AI-Powered Automatic Follow-Up After Customer Inquiry

A customer contacted support, the issue was resolved, but a week later they churned to a competitor. Why? Because the follow-up was either missing or templated. We built an AI system that automatically sends personalized follow-up messages, boosting retention by 15–20% and CSAT by 10 points. The system analyzes the conversation, extracts commitments, and generates a message in the right channel at the right time. It accounts for the customer's time zone, interaction history, and selects the optimal sending moment (typically 9:00–18:00 local time). For a typical SaaS client with 50K users, the AI follow-up system saves up to $50,000 per month in prevented churn.

Why templated follow-ups fail

80% of customers don't open standard "thank you for your inquiry" messages. We use LLM-based generation (GPT-4, LLaMA 3) with ticket context — this yields +40% response rate. Incorrect timing kills conversion: follow-ups on weekends or at night annoy. The system determines the customer's time zone via geo-IP or profile and schedules delivery during business hours. Closed loop without feedback: if the customer doesn't respond, the system escalates; if they respond with a problem, a new high-priority ticket is created. This approach reduces churn by 15% in the first 30 days.

How AI selects the right channel and message

The channel is determined by history: if the customer communicated via email — send email, if via chat — push. The text is generated via an RAG pipeline that uses contextual bandits for optimal timing optimization:

def generate_followup(ticket: Ticket, days_after: int) -> FollowUpMessage:
    prompt = f"""Create a personalized follow-up message.

Ticket: {ticket.subject}
Resolution: {ticket.resolution}
Customer name: {ticket.customer_name}
Days passed: {days_after}

Requirements: brief (2–3 sentences), personal, with a concrete detail from the conversation."""

    content = llm.generate(prompt)
    channel = select_channel(ticket.customer)
    return FollowUpMessage(content=content, channel=channel, scheduled_at=calculate_time())

We use fine-tuning via LoRA on your data — this reduces hallucinations and improves accuracy in mentioning details. ChromaDB vector store holds token-level embeddings with dimension 1536, enabling fast retrieval of similar inquiries. For production, we use INT8 quantization and model distillation, reducing p99 latency to 200 ms and saving GPU memory.

What personalized messages deliver

Personalized messages are opened 1.9 times more often than templated ones. AI-generated follow-ups are 2.5 times more effective in improving CSAT. Comparison of A/B test results on 10,000 tickets (Bayesian A/B testing used):

Data obtained during a pilot project for a SaaS platform with 50K active users. RAG-based personalization yields a 3x higher response rate compared to rule-based systems.

Process overview

1. Data audit — collect ticket history, logs, CSAT surveys. Check quality and completeness. Minimum volume is 5,000 tickets, but 20,000+ is preferable for a stable model.
2. Architecture design — select LLM (GPT-4 or LLaMA 3), vector DB (ChromaDB), pipeline on LangChain.
3. Development and training — fine-tuning via LoRA, RAG setup, CRM integration via REST API. We use MLflow and adopt MLOps best practices to track experiments.
4. A/B testing — compare with current process. Lasts 2 weeks, monitoring response rate and CSAT.
5. Launch and monitoring — deploy on Kubernetes with Ray Serve, connect Grafana dashboard with metrics: p99 latency, GPU utilization, follow-ups per hour. Use GPU memory profiling for optimization.

What's included

Based on our project experience — with 5+ years in AI, 20+ successful deployments, and engineers holding AWS and GCP certifications — we guarantee stable operation under loads up to 1,000 follow-ups/hour. All components use an open-source stack: PyTorch, Hugging Face Transformers, LangChain, Ray for scaling. Our team's combined 30+ years of ML experience ensures reliable infrastructure.

Timeline and pricing

Timeline: from 4 to 8 weeks depending on integration complexity. Pricing is calculated individually — we'll estimate the project within 2 days after the brief. Pilot project (up to 10,000 tickets) — from $5,000.

Order a pilot project on 10,000 tickets to evaluate the effect. Contact us — we'll prepare a custom proposal within two days. Get a consultation on your project today. Message personalization and follow-up automation are at the core; using an LLM for follow-up ensures context-aware messages. For customer reactivation, the system targets inactive users based on their history. Integrating AI in support workflows boosts efficiency and churn prevention.

NLP Development: Text Classification, NER, Embeddings, and Information Extraction

We often receive a task: process 50,000 support tickets — currently all manual. Dataset — 3,000 labeled examples, 12 categories, imbalance: one category occupies 40% of the sample, three at 1-2% each. Baseline accuracy — 78%. Sounds decent until you look at recall for rare classes: 0.31, 0.44, 0.28. These classes — complaints and churn threats — are most important to the business.

This is a typical NLP development project. The problem is not the algorithm but that accuracy is the wrong metric. Our experience across 30+ projects shows: we start by analyzing business metrics and only then choose the model.

Why accuracy is not the right metric for rare classes?

Accuracy ignores imbalance. If the "churn" class appears in 2% of cases, the model can predict "all good" and get 98% accuracy — but the business loses clients. Solution: F1 macro (averaged over all classes) or weighted F1. For NER — strict entity F1 (exact matches only). We guarantee: after choosing the correct metric, model quality becomes measurable and predictable.

Text Classification: From BERT to Distillation

BERT-like models are the standard for classification. ruBERT-base or ruBERT-large from DeepPavlov for Russian. multilingual-e5-large — for multiple languages in one pipeline. XLM-RoBERTa-large — a strong multilingual backbone.

