WhatsApp Product Advice Agents: Answers Before Checkout

A customer sends a photo and asks, “Will this cable fit my monitor?” Another asks whether the blue jacket is available in medium and what it costs in Switzerland. Both look like simple product questions. Only one may be answerable from a normal catalogue without extra evidence.

A useful WhatsApp product advice agent treats the catalogue as its boundary. It clarifies the requirement, identifies the exact variant and retrieves current data before replying. Retailers considering a custom implementation can use AI agent development to connect the messaging channel, catalogue and human handoff around one controlled workflow.

Reliable product advice comes from documented product facts and a stated customer need. This article covers inbound ecommerce questions and the handoff to a product or checkout page. Existing order enquiries belong to a separate workflow. Broader channel choices are covered in the WhatsApp AI agent overview.

---

Separate the Business App from the integration

The WhatsApp Business App supports staff who handle conversations manually. A custom product advice workflow needs the WhatsApp Business Platform so software can receive messages, call the retailer’s catalogue API and route a response or escalation. The App may remain part of day to day work, but it is not the integration layer assumed here.

Ask for a demonstration. Which Platform component receives the message? Which catalogue account and permissions does it use? Does it query live fields or a copied product index? How does a staff member take over, and which source fields can they see?

Define the scope. An agent asked to “answer anything about our products” has no clear boundary when the catalogue lacks evidence; begin with questions the stored fields can actually answer:

• Which listed variant matches a requested size, colour or model?
• What price does the shop currently return for the relevant market context?
• Is this exact variant available for sale at the time of the lookup?
• What material, dimension or feature does the catalogue state?

Missing facts stop the answer. So do contradictory records and questions that require an unsupported compatibility judgement, safety assurance or warranty interpretation. This bounded role complements the wider ecommerce opportunities discussed in AI agents for ecommerce.

---

Turn the shopper’s message into a catalogue query

Staff need to inspect the sequence. Each step should turn the shopper’s wording into a catalogue query or a decision they can test.

1. Classify the question. Identify whether the shopper names a product, supplies a requirement, asks for a price or availability check, or requests an expert conclusion.
2. Ask one focused clarification. If size, colour, model, budget or market context is missing, ask for the field that determines the next lookup.
3. Resolve the exact variant. Match option values to a variant identifier. A parent product title is insufficient when different variants carry different prices or availability.
4. Retrieve controlled fields. Read the selected options, documented specifications, contextual price, sale availability and update time from the retailer’s approved source.
5. Apply recommendation rules. Rank only against requirements the shopper stated and fields the catalogue supports. Record the reason for the suggestion.
6. Answer with source context. Name the exact variant, currency, relevant specification, availability wording and retrieval time. Include the appropriate product or checkout link.
7. Handoff uncertainty. Route missing facts, conflicting fields, regulated products, negotiated prices and expert compatibility questions to qualified staff.

In an illustrative example, a shopper wants a rain jacket in medium under CHF 180. The agent filters those declared constraints and retrieves the exact blue medium variant. For this fictional example, assume the catalogue lists a waterproof rating and a Swiss contextual price of CHF 169. Mountaineering suitability needs separate evidence. The agent should make that claim only when controlled product documentation supports the use.

An illustrative reply using the fictional product, prices and lookup time would read: “I suggested the Alpine Shell, blue, medium because you asked for a waterproof jacket under CHF 180. The catalogue lists a CHF 169 price for Switzerland and marked this variant available for sale when checked at 14:12 CEST. Checkout confirms the final price and availability.”

The buyer has a next step. A reviewer can trace each claim.

---

Read variant, price and availability fields carefully

Shopify is one concrete platform example. Its ProductVariant GraphQL documentation describes a specific combination of product options and exposes fields such as selectedOptions, price, contextualPricing, availableForSale, inventory data and updatedAt. The exact schema and permissions differ on other commerce systems.

Resolve the variant first. A shoe’s parent record may cover several sizes, yet the price or sale availability can differ by variant. If the customer says “the black one,” the workflow may still need size or model before it has a unique answer.

Price needs context too. Shopify documents contextual pricing that can reflect an active market. Include the currency and market context returned by the integration, then link to checkout, where taxes, delivery options, discounts and availability are checked again after the chat lookup.

availableForSale requires particularly careful wording. It indicates whether the variant is purchasable under the configured inventory policy. A policy can permit purchase when physical inventory is exhausted, so the value may describe a backorderable product. It is neither a count of units on a shelf nor a reservation.

An illustrative response says “available for sale when checked at 14:12 CEST.” Use “in stock” only when the retailer has defined that phrase against suitable inventory fields and location rules. “Reserved for you” needs a successful reservation write. A lookup alone cannot support it.

Catalogue data has limits beyond inventory. A description or metafield reports what the merchant stored. It does not establish medical suitability, destination specific legal compliance or compatibility with an unknown device. The agent can say “the catalogue lists” and point to the source. Missing evidence triggers a clarification or handoff.

---

Constrain recommendations and make handoff useful

Explain the selection rule. When a customer names a need, the agent should show how the suggested candidates match it, using the documented product attributes that drove the choice. Broad claims such as “best choice” create risk because they hide the comparison basis.

