Digital and AI transformation
Where a company should use digital tools and AI (including generative AI), what it is worth, how to deliver it, and how to get people to use it.
Key takeaways
- An AI use case is worth what it saves or earns once people use it: hours affected x share of time saved x adoption x value per hour, minus the cost to run it, and only if it is feasible to deliver safely.
- List the candidate use cases, size each on value with adoption included, score feasibility, start with the uses that are high on both, and plan delivery and adoption before promising savings.
- Digital and AI questions now appear often, sometimes as a full case and sometimes as a sub-question in operations or growth cases.
- The strong answer ranks uses by value and feasibility and is honest about how savings are captured.
What this case type is and when it shows up
A digital or AI case asks how a company should use technology such as automation, data analytics, or generative AI (tools that write text, code, or summaries) to cut cost, grow revenue, or improve service. It may be a whole case or a sub-question inside an operations or growth case. The client usually has many possible uses and limited money, people, and data, so the task is to choose, size, and plan delivery.
Key idea
An AI use case is worth what it saves or earns once people use it: hours affected x share of time saved x adoption x value per hour, minus the cost to run it, and only if it is feasible to deliver safely.
The underlying theory, in plain language
Value comes in three forms: cost (hours saved, errors avoided), revenue (more sales, better pricing, new products), and risk or quality (fewer mistakes, better compliance). For productivity uses, size value bottom-up: the hours of work affected, the share of that time the tool saves, the share of people who use it (adoption), and the cost of an hour of work.
Feasibility asks whether it can be delivered well: is the data available and clean, is the technology mature enough for this task, what do regulators and customers expect (especially when AI talks to customers), and how hard is it to connect to existing systems. Tools that help employees are usually easier to launch than tools that face customers directly, because a person checks the output.
Hours freed are not cash saved. The saving becomes real only if the company hires fewer people (for example by not replacing staff who leave), takes on more work with the same team, or moves people to more valuable tasks. Say which one you assume.
Delivery can be built in-house, bought as a ready product, or done with a partner (for example a technology provider plus the company's own data and teams). Compare them on total value over the same period, including the value lost while waiting for launch, and on control of data.
Most AI programs fail on adoption, not technology. Plan training, redesign the work process around the tool, set controls for wrong answers, and measure use from the first week.
What the prompts sound like, from simple to hard
- Simple: where should a Singapore bank use AI in its contact centre.
- Medium: should an Indian IT services firm build, buy, or partner for an AI coding assistant.
- Hard: a European insurer wants to cut claims-handling cost 20 percent with AI; build the plan and the business case.
Finding and narrowing the real problem
Key idea
List the candidate use cases, size each on value with adoption included, score feasibility, start with the uses that are high on both, and plan delivery and adoption before promising savings.
- Which AI uses to pursue, and how
- Key: Value
- Cost: hours x share saved x adoption x cost per hour
- Revenue: more sales or better pricing
- Risk and quality: fewer errors
- Feasibility
- Data available and clean
- Technology mature for this task
- Rules and customer trust
- Connection to existing systems
- Delivery: build, buy, or partner
- Time to launch
- Cost to build and run
- Control of data
- Adoption
- Training and process redesign
- Controls for wrong answers
- How freed time is used
Tailor the branches to the client. Value and feasibility decide the order; delivery and adoption decide whether the value arrives.
Frameworks for this type, each as a thinking tool with its limit
- Value versus feasibility: Plot each use case on value and feasibility; start in the high-high corner. Limit: Scores are only as honest as the sizing behind them.
- Productivity sizing: Hours affected x share saved x adoption x cost per hour. Limit: Freed hours become savings only if headcount or workload changes.
- Build, buy, partner: Compare total value over the same period, including time to launch and control of data. Limit: Hard to estimate build time; software projects often run late.
Methods for solving this type
- List use cases across the value chain
- Size value for each, with adoption included
- Score feasibility
- Choose delivery (build, buy, partner) on value over time
- Plan adoption, controls, and how freed time is used
- Start with a measured pilot
The math patterns it relies on
- Hours x share saved x adoption x cost per hour
- Net value = value minus running cost
- Value over a period minus cost, for each delivery option
- Freed hours / hours per person = people-equivalents
Worked cases
Worked case
AI in a Singapore bank's contact centre
The prompt
A Singapore bank has 400 contact-centre agents. Each costs SGD 60,000 a year fully loaded and works about 1,600 productive hours a year. The team has listed three AI use cases (see the exhibit). Assume 80 percent of agents will use whichever tool is launched. Which should the bank start with, and what is it worth?
Interviewer-led: the interviewer shows the exhibit and asks the questions in order.
