Building AI Products, Part 1: How To Find AI Use Cases Worth Building in the First Place

In February 2024, Klarna announced its AI assistant was handling two-thirds of customer chats. It was doing the work of 700 agents and cutting resolution time from 11 minutes to 2. It looked like the AI success story everyone wanted. Fifteen months later, Klarna was hiring humans again. The CEO admitted that a focus on cost had led to "lower quality" and customers still wanted a person for complex cases. The AI stayed and alongside people.So which Klarna got it right: the one that automated 700 agents' worth of work or the one that hired people back? Picture two chats. "Where's my refund?" has one answer, sitting in a database. "Why was I charged twice?" needs someone to hear the customer out and make a call. AI can handle the first on its own. The second needs a person or at least a person checking.Most AI roadmaps skip that distinction. They list places to add AI, not jobs to split between AI, software and people. The rest of this piece shows how to make that split: a map of where AI value sits, then three tests for any idea. Is it worth it? Is it suitable? How do we build it?First, know where your idea sitsThe AI economy has four layers and value grows as you move from answering to doing:Foundation ("tell me"): the models and compute. OpenAI, Sarvam AI, NVIDIA.Horizontal ("help me"): one pattern sold to every industry. ChatGPT, Freshworks Freddy AI.Vertical ("do this for my industry"): one industry's workflow, data and rules. Harvey in law, Qure.ai in medical imaging.Agents ("get this done for me"): systems that plan, use tools, act and check until the job is done. Once AI can act, permissions, audit logs and approvals become product features.Then run the strategic test: if the AI disappeared tomorrow, what's left?If nothing is left, the product is AI-native, like ChatGPT or Perplexity.If it still works but worse, it's AI-enhanced, like GitHub with Copilot or Gmail's "Help me write".If the core value is intact, it's AI-adjacent, like a payments app whose HR team writes job ads with AI.Each position calls for a different strategy, so know which one you're in before you go further.1. Is it worth it?Using AI is easy. Earning from it? McKinsey's 2026 State of AI survey finds nearly 9 in 10 organisations use AI but only 37% see any EBIT impact(Earnings Before Interest and Taxes) and about 6% earn more than 5% of EBIT from it.Alexa is the cautionary tale: used daily for timers and weather, rarely for shopping. As one former senior Amazon employee told the Wall Street Journal: "We worried we've hired 10,000 people and we've built a smart timer."Heavy use doesn't always equal value. And unlike normal software, every AI answer costs money. Whether you charge per seat, per call, per outcome or a hybrid of these, a flat price can lose money on your heaviest users, so model them first.Before building, run a 60-second TASQ screen:Task ambiguity: does the input vary every time? (Code review, yes. Payroll tax, no.)Automation: does it run end to end, or does someone still copy, check and route the output by hand?Scale: does it happen often enough to matter? A task done 10 times a year rarely is.Quality tolerance: is a mistake caught before it causes harm? (A bad film suggestion, fine. A drug dosage, no)A pass means look closer, not build. Next, put a rupee number on it: Value = volume × value per task × automation rate − costA support bot handling 50,000 tickets a month, resolving 60% and saving ₹40 each, is worth about ₹9.75 lakh a month after ₹2.25 lakh of AI cost. Note that the AI cost applies to every ticket it attempts, not just the ones it solves. One more catch: if nobody leaves the team, you have freed capacity, not saved cash. Be honest about which one you're claiming.2. Is it suitable?The useful question to ask is "where should AI sit in this system?" Split the task into parts, then score each part from 1 to 5 on four dimensions:One right answer? If yes, use software. A scanned GST invoice is the perfect example: AI reads the messy layout, software calculates the tax. AI should never invent a tax figure.Can you write the rule? If yes, write it. If not but examples exist, let a model learn. Fraud teams do both: hard rules, plus a model's risk score, plus a decision.What if it's wrong? A bad email draft costs a minute. An unauthorised money transfer can't be undone. The harder the error is to reverse, the less AI should decide alone.Do we have the data? Data has to exist, fit this task, be good enough and be legal to use. Five years of machine sensor data with failure records is a strong candidate. Three weeks with no failures is not.Add up the scores:16 or more out of 20 is a strong candidate: a sales-email assistant scores about 18.Between 11 and 15, add guardrails.At 10 or less, AI supports and a person decides.Use the total to compare ideas. The answer is rarely yes or no.It's a role: AI can assist (it drafts, a person checks), recommend (it scores, a person decides) or execute (it acts, within set limits). Who should approve a ₹4,500 expense claim? Not the AI. AI reads the receipt, and rules approve it within caps. The pattern that falls out of all this is worth memorising: AI for ambiguity, software for certainty, humans for judgment.3. How do we build it?Pick the simplest route that still has lasting value, and stop at the first yes:Prompt if a good prompt solves it reliably and nothing needs system access.Buy if a product or API already does it and building wouldn't set you apart.Build if the workflow is strategic or your own data is the edge.Mix is where most real products end up: buy the base, build the integrations and rules.Take a WhatsApp bot that refunds up to ₹2,000 for missing food items. Buy the conversational platform. Build the order lookup, per-customer caps, fraud checks and refund rules. Send disputes to a person.Why not just prompt it? Because a prompt alone can't promise facts, policy or privacy. Ask a bare chatbot "where's my refund?" and it may invent a status. Ask it to delete your data and it may say "done" when nothing was deleted. Those guarantees live in the system around the model, not in the prompt.The model is not the moatOn 2 May 2023, Chegg's share price fell 48% in a day. Students wanted homework help, and ChatGPT gave it free, instantly. When a free AI can do your customer's job, features are no defence.Microsoft Copilot (30M+ paid seats) won on distribution: AI arrived inside Word and Outlook, so no new habit was needed.Harvey (used by 80% of Am Law 100 firms) won on workflow: a general model plus legal process, firm documents and lawyer review. GitHub Copilot works because code is quick to check, so the developer's job shifts to review.Put your own idea through itHere is the whole method as a one-page exercise. Pick one painful, high-volume problem from your work and answer six lines:Problem: who has it, and how often?Worth it? Does it pass TASQ and what is it roughly worth in ₹ a month?Suitable? What does it score on the four dimensions?AI's role: assist, recommend or execute?How? Prompt, buy, build, mix or no AI yet?Check: what must you find out before you commit?Klarna eventually found its balance. By Q3 2025, its CEO said the AI was doing the work of 853 agents, while promising customers that "there will always be a human if you want". That's the lesson in one line: let AI carry the volume and keep people where judgment matters.