you're using chatgpt to write prds.
v0 or replit for prototypes
claude for competitive analysis.
does it make you ai ready?
that's like saying
i can use excel so i am ready to be a cfo.
i am watching this play out.
pms with good experience at tier 1 companies. they use ai tools daily for everything. specs, analysis, user research synthesis. they apply to 15 ai roles. get zero callbacks. of course.
and then there young folks
who are 22-25. built one focused ai proof of work over 10 days. solved a specific token optimization problem. three offers, including one from an ai first company.
the difference?
one was using ai.
the other was building with ai.
confused about this distinction?
using ai tools for your daily work isn't proof of ai capability. it's proof you can follow a tutorial. real proof means solving business problems with ai systems. building products where ai is the core value prop. understanding when models fail and designing around it.
but here's where it gets messier.
there are two completely different games being played.
internet first vs ai first
1/ internet first companies
zomato, razorpay, cred, swiggy. these companies existed before chatgpt. ai augments their existing products. makes them faster, smarter, more efficient.
when zomato adds ai powered food recommendations, they're enhancing an existing loop. the core business model hasn't changed. they still make money on delivery fees, ads and commissions.
their pms need to show they can use ai to 10x existing metrics. not create new ones.
2/ ai first companies
sarvam ai, krutrim, cursor, lovable, bolt etc. ai isn't a feature. it IS the product. no ai, no company.
when an ai company sells conversational ai, the model quality directly determines revenue. token costs impact gross margins. latency affects user retention.
their pms need to understand transformer architectures. not to code them. but to make product decisions around them.
most pms build the wrong proof
they show a chatbot for an internet first company. or efficiency improvements for an ai first company. wrong game, wrong proof.
let me show you how.
follow this structure. it’s not easy, but it works. also, majority of you want to crack a role in internet first company. i’ll share how i would do that as well as cracking a ai pm role in a internet first company.
building proof that changes orbits
7 steps. let’s go.
1/ pick your targets
choose 2/3 companies.
real ones. with real problems.
for internet first:
look at established players with clear metrics
zomato (food delivery)
razorpay (payments)
cred (credit/rewards)
for ai first:
emerging ai companies still figuring out their equations
sarvam ai
coderabbit
bolt
lovable.
pick 1 or 2.
2/ decode their growth equation
the ultimate truth.
every business has an equation. even the fuzzy ones. this is where most pms get lazy. they think in vague terms like "improve user experience" or "increase engagement."
fuck that. get specific.
i will show you both ai first example and internet first. let’s start with internet first example as it’s more familiar.
example: zomato
let's break this down properly.
north star:
revenue = orders × aov × take rate
breakdown orders.
↳ orders = new users × activation + existing users × frequency
breakdown every input lever
↳ ↳ new users =
impressions × install rate × signup rate
↳ ↳ activation =
signups × first order rate
↳ ↳ existing users =
total users × monthly active %
↳ ↳ frequency =
app opens × browse to order conversion
go even deeper.
↳↳↳ browse to order =
search × results shown × click rate × cart addition × checkout completion
i can go further,
but you don’t need to.
each of these has sub levers:
↳↳↳↳ search effectiveness = query understanding × result relevance
↳↳↳↳ cart addition = item appeal × price perception × reviews trust
↳↳↳↳ checkout completion = payment success × delivery confidence
see how deep this goes?
this is what i mean by “depth.”
a term growthx members are all too familiar with. depth is what separates a normal proof of work to something that opens doors for you.
now pick ONE lever.
maybe it's browse to order conversion. currently at 12%. industry standard is 15%. that's your gap.
ai first company:
ai led support
north star:
revenue = customers × acv
breakdown each lever
↳ customers =
trials × conversion rate
↳ acv =
seats × price per seat × utilization
level 2 breakdown
↳ ↳ utilization
= conversations × resolution rate
↳ ↳ conversations
= total chats × containment rate
↳ ↳ resolution rate
= successful resolutions / total conversations
but what if this ai company is trying to optimise margins. why? because this is not a foundational ai company. it’s the application layer. here a large cost is still going to the underlying model they are using.
gross margin =
revenue - (token costs + infra costs + human costs)
breakdown token cost
↳ token costs =
conversations × tokens per conversation × cost per token
break it down further
↳ ↳ tokens per conversation
= (input tokens + output tokens) × conversation turns
different lever here.
maybe it's tokens per conversation. currently averaging 4,000 tokens. each token costs money. reduce this by 40% without hurting resolution rate? that's pure margin.
the project here can be a ui improvement, an alternative way to converse or something else entirely.
3/ talk to actual users
not your assumptions.
not case studies.
real humans who feel the pain.
linkedin approach:
search: company name + "product manager" or "marketing head" or "operations"
filter: posted in last month (shows they're active)
message: "working on something that might help with [specific problem]. worth a 15 min call? coffee/lunch on me."
twitter approach:
search: "zomato sucks" or "yellow.ai broken" or "[company] problem"
reply: "felt this pain myself. building something to fix it. can i show you a early version?"
for internet first users,
depending on your lever sharing some example questions.
problem discovery:
"walk me through your last order. where did you get stuck?"
"if you could fix one thing about [app], what breaks your flow?"
"show me your phone. how many food apps do you have? why?"
solution validation:
"what would need to change for you to order 2x more?"
"what makes you abandon an order?"
"if ordering took 30 seconds flat, what changes?"
for ai first products.
trust and reliability:
"when do you stop trusting the ai and take over?"