Fine-tuning for classification: add a classification head on top of the [CLS] token, train for 3-5 epochs with lr=2e-5, weight decay=0.01. For imbalance — weighted CrossEntropyLoss or focal loss with gamma=2.0. Contact us — we will show a code snippet.

Imbalance case study. Dataset — 3,000 examples, imbalance 1:20. Solution: class_weight via sklearn + CrossEntropyLoss. Additionally — augmentation of rare classes via backtranslation (ru→en→ru through MarianMT). Recall for rare classes rose from 0.31 to 0.67 with a slight drop in accuracy (76%→74%). Full NLP development end-to-end took 3 weeks.

Distillation for production. BERT-large gives F1 0.89, but inference on CPU — 180ms. Distillation into DistilBERT or ruBERT-tiny2 reduces latency to 25ms with F1 0.84. Export to ONNX Runtime provides an additional 1.5-2x speedup. DistilBERT achieves 7x lower latency than BERT-large with only a 5% drop in macro F1 – a typical production trade-off.

How to choose between BERT and LLM for your task?

For most classification and extraction tasks, BERT-sized models offer the best trade-off between cost and performance. Shift to LLMs only when the task demands generation, complex reasoning, or zero-shot generalization.

NER: Named Entity Recognition

NER — extracting persons, organizations, locations, dates, amounts, document numbers. For general categories (PER, ORG, LOC), pre-trained models work well. For specialized ones (medical terms, legal concepts) — fine-tuning is needed.

Data annotation. The main cost of an NER project. For a quality model — 500-2,000 labeled sentences per entity type. Tools: Label Studio (open source) or Prodigy (by spaCy creators). IOB2 format — standard.

Architecture. Token classification on top of BERT: each token gets a label (B-PER, I-PER, O). spaCy 3.x with transformer pipeline — a convenient production choice.

Nested entities. Standard IOB models cannot handle nested entities (organization inside an address). For such tasks — span-based NER: SpanBERT or SpERT. More complex but correct.

Post-processing is mandatory. The model predicts tokens — normalized entities are needed. Date — dateparser. Amounts — regex + validation. Names — deduplication via rapidfuzz. Included in our standard delivery.

Sentiment Analysis and Opinion Mining

Binary classification positive/negative works out of the box with BERT. Complexity — aspect-based sentiment analysis (ABSA): "the restaurant has good food but terrible service." For ABSA: aspect extraction (NER) + sentiment per aspect. Joint models BERT-for-ABSA — quality on Russian data is lower due to dataset scarcity. RuSentiment, SentiRuEval — main resources.

For production with simple positive/negative/neutral: distil models are enough. Three classes, balanced dataset, 2,000+ examples — F1 macro 0.82-0.87 in 1-2 days.

Text Summarization

Extractive summarization (select sentences) — TextRank or BM25 without training. Fast, no hallucinations. Good for long documents.

Abstractive (generates new text) — seq2seq: mT5, mBART, FRED-T5, ruT5-large. For production via LLM API (GPT-4, Claude) — often the best cost/quality/speed trade-off.

Embeddings: Vector Representations of Text

Embeddings are the foundation of semantic search, deduplication, clustering, RAG. Quality critically affects downstream tasks.

Models. E5-large-v2, BGE-M3, multilingual-e5-large — strong multilingual embedders. sentence-transformers/paraphrase-multilingual-mpnet-base-v2 — fast option. For Russian: ru-en-RoSBERTa (Skoltech) performs well on semantic textual similarity.

Embedding quality evaluation uses the MTEB benchmark as standard. But top results on MTEB don't guarantee success on a domain dataset — we build domain-specific eval.

Fine-tuning embeddings. If standard models don't give the required Recall@k — contrastive learning on domain pairs with MultipleNegativesRankingLoss. How to perform this for domain data:

1. Collect 500–2,000 semantically similar pairs from your domain.
2. Apply MultipleNegativesRankingLoss with a batch size of 32–64.
3. Train for 1–3 epochs using AdamW (lr=2e-5).
4. Evaluate Recall@k on a held-out domain test set.

This approach yields a 5–15% improvement in Recall@k in practice.

Dimensionality and storage. E5-large: 1024 dim, float32 — 4KB per vector. For 10M documents — 40GB. Quantization int8 reduces to 10GB. FAISS IVF_PQ — more compact but with losses. Included in our deployment recommendations.

Information Extraction

Structured extraction is a frequent task. Examples: key contract terms, technical characteristics, dates and amounts from invoices.

1. Regex + rule-based. For INN, OGRN, amounts, dates — more reliable than neural networks. No data required.
2. NER + post-processing. For variable formats.
3. LLM with structured output. GPT‑4 / Claude with JSON schema — for complex documents. Cost: minimal per document. For 10k+ documents/day — we calculate the economics.

We guarantee a hybrid: regex/NER for typical fields + LLM for edge cases. Our guarantee is backed by years of production experience and more than 30 projects.

Work Stages

Prototype on existing data — 1-3 weeks. Production system with CI/CD — 1.5–2.5 months. Cost is calculated individually — get a consultation for a project estimate.

What's Included

• Model and pipeline architecture documentation
• Access to the model via REST API (FastAPI + ONNX)
• Client team training (2-hour webinar + Q&A)
• Accuracy guarantee on the agreed test set
• Months of post-delivery support (bug fixes, adaptation to new data)

Our Experience

Years of NLP projects from classification to RAG systems. The team includes ML engineers experienced with Hugging Face, spaCy, LangChain, MLOps. We use vLLM, Kubeflow, Weights & Biases — a production stack, not toys. Contact us to evaluate your NLP project within two days — request a free consultation on your text processing pipeline.