How to use AI shopping-assistant conversations to forecast seasonal CX demand

An ecommerce CX team can use AI shopping-assistant conversations as a leading indicator of pre-purchase demand—not as a replacement for its order-service forecast. Classify conversations by intent, measure volume and escalation by time period and product, compare those trends with traffic and sales, and staff a separate sales-assistance queue. Keep order status, delivery, returns, and account issues in the service workflow with their own forecast, SLAs, and capacity plan.

Separate pre-purchase demand from order-service demand

The first step is to create two operational streams: questions that help someone decide whether to buy, and issues raised after an order exists. Forecast them independently because they have different owners, urgency, knowledge sources, and success measures.

Pre-purchase conversations include:

Product comparison and recommendation questions
Fit, size, compatibility, ingredients, materials, or use-case questions
Availability, delivery options, shipping destinations, and expected arrival questions asked before checkout
Price, promotion, warranty, and return-policy questions asked while evaluating a product
Requests for a store or location

Order-service tickets include:

“Where is my order?” and shipment-status requests
Address changes, cancellations, payment problems, and account issues
Returns, exchanges, refunds, damaged deliveries, and warranty claims for an existing order

Do not infer the category from the word “shipping” alone. “Do you ship this product to Alaska?” is usually a buying question; “my package has not arrived” is an order-service issue. Use the conversation stage and the presence of an order identifier as routing signals.

A clean data model should preserve at least these fields for every conversation:

Field	Why it matters for staffing
Intent	Separates product guidance from post-order service
Timestamp and timezone	Reveals hourly, daily, and regional demand
Product, category, or collection	Shows where product knowledge or human expertise is needed
Entry page and channel	Connects questions to campaigns, landing pages, and traffic
Outcome	Distinguishes answered, abandoned, converted, and escalated conversations
Handoff reason	Shows which questions still need a person
Order state	Prevents pre-purchase demand from entering the service forecast

This separation also protects the meaning of your support metrics. A question answered before checkout is not the same workload as a delivery exception, even if both arrive through a chat interface.

Use conversations as a leading signal for seasonal peaks

Conversation volume can reveal buying friction before it appears in order-service tickets. Track pre-purchase questions against site sessions, product views, add-to-cart activity, orders, promotions, and inventory so the CX team can see whether a rise reflects more traffic, more uncertainty, or a product-specific problem.

An on-site assistant is useful here because it captures the shopper’s own wording at the point of comparison. Anagram describes its Site Agent as helping shoppers with conversational product support while they compare options and decide what to buy, and its site positions shopper-question insights as a way to show what customers care about and where they hesitate. See Anagram’s Site Agent overview and its guide to answering shoppers’ product questions.

Build the seasonal signal in layers:

Establish a baseline. Record weekly and intraday pre-purchase conversation volume, unique shoppers, escalation rate, and answer coverage during ordinary trading periods.
Mark commercial events. Label promotions, product launches, media pushes, stock changes, holidays, and major changes to site merchandising. A peak caused by a discount should not be treated as ordinary seasonality.
Group questions by intent. “Which size?”, “which model?”, and “will it work for me?” may require different staffing from general product discovery, even when they concern the same category.
Compare like with like. Compare the same category, channel, weekday, and campaign type where possible. A sitewide average can hide a spike on one high-consideration product.
Watch the handoff boundary. A growing share of conversations escalated to a human is a capacity warning even if total conversation volume is stable.

The result is not a promise that every conversation becomes a ticket. It is a demand indicator: how many shoppers need an answer, which answers they need, and what proportion still requires human intervention.

Turn the signal into a staffing forecast

Forecast pre-purchase staffing from the conversations that need people, not from total assistant volume. Start with the expected number of conversations by interval and intent, then apply the historical escalation rate and the average human handling time for each intent.

A simple planning model is:

Human pre-purchase workload = forecast conversations × expected escalation rate × average handling time

Run the calculation separately for each meaningful segment, such as sizing, technical compatibility, or product comparison. If one question type takes substantially longer or requires a specialist, it should not be represented by the same handling-time assumption as a simple policy question.

