How to respond to PM questions on on AI Economics, In-house vs frontier models and AI vs hiring humans
You can talk about “tell me about a product you shipped.” Something that is talk of the town now in AI is “our AI feature dropped gross margin twelve points, what do you do now?”
That second kind of question is taking over PM interviews at banks, insurers, and asset managers, because AI cost has moved from the IT budget to the boardroom.
Uber blew through its annual AI budget in four months.
Lindy, a 25-person startup, was spending more on its Anthropic API bill than on payroll before switching models and cutting costs 90%.
JPMorgan now spends $2 billion a year on AI and its CEO reports the spend “paid for itself.” This gap between two outcomes is exactly what interviewers are looking for from you.
Here are 20 questions across four themes, with what a strong answer covers. Use them to prepare if you’re the candidate, or steal them if you’re the hiring manager.
Theme 1: Token pricing fundamentals
1. The price per token keeps falling, yet our AI bill keeps rising. Explain that to a CFO in two minutes.
Strong answers should name the Jevons paradox. The price of a fixed level of AI capability has collapsed (Stanford’s AI Index documented a 280x drop for GPT-3.5-level performance in about two years), but cheaper intelligence gets used for more things, and the new things, agents especially, consume tokens at rates chat never did.
Ramp’s enterprise spending data showed per-token prices falling 75% in a year while total bills climbed anyway.
The candidate who says “so plan for both: unit costs fall, total spend rises as long as the product is succeeding” has given the CFO something to budget with.
2. Much of today’s API pricing is subsidized. How do you build a budget on a price that isn’t real?
OpenAI has reportedly spent around $1.35 for every dollar it earned, with losses driven largely by the cost of serving inference. Frontier providers are pricing below cost to win share, which means the number in your model is a promotional price.
A strong answer treats vendor repricing as a scenario to stress-test. what happens to the P&L if input prices double, and which workloads would you migrate first?
3. Why is “price per million tokens” the wrong unit for comparing providers?
Because tokens aren’t a universal cost unit. Providers bill text by tokens, speech by audio duration, and vision by image or feature unit, and even within text, tokenizer efficiency, prompt construction, and output length change the effective cost.
BCG’s benchmarking found cost leadership shifts by workload even when prompts and samples are held constant, so provider-level generalizations lead you to overpay. The right move is normalizing to business units: cost per 1,000 summaries, per 10 hours of audio, per resolved ticket. Interviewers are listening for “cost per outcome,” not “cheapest per token.”
4. Our agentic pilot costs 40x more per task than our chat feature. Walk me through why.
Four forces compound: breadth of adoption, task intensity, context and loops, and model mix. The killer is context: agents carry instructions, history, retrieved documents, and tool outputs forward and often resubmit them every loop, so billed tokens grow roughly with the square of session length.
A session that feels twice as long can cost four times as much. Current reference points make the point vividly: a chat conversation runs half a cent to three cents, while a single agentic task runs $0.50 to $5.00 on frontier models. If the roadmap includes agents, then that needs to be included in the P&L.
5. Which P&L line do tokens hit?
- Tokens that build reusable capability behave like capex.
- Tokens running internal work (engineers coding, analysts modeling) are opex, governed like a budget.
- Tokens consumed inside a customer-facing product are cost of goods sold, and they compress gross margin directly
AI-enabled software margins are resetting to the 65-80% range, and AI-native products to 50-65%, against the 80%+ that SaaS investors are used to. Burying all of it in a generic IT hosting bucket destroys the ability to see margin product by product. .
Theme 2: Tokenmaxxing is dead, governance is next
6. What was tokenmaxxing, and why did it die?
Early heavy adopters, mostly tech firms, treated token burn as a proxy for AI ambition and encouraged maximum consumption.
It died because the ROI didn’t show up; Fortune ran the obituary in May 2026. But the overcorrection is just as bad. Clamping hard caps on token spend starves the highest-return work and biases the org toward narrow, easy-to-measure labor substitution. Neither maximizing nor minimizing the bill is the job. Optimize the ratio between them
7. Define return on AI. Why does the denominator include human cost?
RoAI is economic return divided by the cost of human intelligence plus the cost of tokens. Human cost belongs in the denominator because in nearly every productive workflow, people initiate, steer, review, correct, and approve the work.
That framing blocks the tempting shortcut: cutting headcount to shrink the denominator usually shrinks the numerator faster.
