Chapter 17 of 17
Customer Interaction Automation and Exploration of AI Tools
Covers the design of chatbots, virtual assistants and voice AI, then evaluates RunwayML, Jasper, Synthesia, Lumen5 and Gemini. It ends with a capstone: an end-to-end AI workflow for launching a product, with cases from IRCTC AskDISHA and others.
In this chapter
17.1 Chatbots, virtual assistants and voice AI
| Type | How it works | Example |
|---|---|---|
| Rule-based chatbot | Fixed menus and keyword rules | "Press 1 for balance" style flows |
| Intent-based (NLU) chatbot | Classifies the user's intent and fills slots (date, PNR, amount) | Railway or banking FAQ bots |
| Generative AI chatbot | An LLM writes replies, grounded in the company's knowledge base, and can call tools | Modern support assistants |
| Virtual assistant | Performs tasks for a user across apps | Alexa, Google Assistant, Siri, Copilot |
| Voice AI | Speech-to-text, then the bot, then text-to-speech | Voice bots on customer helplines |
Architecture of a generative support bot
- Channel: website, app, WhatsApp, phone.
- Speech layer (for voice): speech recognition in, speech synthesis out.
- Understanding: detect language, intent and sentiment.
- Retrieval: fetch the relevant policy, order or account data (RAG).
- Generation: the LLM drafts an answer from the retrieved data within set rules.
- Action: call systems to check status, raise a ticket or book a slot.
- Guardrails: block unsafe, off-topic or unauthorised responses; mask personal data.
- Human handoff: transfer to an agent when the bot is unsure, the customer is upset, or the case is complex.
- Analytics: log conversations and measure performance.
Design principles
- Tell users they are talking to an AI.
- Make reaching a human easy.
- Ground answers in approved content, never in the model's general memory, for policies and prices.
- Support local languages; in India, the government's Bhashini platform offers translation and speech services for Indian languages.
- Never let the bot promise refunds, legal outcomes or medical advice outside policy (recall the Air Canada case, Chapter 3).
Key metrics: containment rate (share resolved without a human), first-contact resolution, average handling time, customer satisfaction (CSAT), escalation rate, and accuracy on audited samples.
17.2 Exploration of popular AI tools
| Tool | Category | What it does | Typical users |
|---|---|---|---|
| RunwayML | Video generation and editing | Text-to-video and image-to-video generation, background removal, motion tracking, inpainting of video | Filmmakers, editors, ad agencies |
| Jasper | Marketing copywriting | Blog posts, ads, emails and social copy, with brand-voice settings and templates | Marketing teams, agencies |
| Synthesia | AI avatar video | Turns a script into a video with a digital presenter in many languages | Corporate training, HR, internal communication |
| Lumen5 | Text-to-video | Turns blog posts or scripts into short videos with stock footage, captions and music | Content marketers, social media teams |
| Bard (now Gemini) | General assistant | Google's chatbot, launched as Bard in 2023 and renamed Gemini in February 2024 | Everyone; Google Workspace users |
RunwayML. Founded in 2018 by graduates of NYU's Tisch School of the Arts, Runway co-released the original Stable Diffusion model with the CompVis group at LMU Munich in August 2022. Its tools were used in the Oscar-winning film Everything Everywhere All at Once. In September 2024 Lionsgate partnered with Runway to build a custom video model trained on its own film catalogue (Wikipedia: Runway).
Jasper. Launched in 2021 (originally as "Jarvis"), Jasper focuses on marketing content. Its features include brand voice (learning a company's style), campaign templates and team collaboration.
Synthesia. Founded in London in 2017, Synthesia generates videos with AI presenters from text. Its voice library covered about 140 languages and accents by 2025, and the company reports that over 90% of Fortune 100 companies use it (Wikipedia: Synthesia). Its main use is training and internal communication, where updating a video means editing the script, not reshooting.
Lumen5. Lumen5 reads an article, picks key sentences, and pairs them with stock visuals and music to create a short social video, which the user can then edit.
Bard / Gemini. Google's assistant is covered in Chapter 4. For this chapter, note its value for research with links to sources, integration with Gmail and Docs, and multimodal input.
Ethical points for media tools: get consent before creating an avatar or voice clone of a real person; label synthetic presenters; check the licence terms for commercial use of generated media.
17.3 Evaluating and comparing tools (CO4)
Use a consistent method, so the comparison is fair and repeatable.
- Define the task: for example, "produce a 2-minute product training video in Hindi and English".
- Choose criteria and weights: output quality, ease of use, speed, cost, language support, integration, privacy and commercial licence.
- Run the same task on each tool.
- Score each criterion from 1 to 5, ideally with two or three reviewers.
- Compute weighted totals (Numerical 17.2).
- Check deal-breakers: a tool that fails on privacy or licence is out, whatever its score.
