A woman checking a parcel delivery on her phone, in Arcsen colours

Where is my order? AI agents for retailers, brands and distributors

Much of the customer contact in Gulf and Egyptian retail comes down to a handful of questions. Where is my order? Can I change the delivery slot? How do I return this? Is it in stock at my store? Can I book an appointment? They arrive on WhatsApp, often late in the evening, in Arabic and English, and peak around Ramadan, Eid and White Friday.

Distributors see the B2B version: a grocer sends the sales rep a reorder on WhatsApp, and someone keys it into SAP. AI agents on Salesforce can take much of this work if three conditions hold: the agent reads live data, knows who it is talking to, and hands money, complaints and consent to people.

Four agents worth building first

Salesforce now sells several of these as packaged agents that come with the skills, actions and data models for the job and can be tailored to how each company works. The work is in the tailoring.

  • Orders, delivery and returns, on web chat and WhatsApp. Casey, the help agent Salesforce named in September 2026, has pre-built support for FAQs, returns, account management and human escalation across voice, SMS, WhatsApp and web chat. Service agents only support inbound conversations, so dispatch and delivery notices go out from flows as approved WhatsApp templates, and the agent handles the replies.
  • Products and appointments, for stores and boutiques. Commerce Cloud’s Shopper Agent, generally available since June 2026, carries the conversation from discovery through checkout and into service. A jeweller or a fashion house often wants the conversation to end in a boutique appointment, so ground the agent in the catalogue in both languages and give it an action on the store diary.
  • Clienteling assist, for store staff. Salesforce describes Retail Cloud giving associates AI agents, unified data and real-time context, including product suggestions and messaging replies grounded in your knowledge base. The agent prepares a client brief before a visit and drafts follow-ups that the associate edits and sends.
  • B2B reorders, for distributors. The June release also made the Buyer Agent generally available: it meets B2B buyers in WhatsApp and SMS, and a buyer who asks to repeat an earlier order gets product images, contract pricing and a completed order. For field teams, Consumer Goods Cloud lists a Visit Assistant for each rep, a Sales Planning Agent that flags deviations on retail accounts, and Distributor Management for purchase orders, goods receipts and stock audits.
Agent Channel Data and systems it needs Actions it may take Where a person takes over
Orders, delivery and returns Web chat, WhatsApp, voice Order management or the e-commerce platform, carrier tracking, returns policy by country Order status for a verified customer, delivery slot changes before dispatch, returns within policy, cases Refunds and goodwill above set limits, lost high-value items, payment disputes, complaints
Products and appointments Website, WhatsApp Catalogue in Arabic and English, stock by store, boutique diary Product and care answers, store stock checks, booking and moving appointments Prices outside the catalogue, bespoke orders, authenticity and valuation
Clienteling assist Retail Cloud on the associate’s device Client profile with purchases, preferences and consent, stock across stores Client briefs, suggested pieces, draft messages The associate sends every message, and marketing goes only to clients who agreed
B2B reorders WhatsApp, SMS, rep visits B2B catalogue, contract prices, order history, SAP for price, credit and stock Reorders from approved numbers, price and stock checks, order placement and status Credit holds, orders above a set value, price and promotion exceptions, disputes

What the agents need before they go live

One customer record. Our guide to one customer profile for retail covers the matching and consent work in Data 360, formerly Data Cloud, that agents depend on. Grounding answers in your own content through an Agentforce Data Library needs Data 360 activated first, and the library takes Knowledge articles, uploaded files, web search and custom retrievers.

Identity before personal data. Salesforce’s standard verification subagent sends a one-time passcode by email, valid for five minutes, when an unverified customer reaches a subagent or action that needs it. Many customers in the region shop with a mobile number and no email, so decide how WhatsApp customers are verified before the agent shows an order or an address, and which numbers may order for each B2B outlet.

Live data. Orders, stock, prices and credit stay in order management and SAP, read through actions, usually over MuleSoft. The agent quotes nothing it has not read from a system. Einstein Trust Layer data masking is disabled for service agents, so give each agent only the fields its job needs.

Knowledge with owners. Returns policy, delivery cut-offs, Ramadan store hours and care guides each need an owner, Arabic and English versions, and an end date for seasonal content, so the agent never quotes last year’s Eid cut-off.

