Learning Objectives
By the end of this chapter, you will be able to:
- Define AI for IT 231 and separate narrow AI from general AI.
- Classify AI branches into three groups by what they do.
- Explain business impact through productivity, products, and risk.
- Explain society impact through jobs, access, fairness, and information quality.
- Apply six principles of responsible AI to a Nepali organization.
Detail: data science, ML, and neural networks, cloud, green, and virtual computing, big data and blockchain, social media and digital marketing, IoT.
What AI means in this course
AI is the field that builds systems to do tasks that usually need human intelligence. Reading handwriting. Recognizing a face. Translating Nepali to English. Recommending a product. Flagging a fraudulent wallet transaction. Drafting a reply.
Two distinctions matter.
Narrow AI handles one task well. Every business tool today is narrow. A chatbot that answers fee queries cannot drive a truck.
General AI would handle any cognitive task at human level. It does not exist as a deployed system. Exams confuse the two. Business decisions use narrow AI, with human oversight.
Machine learning, covered in the next chapter, is the usual method for building AI. The system learns patterns from examples instead of following only hand coded rules. Neural networks are one family of machine learning models that power vision, speech, and language.
Three groups: what AI does
Do not memorize seven branches as a flat list. Group them by function. Each group solves a different management problem.
| Group | Question it answers | Branches | Typical business use |
|---|---|---|---|
| A. Understand | What was said or seen | Natural language processing, speech recognition, computer vision | Chatbot for admission queries, OCR for invoices, number plate reading at parking |
| B. Decide | What will happen and what to suggest | Predictive analytics, recommendation systems | Sales forecast, credit scoring, churn prediction, next product to show |
| C. Create and act | What new content or action to produce | Generative AI, robotics and automation | Drafting product descriptions and code help, warehouse packing, RPA for data entry |
flowchart TB
AI[AI: Narrow systems that learn from data]
AI --> U[Group A Understand: language speech vision]
AI --> D[Group B Decide: predict and recommend]
AI --> C[Group C Create and Act: generate and automate]
U --> E[Chat, translate, transcribe, see]
D --> F[Forecast, score, rank, suggest]
C --> G[Draft, draw, move, file]
Rule based software versus learning system
A rule says: if transaction is over Rs 500,000 then flag it. It catches only what the author imagined. An AI fraud model studies thousands of past frauds and learns a pattern: small amount, 2 a.m., new device, new location, repeated retries. It flags Rs 20,000 that the fixed rule misses. The price is explainability. A rule can point to line 41. A model points to weights and probabilities. Managers choose based on cost of error and need for explanation.
Where AI fits in an information system
AI is not a replacement for data quality, process design, or accountability. A model trained on dirty or biased ledgers will automate the dirt and bias at speed. Treat AI as a component inside a system, with inputs, controls, and a named owner.
Impact on business: three lenses
| Lens | Gain | Where it shows | Manager question |
|---|---|---|---|
| Productivity | Automate routine classification, answering, drafting, forecasting | Customer chat, document summary, invoice sorting | How many staff hours shift from typing to checking |
| New products | Services that did not exist without pattern learning | Recommendation engine, route optimization, personalized pricing | Does data advantage create a moat or a dependency |
| Risk | Scale can scale mistakes | Hallucinated facts, privacy exposure, vendor lock in, silent bias | Who reviews, who is liable, how to stop it |
By function:
| Function | AI application |
|---|---|
| Marketing | Segmentation, ad targeting, drafting content, social listening |
| Sales and service | FAQ chatbot with human handover, lead scoring |
| Finance | Fraud detection, credit scoring, invoice and reconciliation |
| HR | CV screening with bias audit, staff query bot |
| Operations | Demand forecast, route planning, predictive maintenance |
Impact on society
Jobs. Routine entry, basic translation, and simple categorization shrink. Oversight, data stewardship, exception handling, and prompt and process design grow. The shift is uneven by sector and by schooling. Displacement is not abstract. It is a cashier whose till becomes an app.
Access. Language models trained mainly on English can underperform on Nepali and Maithili. The same tools, when tuned for Nepali speech and script, open services to citizens who never used English forms.
Fairness. Past hiring or lending bias, encoded in training data, reappears as model output. A loan model that punishes an address because past officers did will repeat the injustice politely.
Public life. Deepfakes, synthetic reviews, and micro targeted misinformation erode trust. Surveillance with face recognition raises questions answered in Unit 9 alongside law and ethics.