Use three planning views rather than one headline number:

Base case: expected traffic, conversation rate, escalation rate, and handling time
Peak case: higher traffic or a campaign-driven increase in a particular intent
Content-improvement case: the expected effect of fixing a recurring answer gap before the peak

Translate workload into shifts using your team’s actual productive capacity, concurrency, breaks, meetings, training, and service-level target. Do not treat scheduled hours as fully available answering time. Review the forecast at a longer seasonal horizon for hiring or contractor decisions, then refresh it at shorter intervals as campaigns and traffic become clearer.

For order-service tickets, keep the existing service forecast separate. Forecast order volume, shipment events, returns, and known fulfillment risks through the system that owns those records. The pre-purchase plan can share people with the service team, but it should retain its own queue, skill labels, and reporting so a sales-assistance spike does not make post-order performance look worse—or conceal it.

Microsoft’s documentation describes the same useful forecasting principle for service operations: forecast volume and representative demand at daily and shorter intraday intervals, while accounting for historical patterns and seasonality. Its case and conversation volume forecasting guidance is a useful reference for the mechanics, even though your pre-purchase stream may need different inputs and service goals.

Route conversations without creating a second ticket swamp

The routing rule should be based on intent and order state. A pre-purchase conversation should stay in the shopping experience when the assistant can answer from approved product and policy information; it should go to a sales-assistance queue when a person needs to compare options, clarify a high-value use case, or resolve uncertainty that blocks purchase.

Use a separate order-service path when the shopper provides an order number or describes an existing-order problem. That path should connect to the systems and policies used for fulfillment, returns, payments, and account support. Do not hand a shopper from a product recommendation flow into an order queue simply because the question contains a service term.

A practical routing table looks like this:

Detected situation	Queue	Useful outcome
Product fit, comparison, or recommendation	Pre-purchase assistance	Product answer, recommendation, or qualified handoff
Policy question before checkout	Pre-purchase assistance or self-service	Clear policy answer and next step
No order exists, but delivery suitability affects the decision	Pre-purchase assistance	Shipping or availability answer from current data
Existing order, delivery, return, refund, or payment issue	Order service	Case resolution and correct SLA
Ambiguous or mixed conversation	Triage	Confirm whether the shopper is deciding or seeking service

Keep the queues distinct even if the same agent handles both. Give the agent the conversation intent, products discussed, unanswered question, and reason for escalation. That context avoids making a shopper repeat a comparison they already completed.

Measure whether the forecast is getting better

A useful dashboard should show both demand and business effect. Conversation counts alone cannot tell you whether staffing improved the experience or whether the assistant simply attracted more questions.

Track these measures separately for pre-purchase and order service:

Conversations by intent, category, channel, and time interval
Human escalation rate and average handling time by intent
First response time and abandonment rate for the human pre-purchase queue
Answer coverage and unanswered-question volume
Product-page, add-to-cart, and purchase activity after assisted conversations
Order-service volume, backlog, response time, resolution time, and SLA performance
Forecast versus actual volume and workload for each queue

Anagram’s published guidance lists questions asked, answer rate, unanswered questions, support deflection, and follow-up clicks as useful measures for an on-site assistant. Treat those as inputs to the operating review, not as proof that staffing demand has disappeared: a high answer rate can coexist with a human queue if the remaining conversations are complex.

Review the unanswered and escalated questions before each seasonal event. If many shoppers ask the same product question, improve the product data, comparison copy, sizing guidance, or policy explanation before adding people. If the answer is available but shoppers still escalate, the issue may be confidence, ambiguity, or a need for tailored advice rather than missing content.

A workable seasonal operating cadence

Run one forecast for pre-purchase assistance and another for order service, then connect them only at the capacity-planning level. This gives CX leaders a way to flex staff without losing the distinction between helping someone buy and resolving an existing order.

Before the season: clean intent labels, identify the highest-volume product questions, verify source content, and build base and peak scenarios. Confirm which issues must always reach a person.

During the season: review conversation volume, escalation, backlog, and forecast variance at a fixed cadence. Reallocate coverage between pre-purchase intents rather than treating every conversation as interchangeable.

After the season: compare predicted and actual workload, identify questions that caused preventable handoffs, and feed the findings into product content and merchandising. Keep the resulting baseline for the next comparable event.

The central rule is simple: use AI shopping-assistant conversations to forecast the demand for buying guidance, while order records and service tickets forecast post-purchase work. That boundary makes the signal useful to both the ecommerce team trying to protect conversion and the CX team trying to protect service levels.