Case studies

Vishal Retail handles 62,000 customer complaints a month in Nagpur. Each one has to be classified into one of eight categories, summarised, and routed to the right team. Three AI vendors have demonstrated; procurement has recommended the one with the lowest quote.

An analyst at Tarang Beverages typed nine words into an AI assistant: "Create a campaign for our new drink." What came back was polished, confident, and wrong in seven separate ways - and it was two days from launch.

Six in the evening. The board pack is due at ten tomorrow. An analyst asks an AI assistant for the key market growth figure and gets a clean, specific, quotable sentence: the Indian premium snack market grew 28% last year.

A monthly sales file is uploaded to an AI analyst. Eleven seconds later it reports that the South region grew 62% and recommends shifting budget and inventory there. Two days later finance finds a Rs.5.45 crore variance against the general ledger.

A consumer electronics distributor across five regions. The AI-built sales dashboard says revenue is up 12%, every regional bar is the same reassuring blue, and the commentary recommends expanding inventory in growth markets.

An interactive, self-contained case for BBA students

Kalyani Stores is a three-branch grocery chain in Pune with 787 loyalty members. Marketing has three offers for the festive quarter and Rs.90,000.

Suvidha Mobile is a regional postpaid network in Pune with 4,195 subscribers. About 13% will stop using their line in the next 60 days. The retention team can make 400 calls a month. Which 400?

Prayag Infotech has 1,400 employees who ask the HR helpdesk 2,800 policy questions a month. An assistant is built to answer them from the company's own documents. The studio walks a student from a fluent invented answer to a system that gets 95% of its behaviour right - and it does that with no change to the model, no additional data and no retraining.

Trikon Industries, a manufacturer in Coimbatore, receives 4,200 supplier invoices a month. A document AI reads each one, extracts six fields, matches them against the approved purchase order and the goods receipt, and routes the result. A person approves the payment and handles the exceptions.

An online retailer's AI customer-service assistant is ready to launch. The team has demonstrated it answering thirty questions perfectly. A red-teamer is given two weeks, a staging environment and a written set of rules.

Two hundred synthetic applications for a sales trainee role. An AI system ranks them in minutes using three years of past hiring data. The dashboard reports 85% faster, 200 of 200 ranked, top 20 ready.

Twelve pilot stores, twelve matched comparison stores, three months. An AI system forecasts demand and proposes purchase orders; planners review and approve. Now somebody has to say what it was worth.

A customer support outsourcer. An AI assistant drafts replies to customer emails. In the lab it drafts in 8 seconds and scores 92% on a 200-item internal test set. Leadership approves a company-wide rollout.

Vistaar Logistics runs four Pune hubs with 1,852 field staff on the roll. Last quarter 484 of them resigned — 26.1%, or 33.8% once the people who joined and left inside the quarter are counted. HR can hold 120 retention conversations before the next quarter starts. Which 120?

Zenith Skills sells a Rs.48,000 data-analytics certificate. Two months of enquiries produced 6,395 people after deduplication, of whom 7.79% enrolled. The counselling team can properly work 1,200 leads before the intake closes. Which 1,200 — and how would anybody know it was worth doing?

Sahyog Cooperative Bank in Nashik took 9,195 loan applications last year, approved 62.5% of them and lost money on 14% of those. The board wants to lend to more people without losing more money. The bank has repayment records for the loans it made - and nothing at all for the 3,444 it refused.

PayNimbus is a payment gateway that processed 79,991 card transactions last month. 799 of them were fraud - one in a hundred, high for the industry and normal for a young gateway with a risky merchant mix. The review team can open 1,200 alerts a month. The existing rules engine raises 5,968.

Setu General Insurance receives about 14,000 policy questions a month. Every answer is in the policy wordings and finding it takes an agent nine minutes. The assistant retrieves the relevant clause and either quotes it or refuses. The hard part is not answering - it is knowing when to refuse.

Ujjwal Broadband handled 57,995 customer contacts last month. Its assistant closed 41.6% of them without a person, reported as a Rs.22 lakh saving. Nearly a quarter of those customers contacted the company again within seven days about the same problem.

Nagrik Seva, a city corporation's grievance cell, receives 11,795 complaints a month and a clerk reads every one to decide which of eight departments it belongs to. Two clerks given the same complaints agree 80.8% of the time. Can a classifier do this, and what does 'right' mean when the humans disagree?

Meridian Textiles is a 1,200-person mill in Coimbatore. HR answers about 300 policy questions a month by hand - leave, notice, travel, safety - and the answers already exist in twelve documents nobody reads.

Anand Auto Components receives about 600 supplier invoices a month. Each one takes a clerk roughly 11 minutes to key in, check and post - Rs.77 an invoice, Rs.554,400 a year. The students automate as much of that as can be automated safely, and work out exactly where "safely" stops.

Mitra, a customer-facing banking assistant, leaked a customer's balance and waived a fee on a Friday afternoon — because a pasted email chain contained an instruction and the assistant obeyed it. The student is brought in to red-team the system properly: map the attack surface, build a fourteen-probe suite, score and rank the findings, write a report an engineer can act on, and retest the fixes.

Two thousand people applied for forty sales-trainee posts. An AI screen read every application and passed half of them to the human panel. The dashboard said 85% faster, and every number on it was correct.