Working in AI

Offer A was from a fourteen-person edtech company in Pune. The title was AI Engineer. The money was ₹6,50,000 CTC. The founder had spent twenty minutes of the interview describing what they would build with language models next year.

Offer B was from the Pune office of a Dutch insurer: a global capability centre with about nine hundred people, a category Sneha had not known existed six weeks earlier. The title was MIS Executive. The money was ₹5,80,000 CTC.

Her mother had one question about each, and it was the same question: which one has AI in it.

What the adverts said when read backwards

Sneha had learned to read the tools list first. The startup’s requirements named a hosted model API, Google Sheets, Zapier and “prompt libraries”, and the responsibilities section ran to fourteen bullets covering dashboards, marketing support, customer research and “AI initiatives”. The insurer’s advert named SQL, dbt, Snowflake and Airflow, listed four responsibilities that described one coherent activity, gave the manager’s name, and described the first ninety days.

She also noticed something she nearly did not mention to anybody: the startup’s advert had been reposted every month since April.

The question she asked in both interviews

Where does the number on your main dashboard come from, end to end?

The founder said it depended who you asked, and laughed, and said that was exactly the problem she would be hired to fix. The insurer’s team lead answered in four sentences: the policy administration system lands in Snowflake by 04:00, twenty-six dbt models run against it, forty-one tests run after that, and if a test fails the dashboard shows yesterday’s data with a banner rather than a wrong number.

One of those answers describes rung two of the maturity ladder. The other describes rung three or four.

The cost on both sides

Taking the startup meant twelve per cent more money, a title that would look serious on a CV, and touching a model API in her first month. It also meant being the only technical person on data, with nobody to review her work, in a company where the first year would be spreadsheets and one integration, and where the advert being reposted monthly suggested she would not be the first person to discover that.

Taking the insurer meant a title that sounded like a clerk’s, no model anywhere near her desk for at least a year, and the slow, procedural culture of a large European employer. It also meant a named manager, a team of six, a reviewer for everything she wrote, and a system she could learn deeply.

The startup wanted an answer within twenty-four hours. The insurer had given her a week.

What she was actually choosing between

Not two jobs. Two first years. One of them would produce a person who had spent twelve months doing four things badly with nobody to correct her, holding a title that would cause a large employer to down-level her at the next application anyway. The other would produce a person who could say, in an interview, exactly where a number came from and what happens when the pipeline fails.

What she could not know was whether the startup would be the exception: the small, chaotic place where somebody learns five years of work in two.

The brief was to reduce failed deliveries. Around eleven per cent of consignments needed a second attempt, each one costing about ₹180 in fuel and driver time, and the operations director wanted to know which deliveries would fail before the van left the depot.

Week one produced no model at all. It produced the discovery that the delivery_status field had only been populated consistently since the previous April, and that before that the depots had used a free-text note. Week two produced a baseline: predict each customer’s historical failure rate. It got about seventy-nine per cent of the ranking right and took an afternoon.

Week three produced the model everybody wanted. Gradient-boosted trees over thirty-one features, and an offline score of 0.94.

Nobody on the team had asked for a neural network and nobody needed one. This is ordinary tabular business data (a few hundred thousand rows, thirty columns, no images and no text) and gradient-boosted trees remain the correct tool for it, which surprises people who have spent a year reading about model architectures.

The question asked at the feature review

A senior engineer went through the feature list and asked one question about each: at the moment we make this prediction, does this value exist yet?

Thirty features survived. One did not. resolution_code was filled in by the depot clerk after a delivery attempt, describing how the attempt had been resolved. In the training table it sat innocently beside the target. At prediction time (05:30, before any van has moved) it is empty.

Removing it took the score from 0.94 to 0.81. Three weeks of work, and the number went backwards in front of a board demonstration that was already in the calendar.

The decision

Demonstrating 0.94 was available. The leaked column would not be visible in the slide, the number was real in the sense that the notebook had computed it, and there was a defensible version of the argument: the demo is about the idea, not the deployment.

The cost of that path was not abstract. The operations director would take the number to the board, the board would fund a rollout, and in production the model would score somewhere near 0.61, at which point the failure would be discovered by the depot supervisors, who are the people whose trust the system needs and who have watched three previous head-office systems arrive and be quietly abandoned.

The cost of the honest path was three weeks of visible progress turning into a lower number, in a company where nobody on the board understood what leakage was and where “it went down” is not an easy sentence to say in a boardroom.

The second problem, which was larger

While arguing about the first, somebody asked what happened to a prediction after it was emitted. The answer: the depot could physically re-check about a hundred and twenty consignments a day out of roughly four thousand. Whatever the model produced, exactly a hundred and twenty deliveries would be looked at.

Which meant the metric everybody had been optimising (overall ranking quality across four thousand rows) was not the metric that mattered. What mattered was how many genuine failures appeared in the top hundred and twenty.

That is the second failure that ends a first year in this role, and it is quieter than leakage: improving a score for six weeks while the business outcome does not move, because nobody followed the prediction to the person who acts on it. A model whose output lands in a queue with a fixed capacity is not being asked to rank four thousand things. It is being asked to fill a hundred and twenty slots well.