Service reach. In health and education, AI can extend scarce expertise. A health post can send a phone photo for a second opinion. A college can personalize practice sets. The gain is real only with connectivity, verification, and consent.
| Benefit | Concern |
|---|---|
| Faster and cheaper services, available 24 hours | Job loss in routine roles without retraining |
| Better decisions when data is sound | Biased or wrong decisions at scale |
| New industries and export services | Privacy and surveillance creep |
| Help for language, disability, and remote access | Misinformation that looks authentic |
| Support for doctors, teachers, farmers | Overreliance on confident but unverified output |
Self-check
Responsible AI in an organization
Principles are controls. Controls need owners.
- Human oversight. A person reviews consequential decisions on loans, hiring, or medical advice.
- Transparency. Customers are told when they speak to a bot and how their data is used.
- Fairness checks. Test whether error rates differ by gender, caste, province, or age. Fix data or thresholds before release.
- Privacy. Collect the minimum needed, store it securely, get consent, and keep a retention period.
- Accuracy checks. Verify generative output before publication. Treat the model as an intern who writes fast and sometimes invents.
- Accountability. A named team owns the system, its logs, and its rollback plan.
Nepal worked example: chatbot plus human path for a Kathmandu retailer (NPR, BS 2081)
A two store apparel retailer in New Road runs Instagram and Facebook messages as its storefront. On 2081-07-15 it receives 220 inquiries per day. Owner Ramesh budgets for an AI chatbot to handle FAQs on size, delivery, and return.
Vendor quote for 2081 Shrawan:
- Chatbot setup and Nepali tuning: NPR 45,000 one time
- Monthly SaaS and WhatsApp API: NPR 8,500
- Staff training and prompt library: NPR 12,000 one time
- Expected automation: 65 percent of inquiries, average handle time from 6 minutes to 45 seconds for bot answered chats
- Human agent for the remaining 35 percent and all payment and complaint issues
First year cost = 45,000 + 12,000 + (8,500 x 12) = NPR 159,000. Estimated labour saving: one full time agent at NPR 25,000 per month redeployed to store sales, 12 months = NPR 300,000 in redeployed value. Net = 300,000 - 159,000 = NPR 141,000 gain, if the bot meets 65 percent automation without increasing refunds or complaints.
The guardrails matter more than the arithmetic. Ramesh keeps three: a clear bot disclosure in Nepali, a button that reaches a human in under 30 seconds, and weekly review of bot answers that trigger returns. TheElectronic Transactions Act and consumer protection still apply when a bot promises a delivery date. Savings mean little if the model invents a return policy that the shop must then honor.
The same pattern fits a cooperative that answers member queries, a college admission desk during Shrawan intake, or a clinic that triages common questions. Automate the frequent and low risk. Keep the human path for money, identity, and complaints.
Stream lens: BBA In a BBA cooperative, Group B decides. Ask: can a credit scoring model trained on 5,000 past loans cut default by 1.2 percent without raising false rejection for women and for Karnali members. If not, change thresholds or keep human review for that segment.
Stream lens: BBM In a BBM retail chain, Group A and C decide. Ask: can a vision model count footfall per aisle from CCTV and can a generative tool draft daily shelf restock notes that a supervisor approves in 10 minutes. Speed without approval is a risk, not a gain.
Summary
AI in IT 231 is narrow, data driven assistance that understands, decides, or creates. Its three groups map to business functions and to distinct controls. Productivity and new products are real. So are job shifts, bias, and misinformation. Responsible use needs oversight, disclosure, testing, privacy, verification, and ownership. Cost math helps, but guardrails decide whether the math holds.
Key Takeaways
- Narrow AI does specific tasks well. General AI is not the business case.
- Groups A, B, and C tell you which problem you are solving and which data you need.
- Business value shows as productivity, new services, and risk that scales.
- Society feels AI through jobs, language, fairness, and trust in information.
- Responsible AI needs a human in the loop, fairness and accuracy tests, and a named owner.
Discussion Questions
- Name one SME task that is a good AI candidate and one that should stay human. Why does the grouping explain the difference.
- Why is the model said so a weak control for audit or appeal.
- How does AI rely on cloud and big data yet remain a separate decision about risk and accountability.
Exercises
Short answer
- Define narrow AI and give two Nepali business examples.
- List the three groups and two branches in each.
- State four principles of responsible AI and one check for each.
Long answer
- Discuss business and society impacts of AI with grouped examples. Use one table for business functions and one for benefits and concerns.
- A Nepali bank wants AI for loan approval. Explain benefits, risks, and six safeguards, and propose a human review rule for loans above a threshold.