Strong candidates will also note what the ratio rules out, since a workflow with a huge output volume and no accepted outcomes has a numerator of zero no matter how cheap the tokens were.
8. Your support agent closes 30% more tickets, and escalations are up 20%. Did it work?
No, and the question is testing whether you can tell output from outcome. Closed tickets that bounce back as escalations are motion, not return. The numerator should count outcomes that survive: tickets resolved without reopening, code that ships and survives review, campaign assets that actually get used.
The denominator should include the human time to initiate, review, and correct.
The same workflow can be strongly positive or negative depending on those two adjustments, which is why every high-consumption workflow needs a standing review with the authority to scale, rescope, or kill.
9. You’re asked to cut token spend 30% in a quarter without hurting outcomes. What levers do you pull, in order?
- Stop: pull deterministic work (lookups, rules-based routing, calculations) out of models entirely and give it back to software, which is cheaper and more reliable.
- Route: match task complexity to model tier instead of defaulting everything to frontier. A working cascade sends roughly 70% of traffic to a small open-weight model, 25% to mid-tier, 5% to frontier, at around 15% of the cost of routing everything to the best model.
- Cache: stable prompt prefixes (policies, compliance requirements, brand rules) cut 50-90% off repeated-context costs.
- Govern and train: assign every material workflow an owner, an outcome, and a cost-per-outcome baseline, and teach users to ask for the smallest useful answer.
Candidates who reach for “negotiate with the vendor” first are answering a procurement question, not a product question.
10. When is the cheap model the expensive choice?
When it fails. A weaker model that needs three retries to reach an acceptable answer costs more per successful outcome than a stronger model that gets there once, and that’s before counting the human time spent reviewing the failures.
This is the single most useful corrective to headline price comparisons, and it reframes the whole cost conversation around the only unit that matters: cost per successful outcome at the required quality, latency, and risk.
Theme 3: Chinese models vs. homegrown
11. DeepSeek V4 Flash costs $0.14 per million input tokens. Frontier models cost $2 to $3. Why isn’t switching an obvious call for a bank?
Banking sits in the category where the case for Western models is strongest.
IMD’s researchers writing in HBR conclude that applications demanding high accuracy, transparency, and regulatory defensibility, specifically pharma, banking, and government, fit better on Western models, while Chinese models win in retail, consumer goods, media, and routine tasks like summarization and basic coding.
Security researchers found weak encryption, potential injection flaws, and undisclosed data transmissions in DeepSeek’s R1, and S&P has warned that banks using generative AI in lending without thorough model validation face elevated credit risk.
12. Under what conditions could a US financial institution use a Chinese open-weight model at all?
About 15% of financial services firms already report using DeepSeek, per Cambridge’s 2026 Global AI in Financial Services Report.
Self-host the open weights inside your own perimeter so no data touches a Chinese API, run the model through the same validation and documentation you’d apply to any credit or risk model, restrict it to workloads where an error is cheap, and encode one rule in software rather than policy.
Regulated workload never fails over to an external provider. Availability never overrides sovereignty. “Never use Chinese models” and “switch everything to DeepSeek” are not so great answers.
13. What is the 3C framework, and what does it tell you about where Chinese models win?
Customization, cost leadership, and calibration.
Chinese providers build vertically integrated stacks tuned to local needs, design models with cost-efficiency as a first principle rather than an afterthought, and obsessively calibrate for deployment in messy real-world settings.
The strategic implication is the hybrid pattern already visible at multinationals: Nestlé found Western LLMs better for logistics but Chinese platforms better for customer analytics; BMW is putting DeepSeek in its China-market vehicles while running different stacks elsewhere. The framework question behind the question: can this candidate think in portfolios rather than picking a team?
14. When should a financial institution run its own models instead of renting?
Data sovereignty (prompts and outputs never leave the perimeter, which for regulated data is often the only compliant option), vendor risk (owned weights cannot be repriced, deprecated, or rate-limited), and fixed-cost economics that make always-on monitoring agents affordable when per-token pricing would kill them.
Open-weight models now sit within a few benchmark points of commercial APIs, which closed the quality objection. Then the counterweight, because interviewers distrust advocacy: below serious volume, self-hosting is usually more expensive once you count serving engineers, utilization risk, and the opportunity cost of running infrastructure instead of shipping product. The break-even conversation starts around ten million requests a month.
15. Cursor went from API wrapper to training its own frontier model. What does that ladder teach a bank’s product team?
Cursor bought first (validating the use case on rented frontier models while accumulating proprietary usage data), then fine-tuned an open-weight model on that data, and only trained from scratch once it had passed $2 billion in ARR and the volume math was trivial.
