Case Study: AI Assistant for an Online Home Goods Store - Bot, Product Texts, and Inquiry Review in Bitrix24

Company Profile and Starting Point

This mid-sized online store sold home goods and renovation supplies through its own website, several marketplaces, and a small wholesale channel for construction crews and interior designers - with Bitrix24 already handling orders, open channels, the sales funnel, and stock data before AI was introduced.

The product catalogue was wide and updated in batches. A small customer-care team handled a continuous inbound stream from the website chat, messengers, and social channels - including nights and weekends when no agents were online. The wholesale side received order specifications as PDF files and scanned documents.

How the order collection from multiple sources, marketplace sync, and processing funnel are set up is covered separately in the Bitrix24 for E-Commerce and Online Retail article. This case is strictly about the next step: adding AI on top of a process that already worked.

What was slowing things down before AI

AI Scenarios We Introduced and How They Work in Bitrix24

Five distinct AI scenarios were layered onto the existing Bitrix24 setup - a customer bot, chat summaries, text generation, wholesale lead scoring, and document recognition - each connected to the CRM data that was already there.

A customer message travels from the first touchpoint to resolution or human handoff through the following path:

A customer writes in any connected channel; an AI bot queries the company knowledge base via RAG and either answers or passes the full conversation with a summary to a human agent, while CoPilot handles text drafts and a Vision-capable model processes incoming PDF specs.

24/7 Customer-Facing Bot in Open Channels

The bot handles stock availability, delivery timelines, return and warranty conditions, and wholesale terms at any hour - answering only from the documents loaded into its knowledge base, not from general internet knowledge.

The bot was built using a Smart Auto-Reply for Open Lines scenario from the Alaio Vibecode catalog - this is a Bitrix24-connected app, not a built-in CoPilot feature. The knowledge base was loaded with the company's own policy documents, FAQ, and terms. The retrieval mechanism (RAG with an embeddings model) finds the relevant fragment and constructs an answer from it.

The escalation boundary is explicit: routine questions (availability, lead times, conditions) stay with the bot; disputed claims, nuanced returns, and financial questions go straight to a human agent, together with the full conversation history. The customer does not need to restart.

One bot instance can run with different instructions on different lines - for example, one configuration answers delivery questions on the website widget while another handles order-status queries in a messenger channel.

Chat Summary on Handoff to a Human Agent

When the bot escalates a conversation, the receiving agent sees an automatic summary - who contacted, what the question was, and what has already been established - so the agent continues the conversation rather than restarting it.

This used Bitrix24 CoPilot's built-in chat analysis: CoPilot reviews the customer chat in the CRM card and fills the form fields once the dialogue has enough content (at least 1,000 characters, with 30 seconds elapsed since the last message, per official documentation). The summary appears directly in the inquiry card.

This was described by the team as the fastest win of the entire project: minimal configuration, immediately visible result.

AI-Drafted Product Descriptions and Review Replies

CoPilot generates first-draft product descriptions and review replies; a human editor reviews and publishes every piece before it goes live - AI handles the structure and tone, not the final word.

Product specifications come from the accounting system - the AI does not invent parameters. CoPilot's role (available in Sites and Stores, per the pricing page) is to take those raw specs and produce readable, structured copy. For marketplace review replies, the model suggests a response in the appropriate tone for a negative or neutral review; the content manager edits and posts.

This is worth stating plainly: the review reply or description is a draft, not a finished output. Skipping the human check is a quality risk.

Bulk Text Generation for New Product Batches

New product batches were processed as planned bulk runs rather than one card at a time, so drafts for a whole batch were ready for review together. Because Essentials plans have AI usage limits, the volume of these runs was planned against the plan the store was on.

When a store has a wide catalogue updated in batches, text generation volume grows quickly, so the team planned bulk runs per batch and kept the human review step for every draft. Note that on Essentials plans there are AI usage limits; "Unlimited AI" starts from Standard Vibe+ (per bitrix24.com/prices).

Wholesale Lead Scoring and Repeat-Purchase Prediction

A lead-scoring app connected to Bitrix24 evaluates wholesale inquiries by deal probability, so managers start each day with the highest-potential requests at the top - not with the oldest ones.

This was implemented using a Lead Scoring app built on Alaio Vibecode (available in the Vibecode solutions catalog), not a native CoPilot feature. For existing wholesale accounts, the system tracks purchase rhythm and signals in advance when a regular buyer is likely due for a reorder. This is the AI repeated-sales logic applied to the wholesale segment.

For more on lead management in Bitrix24, including capture and distribution setup, see the dedicated article.