17.4 Capstone: designing an end-to-end AI workflow (CO5)
Scenario. A Kanpur-based leather goods startup is launching a new laptop bag online next month. It has three staff and a small budget.
| Stage | Tool | AI task | Human checkpoint | Measure |
|---|---|---|---|---|
| 1. Research | ChatGPT or Gemini (with web search) | Summarise competitor prices and features; draft buyer personas | Founder checks sources | Persona sign-off |
| 2. Product copy | ChatGPT or Jasper | Product title, description, 3 benefit bullets in English and Hindi | Verify claims (no false "genuine leather" grades) | Listing approved |
| 3. Landing page | GitHub Copilot | Generate HTML/CSS and a contact form | Developer tests on mobile; security check on form | Page speed, form works |
| 4. Visuals | DALL·E or Midjourney, plus product photos | Lifestyle backgrounds and social graphics | Images must show the real product accurately | Brand consistency check |
| 5. Video | Lumen5 and Synthesia | Short promo from the launch blog; how-to-use video with avatar | Script and avatar reviewed; AI label added | Video completion rate |
| 6. Campaign | Jasper or ChatGPT; email platform | Social calendar, ad variants, email drip | Founder approves schedule and offers | CTR, open rate |
| 7. Support | Generative chatbot on website and WhatsApp | Answer FAQs from policy; take order queries | Handoff to staff for returns and complaints | Containment, CSAT |
| 8. Review | ChatGPT on exported analytics | Summarise results; suggest improvements | Team decides next steps | Sales vs target |
Design rules applied: AI drafts and humans decide; every customer-facing claim is checked; personal data stays in approved systems; and each stage has a metric so the workflow can be improved.
17.5 Worked numericals
Numerical 17.1: Chatbot cost and containment (illustrative costs).
A company handles 30,000 chats a month. A human-handled chat costs ₹60. With a bot, every chat starts with the bot at ₹5 per chat, and 65% are resolved there; the other 35% go to humans.
If containment fell to 40% because the bot answered poorly, the cost would be 1,50,000 + 18,000 × 60 = ₹12,30,000, and unhappy customers would add hidden costs. Quality drives savings.
Numerical 17.2: Weighted comparison of video tools (class exercise scores).
A training team compares two approaches for 20 short training videos.
| Criterion | Weight | Avatar video tool | Text-to-video tool |
|---|---|---|---|
| Fit for training content | 0.30 | 5 | 3 |
| Language support | 0.25 | 5 | 3 |
| Ease of updating | 0.20 | 5 | 4 |
| Cost | 0.15 | 3 | 4 |
| Visual variety | 0.10 | 3 | 5 |
For training videos the avatar tool wins. For social media teasers, where visual variety would carry more weight, the result could reverse.
Numerical 17.3: Video production cost (illustrative).
Twenty training videos cost ₹40,000 each at a studio, so ₹8,00,000. With an AI video platform at ₹1,50,000 a year plus 4 staff-hours per video at ₹500 an hour:
The saving is ₹6,10,000, about 76%, and later edits cost almost nothing.
17.6 Industry use cases
- IRCTC AskDISHA (India). Indian Railways' ticketing arm IRCTC launched AskDISHA, built by CoRover on Microsoft Azure, in October 2018. It answers questions on bookings, refunds and PNR status by voice and text in English and Hindi. Microsoft reported about 150,000 queries a day, 90% response accuracy, and a 70% fall in queries through social media, phone and email (Microsoft, January 2021).
- Klarna AI assistant. See Chapter 8: 2.3 million conversations in the first month and resolution time cut from 11 to under 2 minutes, by the company's account.
- Runway in film. Runway's AI tools were used in the visual effects work for Everything Everywhere All at Once, and Lionsgate commissioned a custom model trained on its own catalogue in 2024.
- Synthesia in corporate training. Companies use AI presenters to produce and update multilingual training videos without filming.
17.7 Review questions
- Draw and explain the nine-part architecture of a generative support chatbot.
- Why must a support bot offer easy human handoff? Relate your answer to the Air Canada case.
- Compare RunwayML, Synthesia and Lumen5 for (a) a corporate compliance course, (b) a 30-second Instagram teaser, (c) a short film's visual effects.
- Evaluate Jasper against a general assistant such as ChatGPT for a marketing agency.
- In Numerical 17.1, what containment rate makes the bot break even with the all-human cost? (Answer: 30,000 × 5 + (1 − c) × 30,000 × 60 = 18,00,000 gives c ≈ 8.3%.)
- Design an end-to-end AI workflow, with tools, checkpoints and metrics, for a college to run its annual admission campaign.
Knowledge sources
- Microsoft Source Asia, Powered by Microsoft Azure and developed by CoRover, IRCTC's AI chatbot AskDISHA enhances user experience, 5 January 2021.
- Wikipedia: Runway (company) and Wikipedia: Synthesia (company), accessed 23 October 2026.
- Product sites: Runway, Jasper, Synthesia, Lumen5, Gemini.
- Government of India, Bhashini: National Language Translation Mission.
- OpenAI, Klarna customer story, 2024.
Independent study notes for learning purposes. Not affiliated with any university or course provider. Facts and case-study figures are linked to their original sources. ChatGPT, Gemini, Copilot, Jasper, Synthesia, Runway, Lumen5 and other product names are trademarks of their respective owners.