Build an explicit safe answer list. Straightforward material, dimensions, colour, price and variant options can be eligible when the source field is present and current enough. A separate escalation list should include:

• compatibility with a customer owned item when no approved matrix identifies the exact model;
• health, safety, installation or regulated use questions;
• conflicting or missing product fields;
• warranty interpretation and commercial exceptions;
• negotiated pricing, reservation or any action that changes stock or an order.

Give staff the customer’s stated need, variant candidates, catalogue fields already checked and the missing evidence, so they can continue without asking the shopper to repeat the conversation. The gaps also guide catalogue updates. If employees repeatedly answer the same compatibility question from manufacturer documentation, the retailer can decide whether to add a governed compatibility matrix.

Count corrections. Review whether the chosen variant was accurate, whether cited fields supported the reply and whether staff changed the recommendation. A locally measured accuracy threshold is more useful than a generic automation claim because catalogue quality and product risk differ sharply among retailers.

Integration with a CRM or product information system may help staff see past context and source data. AI agent CRM and ERP integration explains the wider system design. For this workflow, keep permissions limited to the reads and handoff writes the product question actually needs.

---

Keep service replies separate from promotion

The current WhatsApp Business Messaging Policy permits free form Platform replies within 24 hours of the user’s last message. Outside that customer service window, a business needs an approved template to initiate the next message. The policy also requires automated experiences to provide a prompt and clear escalation path.

An inbound question about the price of a named jacket can fit a service interaction. A later message promoting matching trousers has a different purpose. Meta’s current Business Platform pricing guidance includes related product suggestions among marketing examples, while it describes incoming enquiries handled through an agent or conversational AI as service messages. The workflow should classify what the message does at send time and apply the appropriate consent, template and category treatment.

Commerce capability is also market and policy dependent. The business remains responsible for the transaction and fulfilment. Design the stable handoff around a product or checkout link rather than promising that every shopper can pay inside WhatsApp. Checkout can validate final price, availability, tax, delivery and payment options in the customer’s actual context.

Current pricing is per delivered message and varies by category and recipient market. Meta currently lists service messages as free of its charge, but that does not remove provider, model, integration, monitoring or staff costs. Use the live pricing source during budgeting and recheck rules before launch because platform policies and rates can change.

If an external AI provider participates, assess the current Business Solution Terms and data arrangement for the exact phone numbers, markets and use. The project should document what conversation and catalogue data reaches each provider, how long it remains there and whether any model improvement use is proposed. Product advice for one retailer should stay within that retailer’s controlled business workflow.

---

Measure capacity without assuming more sales

Suppose an illustrative retailer receives 15 pre purchase questions per weekday. If staff spend an assumed 4 minutes locating the exact variant and replying to each question, the total is 60 minutes. If the agent answers an assumed 9 questions from complete catalogue records and routes 6 uncertain cases to staff, the remaining manual lookup and reply time is 24 minutes. The illustrative difference is 36 minutes of gross recoverable capacity that day.

That calculation is neither a conversion benchmark nor a wage saving. Review time, exception handling, corrections, catalogue maintenance, monitoring and integration operations consume part of the capacity. A pilot should measure question volume, handling time, answer support rate, correction rate, handoff time and completed checkout outcomes separately.

Sales attribution needs a comparison. A shopper may have purchased anyway, and a click on a checkout link does not prove incremental contribution. The capacity case is measurable from the workflow itself and can justify a pilot where repetitive catalogue lookup occupies staff time.

Orange ITS uses illustrative planning ranges of CHF 5,000 to 18,000 for a narrow pilot with one integration and limited hardening, and CHF 18,000 to 60,000 for a single production agent. Scope depends on catalogue complexity, markets, languages, monitoring, evaluation and staff routing. A retailer still needs a specific quote. Ongoing provider, model and support costs need separate estimates.

The comparison should use net recovered capacity after overhead and the retailer’s own value for that time. Where staff mainly handle uncertain product questions, the automation share may be too small. Where employees repeatedly search complete catalogue fields for price, variant and availability answers, the case is stronger.

---

Set pilot acceptance criteria around answer quality

A pilot needs a representative test set before it speaks to customers. Include simple variant questions and difficult cases: ambiguous photos, missing sizes, two similar model names, backorderable products, market specific prices, outdated specifications and compatibility requests with no approved matrix.

Define acceptance criteria that a reviewer can score:

• the selected variant matches the shopper’s stated attributes;
• every factual answer maps to an approved catalogue field or controlled document;
• price includes the correct currency and available market context;
• sale availability is time bounded and never described as a reservation;
• unsupported compatibility, safety and commercial decisions reach the right staff queue;
• the message uses the correct WhatsApp window, category and template treatment.

Repeat the tests after changes. A revised catalogue schema, inventory policy, market configuration or prompt can change the answer, so keep sampling resolved conversations and inspecting staff corrections in production. A high automatic answer count has little value if the system chooses the wrong variant or overstates availability.

The retailer is ready when product records are complete enough to answer recurring questions, exact variants can be resolved, and staff own the remaining judgements. The successful agent leaves the customer with a cited fact and a viable checkout path while preserving human authority where the catalogue runs out.