The transferable lesson is that build-vs-buy is a ladder you climb as data, volume, and strategy align, and that you should know your migration trigger before you hit it.
When inference became a fixed cost, previously impossible features (parallel agents, an always-on reviewer) became flagship products. Owning the economics changed what the product could be.
Ask yourself what an always-on fraud reviewer or portfolio monitor would look like if inference were free at the margin.
Theme 4: Strategy, pricing, and the CFO conversation
16. Price an AI-powered feature for a wealth management product.
Four workable patterns: bundle AI into higher tiers, sell usage allocations with transparent overage, price per outcome (per completed analysis, per generated report), or reserve it for enterprise deals entirely. For agentic features, outcome pricing is gaining because the client cares about the completed task and would be baffled by a token bill.
The risk is that you absorb cost variability, including retries and failures, so you need real confidence in your cost structure first.
17. Your most loved AI feature is bleeding money. Adoption is 70%, NPS is great, and gross margin fell from 82% to 70%. What do you do?
First, diagnose before touching anything. Usage variability in AI products is extreme, and the profitability spread between a casual user and a power user at the same price point can be 25x.
Your heaviest users are usually your loudest advocates, so blunt throttling has a real cost.
Cascade routing so cheap requests hit cheap models, prompt caching, usage tiers that convert the power users into your best customers instead of your worst, and honest repricing if the unit economics still don’t close. The candidate should say “product-market fit is not business-model fit.”
18. What goes on the one-page AI business case you’d hand a CFO?
Cost per user per month, and better, cost per successful outcome, with the trajectory at scale. Gross margin impact and where it stabilizes. Break-even usage level and the timeline to reach it. Cost variability exposure: what happens if usage doubles, and what happens if the vendor reprices. And competitive necessity: what it costs to not invest.
19. Is the AI capex boom a bubble, and why should a product manager care?
JPMorgan’s analysis puts the revenue needed to justify current infrastructure spend at roughly $650 billion a year against industry revenues near $75 billion, while hyperscaler capex runs past $700 billion. That’s the shape of every capital cycle since railways.
But the PM-relevant part is the two scenarios and their opposite implications for your token strategy: if the bubble pops, capex slows and cheap subsidized inference may persist;
If providers race to justify valuations instead, token prices rise and today’s economics break. A strong candidate designs for both, keeping workloads portable and knowing the switching cost per workflow in advance.
20. What does good look like at the end state?
JPMorgan. Roughly $2 billion a year of AI spend inside a $19.8 billion tech budget, described by its CEO as having paid for itself, with function-level receipts.
Engineers about 10% more efficient, per-unit fraud costs down 11%, research time on automated tasks down over 80%, 400+ use cases in production, and an internal LLM platform on the bank’s own data behind its own controls.
Two details worth stealing: the bank reclassified AI from discretionary innovation to core infrastructure, and the $2 billion payoff sits on top of a fourteen-year capability build with about 2,000 AI specialists.
Budget your credibility accordingly. Dimon expresses AI in CFO language, dollars in and dollars out by function, and admits openly that measurement is hard, which is exactly why his numbers are believed.
The thread through all twenty
Every question above reduces to the same discipline: measure cost per successful outcome, not cost per token, and manage AI spend as a portfolio rather than a line item.
The candidates who get offers are the ones who can hold two ideas at once, in every section: prices are collapsing and bills are exploding, Chinese models are dangerous and useful, the cheap model is sometimes the expensive one, and the best AI feature you’ve ever shipped might be the one eating your margin.
If this was useful, share it with someone prepping for a PM interview. And if you’re the interviewer: ask question 17. It has the best signal-to-word ratio of anything on this list.
Sources:
BCG’s two-part Return on AI series and Cloud Cover pricing analysis
“How Savvy Companies Are Using Chinese AI,” Harvard Business Review, Sept-Oct 2025
Cambridge Centre for Alternative Finance, 2026 Global AI in Financial Services Report
American Banker and S&P Global on DeepSeek risk; Risk.net on banks and DIY AI
Fortune on the end of tokenmaxxing;
DeepSeek published API pricing, July 2026;
Bloomberg interview with Jamie Dimon;
reporting on Uber, Lindy, and Cursor from CNBC and company disclosures.
Upcoming book: Product Manager to Product Leader in the AI economy by Anil Jaising and Bernie Maloney