Clarifying questions, with the interviewer's answers
- Is the goal lower cost or better service?Answer: Both, but the board wants savings visible within a year.
- What rules apply to AI that uses customer data?Answer: The bank has an approved setup for internal tools; customer-facing AI needs extra review, human checks, and more time.
- How should value be counted?Answer: Agent hours freed, at fully loaded cost, and only if the bank can reduce hiring.
A hypothesis to say out loud: Tools that help agents are faster and safer to launch than tools that talk to customers. My hypothesis is that after-call summaries give the best mix of value and feasibility, so they should go first.
The structure
- Rank the use cases on value and feasibility
- Cost of an agent hour
- Key: Value of each use case, with adoption
- Feasibility and running cost
- What happens to the freed time
The exhibit
| Use case | Agent hours affected a year | Share of time saved (%) | Feasibility (1 low to 5 high) |
|---|---|---|---|
| After-call summaries written by AI | 160,000 | 50 | 5 |
| Answer suggestions during calls | 320,000 | 15 | 3 |
| Chatbot that handles simple requests | 128,000 | 40 | 2 |
Working it through
1. Cost of an agent hour
Fully loaded cost divided by productive hours.
Cost per agent hour (SGD):60,000 ÷ 1,600 = 37.52. After-call summaries: hours saved
160,000 hours of after-call work a year, half of which the tool can save, with 80 percent adoption.
Hours saved a year:160,000 × 0.5 × 0.8 = 64,0003. After-call summaries: value
Hours saved times cost per hour.
Value (SGD a year):64,000 × 37.5 = 2,400,0004. Answer suggestions: value
320,000 hours of calls, 15 percent faster, 80 percent adoption.
Value (SGD a year):320,000 × 0.15 × 0.8 × 37.5 = 1,440,0005. Customer chatbot: value
128,000 hours of simple requests, 40 percent handled by the chatbot, 80 percent of the planned volume reached.
Value (SGD a year):128,000 × 0.4 × 0.8 × 37.5 = 1,536,0006. Net value of summaries
Interviewer: "Licences and model usage for the summary tool cost about SGD 400,000 a year."
Net value (SGD a year):2,400,000 - 400,000 = 2,000,0007. What the freed time means
Convert the hours saved into agents. Interviewer: "The bank loses about 60 agents a year through normal staff turnover." So it can capture the saving by not replacing some leavers, without layoffs.
Agent-equivalents freed:64,000 ÷ 1,600 = 40
What the exhibit shows
After-call summaries have the highest value and the highest feasibility. The chatbot has similar value to answer suggestions but the lowest feasibility, because it talks to customers directly.
The recommendation
Start with AI-written after-call summaries. First, they have the highest value: about SGD 2.4 million a year, or SGD 2 million after running costs. Second, they are the most feasible, because an agent checks every summary before saving it. Third, the freed time, about 40 agents' worth, can be captured by not replacing some of the roughly 60 agents who leave each year, so no layoffs are needed. Plan answer suggestions next, and prepare the chatbot in parallel through the bank's review process.
Risks: Adoption may be below 80 percent if summaries need heavy editing; Savings disappear if the bank replaces every leaver anyway; Summaries with wrong details could create complaints.
Next steps: Pilot summaries with 50 agents for six weeks and measure minutes saved per call and edit rates; Agree with the workforce team how many leavers will not be replaced.
A strong candidate
Sized each use case with adoption, compared value with feasibility, netted off running cost, and said how freed hours become savings.
A weak candidate
Recommended the chatbot because it is the most visible, and counted all freed hours as cash saved.
Worked case
An Indian IT services firm: build, buy, or partner?
The prompt
A large Indian IT services firm wants an AI coding assistant for its developers. Should it build one, buy one, or partner with a technology provider?
Candidate-led: you drive the case and ask for data; the interviewer answers only what you ask.
Clarifying questions, with the interviewer's answers
- What share of developer time is spent writing code, and how much does the assistant speed it up?Answer: About 40 percent of time is coding; pilots show about 10 percent faster coding.
- How many developers would use it?Answer: 10,000 developers; about 70 percent used it regularly in the pilot.
- Are there limits on where client code can go?Answer: Yes. Many clients require that their code stays inside the firm's own cloud setup.
A hypothesis to say out loud: Speed to launch matters because every month of delay loses value. My hypothesis is that buying or partnering beats building, and that the data rule on client code will decide between the two.
The structure
- Compare three-year net value of each option
- Yearly value once live
- Buy: fast, licence cost, data leaves the firm
- Build: slow, high upfront cost, full control
- Key: Partner: medium speed and cost, data stays inside
Working it through
1. Yearly value once live
Candidate: "What does a developer cost?" Interviewer: "About INR 1,500,000 a year (INR 15 lakh), fully loaded." Candidate: "Then value is developers x cost x coding share x speed-up x adoption." That is INR 42 crore a year.