Guardrails, and where a person takes over

Limits should be rules the agent cannot talk its way around. Agent Script, the language for building agents in Agentforce Builder, mixes conditional logic with prompts, so a refund limit can be a fixed rule inside a natural conversation. Since April 2026 Salesforce calls topics subagents, and each subagent should state what it may do and when it hands over.

Rules that stay with people in retail and distribution.

  • Refunds, credit notes and goodwill above the limits finance has set, and every exception to the returns policy.
  • Complaints, as cases with the conversation attached and a named owner.
  • Marketing consent: the agent records an opt-out at once and never opts anyone in. Wording and purposes belong to marketing and legal.
  • Prices, discounts and trade terms not already in the system, credit holds, and B2B orders above a set value.

Every handover carries the conversation and what the agent gathered.

Testing and measuring, before launch and before each peak

The Agentforce Testing Center generates test cases from an agent’s subagents and actions, builds question-and-answer tests from knowledge, or takes your own scenarios, and scores answers on measures such as completeness and instruction adherence. Salesforce advises running it only in a sandbox, because tests can change CRM data.

Arabic needs its own plan. Salesforce’s language support page lists Arabic as Beta and warns that languages in beta may have some inconsistencies when retrieving information. Test the way customers write: Modern Standard Arabic, Gulf and Egyptian dialect (وين طلبي؟ and الأوردر بتاعي فين؟ both ask where an order is), and Arabic typed in Latin letters. Native speakers review the answers, an Arabic-speaking person stays one step away, and our notes on what to test in Arabic go further.

After launch, measure by language and channel: containment, handovers where the person had to ask again, complaints and refunds that follow an agent conversation, and repeat contacts about the same order.

Before Ramadan, Eid and White Friday, rehearse the season in the sandbox: cut-offs, store hours, sold-out offers and gift returns, in both languages. Staff the human queue for the busiest hours and freeze rule changes during the peak. Our SitecoreAI guide to retail peak seasons covers the website side.

Our retail and distribution work

Arcsen is a Salesforce Summit partner with 170+ consultants and 400+ certifications and Arabic-first delivery from Dubai, Riyadh, Jeddah, Doha and Cairo. The work below is the kind of foundation these agents stand on.

  • Azza Fahmy: we are implementing Sales Cloud, Marketing Cloud and Data 360 for the Egyptian jewellery house, with one customer profile behind every sale and campaign.
  • Coca-Cola KSA: we are implementing Sales Cloud for Coca-Cola in Saudi Arabia, integrated with SAP so sales and the ERP work from one set of data.
  • Brother: the UAE distributor channel runs self-service on a partner portal, with omni-channel customer service on one desk.

Where to start

  1. Count contact reasons across WhatsApp, chat and calls, and pick the agent that covers the most routine questions.
  2. Write the money rules: refund, goodwill and order-value limits, signed off by finance.
  3. Decide verification for WhatsApp customers and B2B buyers.
  4. Connect the first agent’s data, and name an owner for each knowledge article.
  5. Test in both languages in a sandbox, then launch on one channel in one market before the next peak.
Planning your first retail agent? The free Agentforce Readiness Assessment takes ten working days and gives you three use cases and a costed pilot plan. See our retail and distribution work and how we deliver Agentforce in Arabic and English, or talk to our retail and distribution team.

Frequently asked questions

Can an Agentforce agent answer retail customers on WhatsApp in Arabic?

Yes, with a Beta label for Arabic. Test in both languages, have native speakers review the answers and keep a quick route to an Arabic-speaking person.

Can an agent issue refunds?

Within limits you set as fixed rules. It can start returns inside policy, and anything above the limit, and every complaint, goes to a person.

Can retail customers reorder from a distributor on WhatsApp?

Yes. Commerce Cloud’s Buyer Agent, generally available since June 2026, takes B2B orders in WhatsApp and SMS with contract pricing. For distributors on SAP, the order still passes SAP’s price, credit and stock checks.

Do we need Data 360 for retail agents?

For an Agentforce Data Library, which grounds answers in your own content, yes: Data 360 must be activated first.

How do we prepare an agent for Ramadan and White Friday?

Rehearse the season in a sandbox with the Agentforce Testing Center, load seasonal knowledge with end dates, staff the human queue and freeze rule changes during the peak.

Sources