Document Recognition for Wholesale PDF Specs

A Vision-capable model app reads wholesale order specifications from PDF files and scanned images and populates the CRM card fields; the manager reviews and corrects the result rather than re-entering everything manually.

Construction crews and interior designers typically send order lists as PDFs, spreadsheet exports, or even photographs of handwritten lists. A Vision-capable app built on Alaio Vibecode extracts line items and quantities and maps them to the CRM card. The manager's job shifts from data entry to verification - faster and less prone to input errors.

Internal Wholesale Portal Built with Alaio Vibecode

A team member described the requirement in plain language - a simple portal where wholesale clients can see their orders and statuses - and Alaio Vibecode assembled a working app without a developer.

This is a good illustration of what Alaio Vibecode is suited for: quick internal tools and idea prototypes, not production-grade custom platforms. As the official Vibecode documentation notes, the platform "works best for quick internal apps, idea testing, and tools that solve immediate team needs" and "does not replace development for every task." A Bitrix24 partner can help define the task, build the app in Alaio Vibecode, test the scenario, and deliver a ready-to-use tool.

What Changed in Daily Work

After the AI layer was added, after-hours inquiries were answered immediately, agents handled complex cases rather than routine ones, and new product batches went live with copy on the day they were loaded - not weeks later.

Specific changes observed:

• Night and weekend inquiries - the bot closes routine questions at the moment they arrive, regardless of the time.
• Agent context on pickup - the agent receives a summary and continues where the bot left off. No re-reading, no repeat questions to the customer.
• Product copy availability - new batches get draft descriptions as soon as specs are loaded from the accounting system, rather than waiting for a content manager's available slot.
• Wholesale queue order - the scoring app surfaces high-probability and high-value requests first. A large wholesale order no longer sits behind dozens of smaller retail tickets.
• Team focus - the customer care team handles disputes, complaints, nuanced situations, and relationship-building. The bot handles the volume.

Limits: Where AI Did Not Help

Not every problem yielded to AI, and it is worth being direct about that.

• Disputed or emotional complaints - the bot correctly escalated these, but the underlying complaint-handling process still needed human judgment and, in some cases, process redesign. AI did not fix a weak returns policy; it only got the message to the right person faster.
• Thin knowledge base - the bot's quality is exactly equal to the quality of the documents in its knowledge base. Contradictory policies, outdated delivery conditions, and missing product categories all surfaced as bot errors in the first weeks of live traffic.
• Complex wholesale negotiation - lead scoring flags the opportunity; it does not close it. High-probability wholesale leads still required experienced account management.
• Image-heavy or handwritten specs - Vision model accuracy on poor-quality scans or ambiguous handwriting was imperfect. Managers still needed to check every extracted field carefully.
• Data processing considerations - CoPilot uses third-party AI providers such as OpenAI; requests are processed in the USA and providers may use content under their own terms (per Bitrix24 AI Tools Terms of Use). Teams handling sensitive customer data should review the AI data residency and compliance guide before deployment.

How a Similar Company Can Start

Start with the knowledge base and the handoff summary - those two steps deliver visible results with the least configuration overhead and reveal where the bot will struggle before it faces real customers.

Checklist for a similar e-commerce team:

• Audit and clean the knowledge base before launching the bot. The bot answers exactly what is in those documents. Outdated delivery terms, contradictory return conditions, and missing product categories will appear verbatim in customer replies.
• List every routine question type and decide honestly which ones the bot can close alone. "When will my order arrive?" is routine. "You sent the wrong item - what are you going to do about it?" is not.
• Launch the bot in one channel first and review real conversations, not test scripts. Live traffic exposes knowledge-base gaps and unclear phrasings that structured tests miss.
• Enable CoPilot chat analysis and summary on handoff from day one. It requires minimal setup and delivers an immediately visible improvement in agent handoffs.
• Treat every AI-generated text as a draft that requires human review before publication. This applies to product descriptions, review replies, and any customer-facing copy.
• Plan bulk text generation against your plan. Essentials plans have AI usage limits; "Unlimited AI" starts from Standard Vibe+.
• Add lead scoring and document recognition after the core bot is stable. These require clean CRM data and some calibration time; introduce them when the team is comfortable with the first layer.
• For custom internal tools, explore Alaio Vibecode with a partner. A Bitrix24 Gold partner can scope, build, and test small apps without full development engagement - see the Bitrix24 Vibecode overview for what the platform covers.

For context on AI call analysis in Bitrix24 - transcription, script scoring, and coaching - that topic is covered in a separate guide.