"show me the last time the ai failed you"
"what tasks do you never give to the ai?"
value perception:
"what would the ai need to do for you to pay 2x?"
"when does the ai save you the most time?"
"what if the ai could do X? how much is that worth?"
document everything.
record if they let you. patterns emerge after 5 conversations. by 10, you know exactly what to build.
4/ understand market
study the market they operate in.
what are the constraints? what's changing? where are the opportunities everyone else missed?
analyze their competition.
not feature comparisons. understand how competitors solve the same customer problems differently. find the gaps.
5/ build something real
not a deck.
not a notion doc.
build something real.
internet first proof example
you picked zomato's browse to order conversion. and say you want to build an ai powered food mood matcher
your features might include
analyzes order history + time + weather. the product predicts what user wants before they browse. we show them 3 options max and there is one tap ordering. we can measure: browse time reduction, conversion lift
ai first proof example
maybe you wanted to crack the cost of tokens per conversation for ai support product. and say you built dynamic context compression. so basically we compresses historical chat without losing accuracy
2 tips for building:
speed over perfection
hardcode what you can
it’s okay to fake the data pipeline
focus on core experience
polish comes later
show the process
capture decision moments
document tradeoffs
show iterations based on feedback
6/ test with users
if you can do this.
then i can guarentee your proof of work becomes a proof of value.
showing > asking.
always.
ask the same 5 users from research phase. get 20 minute sessions max. maybe record everything (with permission). capture their feedback. ensure you’re observing without guiding. measure the time saved, delight moments, confusion points.
document what matters
exact quotes, moment of delight, also the confusion points (where they hesitate) and feature requests (what's missing)
7/ package and send
one page. that's it.
problem → solution → user feedback → impact.
email structure:
subject: reduced [metric] by X% for [company]
hi [name],
built something over the weekend that solves [specific problem].
the issue: [one line problem statement with number]
what i built: [one line solution description]
early results: [key metric improvement]
5 users tested it. 4 would pay for it. one called it “magic.”
demo video (2 min): [loom link]
try it yourself: [live link]
worth a 15 minute call?
[your name]
p.s. - [one interesting learning from building this]
finding the right person.
but not hr. never the hr at this stage.
for internet first:
head of product, growth pm, ceo (if under 100 people)
for ai first:
founding team, head of ai, product lead
please avoid
generic chatbot wrappers
everyone can call openai’s api. showing another customer service bot isn’t proof. it’s homework.
over engineered solutions
model fine tuning for a simple classification problem. rag for a basic faq. kubernetes for a weekend project. choosing complex tech to look smart is the fastest way to look amateur.
missing the business
“built with langchain and pinecone” means nothing.
“can improve business metric by XX%” means everything.
nuances by company type
internet first companies
internal transitions are real. you’re competing with engineers or technial pms moving to ai pm. they can already build. plus current pms are also very excited to build in ai right now. if you build the right proof of work, you can actually break in.
but the nuance with ai first is a lot more. let’s dig in.
ai first companies
i looked job descriptions from cursor, lovable, replit, openai, anthropic, harvey. then looked at questions asked in those interview rounds.
some things you should know.
1/ attention mechanisms.
token limits. fine tuning vs few shot vs rag. context windows. embedding spaces. not to test engineering skills, but to see if you understand the constraints you’re designing within.
example question:
“user complains the ai forgot something from earlier in the conversation. what might be happening?”
good answer:
“context window limit. need to implement sliding window summarization or selective memory.”
2/ failure mode thinking
“what happens when the model hallucinates?”, “how do you handle variable latency?”, “what’s your graceful degradation strategy?” they want to see you’ve thought about when ai breaks. because it will.
3/ cost consciousness:
every product decision has a token cost. example: “we need to add conversation memory. how do you implement it?”. bad answer: “store everything, retrieve everything”. good answer: “semantic compression, retrieve only relevant context, approximate with embeddings when possible”
4. interview process differences
round 1: product sense with ai constraints. standard product questions but you must factor in ai limitations. “design a feature” becomes “design a feature where the ai might be wrong 10% of the time.” they want to see how you think about failure and not just success.
round 2: technical depth on ml/ai concepts. no coding, but deep understanding.
round 3: live product critique. they’ll show you their actual product. ask what’s broken. what’s expensive. what’s limiting growth. they want to see if you can spot ai specific issues. they want you to think in probabilities, not certainties
the uncomfortable truth
ai is not changing pm roles.
it’s splitting them.
type 1: ai architects
designing systems where models, data, and users create compounding value. these pms think in feedback loops, data flywheel effects, model marketplace dynamics. they’ll work at ai first companies or lead ai transformation at internet first ones.
type 2: automation specialists
making existing products 10x better with ai augmentation. these pms excel at finding leverage points, building practical solutions, measuring real impact. they’ll thrive at internet first companies looking to stay competitive.
type 3: the obsolete
using chatgpt to write specs and thinking they are ai ready. writing requirements for features they don’t understand. managing timelines for projects they can’t evaluate.
they’ll be automated, eventually.
harsh? maybe. true? yes.
your move
pick two companies.
understand their equations. map out every variable. build something that moves one lever. just one.ship it in the next 10 days.
in 6 months,
having “used chatgpt for product specs” on your resume will be like listing “proficient in microsoft word.”
but shipped something
that solved a real business problem with ai? that reduced costs by 40%? that opened a new revenue stream?
this is the proof of work
that changes orbits.