Value per year (INR):10,000 × 1,500,000 × 0.4 × 0.1 × 0.7 = 420,000,0002. Buy
Candidate: "What would a ready-made product cost, and how fast?" Interviewer: "INR 1,500 per developer per month, live in one month, but code is processed in the vendor's cloud." Over three years, with two years and eleven months of value:
Buy: 3-year net value (INR):(3 - 1 ÷ 12) × 420,000,000 - 3 × 10,000 × 1,500 × 12 = 685,000,0003. Build
Interviewer: "Building would take about 12 months and INR 250 million, then INR 60 million a year to run." So only two years of value in the three-year window.
Build: 3-year net value (INR):2 × 420,000,000 - (250,000,000 + 2 × 60,000,000) = 470,000,0004. Partner
Interviewer: "A provider would run its model inside our cloud with our code, live in about four months, for INR 100 million a year." That gives two years and eight months of value.
Partner: 3-year net value (INR):(3 - 4 ÷ 12) × 420,000,000 - 3 × 100,000,000 = 820,000,0005. Test the swing assumption
Candidate: "Adoption drives everything. If only 40 percent use it, does partnering still pay?" It does, but by much less.
Partner at 40 percent adoption (INR):(3 - 4 ÷ 12) × 10,000 × 1,500,000 × 0.4 × 0.1 × 0.4 - 3 × 100,000,000 = 340,000,000
The recommendation
Partner with a technology provider that runs its model inside the firm's own cloud. First, it has the highest three-year net value, about INR 82 crore, against INR 68.5 crore for buying and INR 47 crore for building. Second, it meets client rules, because code stays inside the firm, which the ready-made product does not. Third, it launches in about four months, far faster than building. Adoption is the swing factor: at 40 percent use, net value falls to about INR 34 crore, so the plan must include training and usage targets from day one.
Risks: Adoption may stall if developers do not trust the suggestions; The provider's price may rise after the first contract; Faster coding only pays if the firm uses the time for more client work.
Next steps: Negotiate a three-year contract with price protection; Launch with two business units and track weekly usage and code quality.
A strong candidate
Asked for the data needed, compared options on value over the same period including launch delay, used the client-code rule to break the tie, and tested adoption.
A weak candidate
Chose to build "for control" without counting a year of lost value, or bought the product without asking where client code would go.
Prompt: "Where should this bank use AI?"
Weaker answer
Lists many exciting AI ideas, recommends a customer chatbot first, and counts every freed hour as money saved.
Stronger answer
Sizes each use case with adoption, scores feasibility, starts with agent-assist summaries worth about SGD 2 million a year net, and explains how freed hours become savings.
Why the stronger answer wins: The strong answer ranks uses by value and feasibility and is honest about how savings are captured. The weak one chases novelty and overstates value.
Common mistakes, traps, and curveballs
- Counting freed hours as cash saved without saying how
- Leaving out adoption, so value is overstated
- Forgetting running costs such as licences and usage fees
- Starting with the most exciting customer-facing use instead of the most valuable feasible one
- No controls for wrong answers
- Treating it as an IT project instead of a change in how people work
Digital and AI questions now appear often, sometimes as a full case and sometimes as a sub-question in operations or growth cases. Firms with large digital and AI units may go deeper on data and delivery. Formats differ by office and change over time, so check the current process for your target office.
Practice
A Dubai insurer's claims team has 50 staff working 1,800 hours a year each. 30 percent of their time is document review, and AI can save half of that. How many hours a year can it save, assuming everyone uses it?
A tool could save EUR 2,000,000 a year if everyone used it. Only 60 percent of staff use it, and it costs EUR 500,000 a year to run. What is the net value, in EUR a year?
An AI project in the UK costs GBP 600,000 to set up and saves GBP 400,000 a year after running costs. What is the payback, in years?
A use case has high value but low feasibility; another has medium value and high feasibility. Which usually goes first?
An AI tool frees 20,000 hours a year. When does that become a cash saving?
Why include adoption in AI sizing?
Rank AI uses on value and feasibility, size value with adoption included, and say how freed hours become real savings.
Sources for this lesson (1)
- Recognized public explanations of case-interview concepts and frameworks
My notes on this lesson
0 of 5,000 characters. Saves automatically.
Try the 6 remaining checks and drills above to complete this lesson (0 of 6 done).
Spotted something wrong or out of date? Report a mistake. We check every report and correct the page.