An AI agent for e-commerce store desks answers “where is my order,” starts a return from the published policy, and chases an abandoned cart on your script. In San Diego that usually reads Shopify or ShipStation and writes to Gorgias or Klaviyo. A person still refunds money and still decides a warranty exception.
Nidrosoft is the San Diego studio of Cyriac Ndo Zeh, a design engineer who implements AI for small and mid-size businesses: audits, custom builds, phone and receptionist agents, training, and systems the team can own.
Written by Cyriac Ndo Zeh, Founder of Nidrosoft — 12+ years building software and AI products for San Diego small businesses. Also on LinkedIn.
Key takeaways
- The first three jobs are order-status (WISMO), returns intake, and abandoned-cart recovery on your words.
- Shopify, WooCommerce, BigCommerce, Gorgias, Klaviyo, and ShipStation stay the record. The agent reads and writes fields those systems already accept.
- A two-person support pair that still types every “where is my order” usually gets 12-20 hours back in a week once the ticket tags are honest.
- Same-day chatbot widgets can greet tonight. Custom work starts when
FulfillmentStatusandReturnReasonmust match the warehouse. - This page names no store clients. The transcript below is an illustration, not an order file.
- Book a free 15-minute AI readiness call if the leftover is last week’s open WISMO pile.
Contents
- WISMO tickets that eat the morning
- Three jobs for a store desk
- Sample order-status transcript
- Hours a two-person support pair usually gets back
- Shopify, Gorgias, Klaviyo, and the 3PL
- AI agent for e-commerce store warehouses across San Diego
- What finance still refunds
- Build the policy the page already states
- Frequently asked questions
- Start with last week’s open tickets
WISMO tickets that eat the morning
A San Diego e-commerce morning still starts in Gorgias or a shared inbox. Overnight USPS scans stalled in Los Angeles. A guest checkout cannot find the order. A wholesale buyer in Kearny Mesa wants a carton count. The founder is at the 3PL in Otay. By 10 a.m. the two-person support pair has answered the same tracking question 40 times and has not touched the return that will become a chargeback.
The leak is structured lookups and honest policy, not a brand-voice bot on the homepage. “Where is my order” (WISMO) tickets that a person retypes from Shopify. Returns that stall because nobody asked for photos. Abandoned carts that get a cute Klaviyo flow and then a human who also emails the same person. Local pickup orders that sit on a shelf in Mira Mesa because the buyer called the wrong number.
Start with last week’s ticket tags and the fulfillment exception list, not with a brand voice. Write each inbound reason as an intent. Order status. Address change. Cancel. Return or exchange. Warranty. Wholesale or B2B. Local pickup. Spam. Each intent needs a required field list, a store action, and a transfer target. Chargebacks and “I will sue” language transfer. The agent does not invent a tracking number.
AI implementation for small business starts with a workflow the team already owns. Adjacent floor work is AI agent for retail store. If the founders want a room before a build, AI workshop for small businesses has been run in La Mesa.
Fit is companies roughly 30 to 250 people, plus a smaller shop if one store, one inbox, and one fulfillment path are clear. A Shopify brand with two support seats and ShipStation is a clear workflow if you name which pile is first. A holding company that wants “AI for all our stores” with no inheritor is not clear yet.
Three jobs for a store desk
Use case 1: order status that reads the live record. The agent asks for OrderNumber or email plus ZIP, then reads FulfillmentStatus, Carrier, TrackingNumber, and DeliveredFlag from Shopify or the 3PL feed you already trust. If the order is unfulfilled, it repeats the published ship-by window. If the scan is stale, it offers the exception path you already use (wait 24 hours, then a claim). It does not invent a delivery day. It does not say “it is on the truck” unless the scan says that.
Use case 2: returns intake from the published policy. Collect OrderNumber, SKU or product name, ReturnReason from a closed list, PhotoAttached if you already require it, and RefundOrExchange. The agent repeats the window and the restocking rule that are already on the returns page. It creates the return in Loop, Returnly, or the Shopify return you already run, or it writes a Gorgias ticket with those fields if you have no return app. It does not grant an exception. It does not refund shipping unless the policy already says so. Used or missing items transfer.
Use case 3: abandoned-cart recovery on your script. Treat each cart as a row: CartId, EmailOrSMS, LastAttemptAt, CartStatus (open, attempted, recovered, opted-out, do-not-contact). Band A, one hour after abandon if you already do that. Band B, 24 hours. Band C, day three. Use the copy the founder already approved. Do not invent a discount code. If a code exists, a person still creates it in Shopify. Honor STOP and the list-unsubscribe you already show. Wholesale carts and custom-quote carts stay off this flow.
Those three jobs share a help address and sometimes a phone number. They do not share a tone. One agent that uses cart-copy on a missing-package ticket will sound like a hustle when the buyer is already angry.
