an express replacement from the store if it has stock.
- Non-returnable — fresh produce or apparel outside the window: the refusal explained, with the policy cited.
Below your cap it just happens. Above it, one click from a person.
Every refund and credit goes through one policy gate. Set an automatic limit — ₹300 per refund, ₹100 per goodwill credit — and the agent settles anything under it on the spot. Anything above lands in a Needs-approval queue with the evidence laid out: the packing scan, the rider's GPS at drop, the gateway record. A ₹310 refund is one click; a ₹2,460 “marked delivered, not received” claim whose rider was 1.4 km from the address is one click, and the rider is flagged.
Refunds and credits on an order can never add up to more than the order was worth, however many channels the customer tries.
Trust is a signal, not a gut feeling.
A customer with six missing-item claims in thirty days is not treated like one with none. Refund history sets a trust level on every customer, and the policy reads it: good customers get the automatic refund, customers on watch get asked for a photo, and risk-level customers go to a Risk review queue — politely, with the same facts in front of the reviewer.
The questions that need another system, answered from that system.
A UPI debit with no order is answered with the reversal date from the gateway, and escalated to Payments if that date passes. A coupon that did not apply is credited for the difference. A rider who asked for ₹50 extra on a COD delivery gets a complaint logged and a goodwill credit for the customer, and Last-mile Ops sees it on their board.
The order stream opens the case before the customer does.
A rider stuck in rain, an item out of stock at packing, a payment debited twice — these show up in your order events before they show up in your inbox. Settl opens the case from the event, decides what the policy allows, and messages the customer first: a new ETA and the delay credit, or a substitution they can accept with one reply.
WhatsApp, in-app chat, voice, email, social
The channels your customers already use, on one case per order issue.
Status, rider, packing record, proof of delivery
Read through an adapter to your order system before the agent says a word.
Refund, partial refund, wallet credit, redeliver, reassign rider, substitute
Reads run automatically; writes follow your policy and its cap.
Patterns across cases
Short-picks at one store, a bad batch across stores, one rider's not-received claims, a UPI spike.
English, Hindi, Hinglish on voice
Callers switch mid-sentence; chat and WhatsApp answer in the customer's language.
Containment, CSAT, hotspots, leakage
Delay hotspots by zone and rider, refund leakage by store, and plain-English questions over all of it.
Frequently asked
Can the AI issue refunds on its own?
Yes, under the cap you set, and only when the order's data supports the claim — the packing scan, the proof of delivery, the photo. Above the cap, or when the customer's refund history puts them on review, a person approves with one click and the evidence in front of them.
How does it know the order is actually late?
It reads the order's status and the rider's live position through an adapter to your order management system, and compares them with the promised window in your policy. The answer is the real ETA, not a script.
What about customers who claim missing items every week?
Each customer carries a trust level built from their refund history. Risk-level claims go to a Risk review queue instead of being refunded automatically, with the customer told their claim is being checked.
Which order systems does it connect to?
Any system with an API, through a swappable adapter: order status, items and packing, rider location, payments and wallet. Our team builds the adapter during the first week.
Does it work on WhatsApp?
Yes. Connect your WhatsApp Business number through Meta's Cloud API, Gupshup or Twilio, and the same agent handles WhatsApp alongside app chat, calls and email.
How long until it is live?
About two weeks: a week to connect the order system, channels and policies and rehearse against past conversations, then copilot mode on live tickets. Intents such as where-is-my-order and delay credits usually go autonomous first.