What is AI?
Artificial Intelligence is the engineering of systems that do things we would call intelligent if a person did them: reading a handwritten PIN code off an envelope, spotting a fraudulent UPI transaction, planning a delivery run through Pune traffic, or answering a customer's question in Marathi. It is a field, not a single technique.
Narrow AI, general AI and the hype in between
| Term | What it means | Where it stands |
|---|---|---|
| Narrow (weak) AI | Very good at the one family of tasks it was built for | Everything in production today |
| General AI (AGI) | Matches a human across any intellectual task and transfers skill between them | A research goal, not a product |
| Superintelligence | Far beyond human ability in every domain | Speculation |
A model that writes essays, passes exams and drafts code is still narrow AI: astonishingly broad within language, and helpless the moment it has to lift a carton. Keeping the distinction in mind saves you from both over-promising to a client and dismissing the technology as a toy.
The main subfields
- Machine learning — learn the rules from examples instead of writing them.
- Natural language processing — text and speech: classification, extraction, translation, generation.
- Computer vision — images and video: recognition, detection, OCR.
- Robotics and control — perception plus action in the physical world.
- Search and planning — find a sequence of moves that reaches a goal (routing, scheduling, timetabling).
- Knowledge representation and reasoning — encode facts and rules a machine can reason over.
Generative AI, the part everyone talks about, sits mostly inside NLP and vision: a very powerful form of narrow AI trained on enormous datasets.
Two ways to build the same feature
Say a kirana billing app must flag "large" invoices for the owner's approval. Version one is a hand-written rule; version two learns the threshold from history.
# 1. Symbolic AI: a human states the rule
def needs_approval(invoice_total):
return invoice_total > 25000 # Rs 25,000, decided by the shop owner
# 2. Machine learning: the rule is learned from labelled examples
from sklearn.tree import DecisionTreeClassifier
totals = [[450], [1200], [8000], [26000], [31000], [90000]]
reviewed = [0, 0, 0, 1, 1, 1] # did the owner actually review it?
model = DecisionTreeClassifier(max_depth=1).fit(totals, reviewed)
print(model.predict([[24000], [40000]])) # [0 1]
Both are AI. The rule is transparent and free to run; the model finds the threshold itself and keeps up as the business changes — but it needs data, and it can be wrong in ways nobody wrote down.
Where you already meet AI in India
Cash-deposit machines reading amounts, e-commerce apps ranking search results, keyboards suggesting the next Hindi word, telecom firms predicting which tower will fail, tax portals matching invoices, hospitals triaging X-rays. Almost none of these are chatbots — worth remembering when a client says "we want AI".
Common mistakes
- Treating "AI" and "a chatbot" as the same thing, and ignoring cheaper classical techniques.
- Believing that a fluent answer implies understanding; fluency and correctness are separate properties.
- Starting from a model instead of from a measurable decision the model is supposed to improve.
- Planning products around AGI arriving next quarter.
Practice
List five decisions taken every day in a business you know well. For each, write down whether it is a fixed rule, a pattern learned from history, or a human judgement call — and which subfield above would apply if you automated it.