Sample order-status transcript
The exchange below is an illustration. It is not a Shopify order and it is not your carrier mix.
Agent: This is the support line for Shoreline Supply. I am an automated assistant. I can check an order, start a return, or take a pickup question. I cannot issue a refund on this call. Do you have an order number?
Caller: Yes, 10482. It was supposed to be here yesterday. I am in North Park.
Agent: I will look up 10482. Please confirm the email on the order or the shipping ZIP.
Caller: 92104. Email is sam.ortega at proton.me.
Agent: Order 10482 matches 92104. Status is fulfilled. Carrier is USPS. Tracking ends in 9031. The last scan is “arrived at regional facility, Los Angeles” from yesterday 4:10 p.m. I do not have a delivered scan.
Caller: So you lost it. I need it for a shoot Friday.
Agent: I cannot say the package is lost from one stale scan. Our published path is to wait one more business day, then open a claim if there is still no scan. I can also check whether a local pickup replacement is on the shelf in Mira Mesa if that product is in stock. Do you want the claim timer, a stock check, or a person?
Caller: Stock check. Can you overnight it?
Agent: I can note a replacement request. Overnight is a paid exception a person approves. I will not charge a card on this call. The Mira Mesa counter is open 10 a.m. to 4 p.m. weekdays if the SKU is on the shelf. I will flag Friday shoot on the ticket.
Caller: Fine. Also I want to return the extra straps from the last order.
Agent: I can start a return if that order is inside the published 30-day window. What is that order number, and are the straps unused?
Caller: 10311. Unused, tags on.
Agent: 10311 is inside the window. I will create a return for the strap SKU, unused, refund to the original payment method as the policy states. Shipping on returns is the buyer’s cost unless we marked the item damaged. I will email the label if your store issues labels, or the returns URL if you self-ship. Which product is damaged, if any?
Caller: Nothing damaged. Just extra straps.
Agent: Return started for 10311 straps. Ticket opened for 10482 stale scan and Friday flag. A person reviews overnight. Reply STOP to leave SMS.
Nine exchanges. The agent read a scan, refused a lost-package verdict, refused an overnight charge, and started a return from written policy. That refusal is the product.
Hours a two-person support pair usually gets back
Ground the hours in a desk you can watch. WISMO is repetitive. If a pair handles 60 status tickets a day at 6 to 10 minutes each, that is 6-10 hours before anyone touches a warranty. An agent that authenticates and pastes the live scan usually cuts those tickets to the exceptions (no scan, wrong address, guest checkout mismatch). Shops that measure this honestly often see 6-10 hours back on WISMO alone.
Returns are the second pile. Photo chasing and “is this inside 30 days” math is 8 to 15 minutes a ticket. Twenty returns is three to five hours. Structured intake that creates the return object usually returns 3-5 hours a week, and still leaves exceptions on a person.
Cart recovery is lighter on the clock and heavier on the mess. If a human also emails carts that Klaviyo already touched, you are paying twice. Putting Band A through C on one queue and suppressing when an order lands usually returns 2-4 hours a week of “did we already send that?” plus whatever revenue you already attribute in Klaviyo. Do not invent an attribution study.
Wholesale and local pickup exceptions are bursty. Structured PickupWindow and PONumber usually save 1-2 hours.
Add those piles and a two-person San Diego support pair typically lands in a 12-20 hours per week range. A founder doing support at night may see 8-12. A ten-seat team may see 20-plus on WISMO and still need people on chargebacks. Do not budget a layoff from a blog post.
Same-day chat widgets (Gorgias macros, Shopify Inbox, many “AI support” apps) can greet tonight. Praise that. Buy one if the fire is an empty chat bubble. Custom work starts when the 3PL scan and the return app must agree and a fail-closed refund rule is written.
Shopify, Gorgias, Klaviyo, and the 3PL
Name the software you already pay for.
| System | What stores use it for | What an agent may write | What stays human |
|---|---|---|---|
| Shopify | Orders, inventory, pickup | Ticket fields, return if enabled, order note | Refunds, discount codes, inventory adjustments |
| WooCommerce / BigCommerce | Same shape | Same, via the API you already have | Same |
| Gorgias / Zendesk / Freshdesk | Tickets | Tag, macro, required fields | Macros that refund |
| Klaviyo | Flows, carts | Suppress, log attempt | New percentage-off codes |
| ShipStation / Shippo | Labels | Read status | Void a label |
| Loop / Returnly / Happy Returns | Returns | Create return if policy matches | Exceptions |
| Amazon / Etsy / wholesale portals | Other channels | Only if that channel is in scope | Channel policy fights |
| QuickBooks / NetSuite | Books | Nothing that posts | Credits |
The 3PL (a Kearny Mesa warehouse, an Otay dock, or a national 3PL with a San Diego overflow) is a system. If their scan is the truth, the agent reads their feed, not a hope. If they batch-scan at 4 p.m., say so in the published ship window. Do not let the agent promise same-day when the dock already closed.
Do not replace Shopify in a six-week project unless that conversion is the actual job. AI implementation San Diego is the county door. The habit page is AI implementation for small business.
TokenTra is a public Nidrosoft system that helps teams cut 10-30 percent of wasted AI spend. If you later run a lot of model calls on support, that invoice is a leftover. It does not refund a customer. NanoBrief can turn a messy internal brief in under two minutes versus three-plus hours. It does not write a product description you then publish without a person.
AI agent for e-commerce store warehouses across San Diego
San Diego e-commerce is often a warehouse with a storefront, not a coastal boutique. Kearny Mesa and Miramar hold a lot of pick-pack. Otay and San Ysidro hold cross-border cartons and delay stories the agent must not invent. Carlsbad and Oceanside brands add tourism gift volume in summer. Downtown and North Park brands add local pickup that collides with the warehouse clock.
Local pickup fields that pay: PickupLocation, ReadyFlag, IDRequired, HoldUntil. The agent may say the order is ready if Shopify says ready. It may not extend a hold past the published window without a person. Tourism buyers who want weekend pickup need the Saturday hours you already post.
Spanish-first buyers show up on South Bay pickup and on many wholesale counters. Offer Spanish on status and returns. Complete the thread in one language. Do not live-translate a warranty that limits coverage.
Cross-border addresses (Tijuana ship-to, broker, or “hold at Otay”) transfer. A model that guesses a broker fee will create a finance mess.
Identify as AI. California recording and TCPA rules still apply if you call or text. Honor STOP. Ask counsel. This page is not legal advice and it is not a card-network chargeback manual.
If you also run a retail counter, keep the floor script on the retail article. This page stays on the order object.
What finance still refunds
A person still refunds. The model does not push a credit to Shopify Payments or Stripe. The model does not add a discount code. The model does not tell a buyer to do a chargeback.
Human gates:
- Money: refunds, partials, shipping credits, wholesale terms. A person posts.
- Exceptions: late returns, used items, “it smells,” warranty beyond the page. A person.
- Fraud and promo abuse. A person.
- Wholesale pricing. A person.
- Channel disputes (Amazon A-to-z). A person.
Inventory writes (-1 on a SKU) stay with the warehouse or a person. A model that decrements on a chat will desync the 3PL.
Nexuvo-shaped inventory work is a dealer product, not a Shopify app. Protectron-shaped European paperwork is for companies that sell into Europe. If you do, name that leftover. If you do not, skip it.
Build the policy the page already states
The method is Embed, Diagnose, Scope, Build, Own. Embed means a morning in Gorgias and a walk of the Mira Mesa or Otay shelf, not a brand workshop. Diagnose means last week’s tags and the exception list. Scope means one pile: WISMO, returns, or carts. Build means the agent plus the read/write. Own means your support lead can change the 30-day sentence when you change the page.
A free 15-minute AI readiness call is the door. A scoped audit often lands in the low thousands. A first custom agent or integration often lands in a mid four- to five-figure project band depending on Shopify apps and the 3PL feed. A $50-per-month chatbot is a different leftover.
Cyriac has 12+ years building products, 125 shipped, 16 public, and time at Anthropic, Microsoft (Bing and Edge), Intuit, and Gap Inc. Live systems you can open include Omnira Dental, Octana, Eviction Wizard, Guidera, TokenTra, Protectron, and Nerlude. We will not invent a San Diego brand that uses them.
Bring last week’s ticket export, the returns URL, and the 3PL scan delay you already know. If the returns page and the agent disagree, the page wins until a person edits both.
Frequently asked questions
Can the agent issue a refund?
No. It can start a return that matches the published policy. Finance or a named support lead still pushes the credit. Chargebacks stay on a person.
What if Shopify and the 3PL disagree?
The agent should say the two records do not match and open an exception ticket. It should not pick a winner. Your warehouse lead decides.
Will this replace a support hire?
Plan for hours back on WISMO and clean returns, not for an empty seat. People still do warranties, wholesale, and angry threads. A shop that cannot hire a third seat may use the hours on the exception pile.
Do we have to leave Gorgias?
No. Use the inbox you already pay for. Custom work is the order read and the fail-closed refund rule.
Can it write product copy or ads?
That is a different leftover. This page is the order desk. If you want training on drafts, the workshop is the door. Do not let a support agent publish copy.
Start with last week’s open tickets
Export Gorgias tags before you shop a widget. Count WISMO, returns, and carts that a human also touched. That count is the scope. Book a free 15-minute AI readiness call and bring the store and 3PL names you already pay for. Hello sits at cyriac@nidrosoft.com if you want a written note first.