What the three weeks were actually for

The engineer’s own summary afterwards was that roughly one week of the three had been modelling. The rest was establishing what the target field meant before April, arguing about whether a re-attempt on the same day counted as a failure, writing the feature computation so that the training path and the serving path used the same function, and finding out what happened to a prediction after it was emitted.

That proportion is not a sign of a badly run project. It is the job. Model training is somewhere between five and fifteen per cent of the effort in a deployed system, and a candidate who describes their week the way the internet describes this role has usually not deployed anything.

His aunt runs the reception desk at a clinic near Arera Colony. Fourteen months of appointments live in a spreadsheet called appointments_master_final.xlsx, maintained by hand, and about a fifth of booked patients do not arrive. The doctors lose the slot; the patients who could have had it wait another week.

It is a genuinely useful problem and Farhan had the data on his laptop within a day. Then three separate questions arrived, in the wrong order.

The mess, which was the interesting part

The first thing the file taught him had nothing to do with prediction. Three months in the middle of the fourteen used day-first dates and the rest used month-first, with nothing marking the change. He found it because Tuesday appointments spiked implausibly in exactly those three months, which is what a day-month swap looks like when you plot it.

He had two defensible options and had to choose one. Re-parse those months with the other convention and keep the rows, accepting that any row where both numbers were twelve or lower could still be wrong and he would never know which. Or drop the three months entirely and lose a fifth of his data. He dropped them, and wrote down why, because inventing a correction would have hidden the fact that he could not tell.

The data question

Patient names, mobile numbers, ages, and a free-text reason column that in about four hundred rows contains a diagnosis. Nothing about that file is his. It belongs to the clinic, it describes identifiable people, and India’s data protection regime does not contain an exemption for “my aunt gave it to me”.

A blurred screenshot would not have helped either: the underlying file would still have been in the repository.

The measurement question

He built forty test cases by hand from the last two months, wrote down for each what a correct prediction would be, and scored his model: eighty-six per cent.

Then he did the thing almost nobody does and scored the dumb version on the same forty cases. Predicting that every patient shows up (a rule requiring no model, no Python and no laptop) was right eighty-two per cent of the time, because most people do turn up.

So the honest headline was four points above a constant. Reporting eighty-six per cent alone would not have been a lie. It would have let every reader assume the baseline was zero.

The decision

Three routes were open, and each cost something.

Publish the real analysis. Strongest evidence: real mess, real decisions, a real user. Also a breach of the clinic’s trust, a probable breach of the law, and a reviewer who spots patient data in a public repository will assume he would do the same with theirs.

Publish with synthetic data. Safe, and honest if labelled. Also much weaker, because synthetic data demonstrates that a pipeline runs and demonstrates nothing about a finding. He would have a project that proved he could write code and proved nothing about judgement.

Ask, in writing, and publish what is permitted. Slowest, and it required a conversation with a doctor who had not asked for any of this and who might reasonably say no, in which case he would have spent three weeks and had nothing.

He also had to decide whether to report the eighty-two per cent baseline at all, three weeks before applications opened. Nobody would have known it existed. He had computed it himself, in an afternoon, for his own benefit, and deleting one line from a page would have turned a four-point improvement into an eighty-six per cent result.

The reviewer he was actually writing for

He kept forgetting who the audience was. A reviewer gives a portfolio project between one and two minutes: the title, the first sentence, whatever number is visible, and (if the page has earned it) the section admitting what is wrong. Almost nobody runs the code, and nobody at all reads a notebook of forty cells.

So the page had to carry the mess, two decisions with the option he rejected for each, one number, and the limitations, in that order and above the fold. The model itself was fifteen lines of scikit-learn and the least interesting thing on the page.

The spreadsheet is the only reason this case has numbers in it. She had started it in month two, on the advice of someone in a Discord server, and it had seven columns.

After eleven months:

• Applications sent: 341
• First conversations: 6
• Technical stages: 2
• Final stages: 0
• Offers: 0

Six conversations from three hundred and forty-one applications is a rate of one point eight per cent. In the same eleven months she had finished two courses, built three projects and learned dbt.

What the funnel said

The leak was at the very top, and everything she had spent the year improving sat below it. Her SQL was not the reason nobody was calling; nobody was getting far enough to find out about her SQL. A funnel that breaks before the first conversation is a statement about targeting and route, not about ability, and it was the one diagnosis she had spent eleven months not making, because improving her skills felt productive and changing her method felt like admitting the year had been wasted.

Run the arithmetic forwards and it is worse. To reach two offers she needed roughly seven finals, seventeen technicals and thirty-four first conversations. At one point eight per cent that is close to nineteen hundred applications. At a warm-approach rate of thirty per cent it is a hundred and thirteen approaches.

Those two numbers describe the same search. One of them is possible for a person with rent to pay and a life to lead; the other is not, and no amount of stamina closes the gap, because the gap is division rather than effort.

The part of the funnel she could not see from inside

The tracker also had a column she had stopped filling in: why she wanted each role. Reading twenty rows back, most of them said some version of it was open. That is a different failure from a leaking funnel and it does not show up in a conversion rate. A search that has drifted into applying to whatever exists produces applications that are visibly generic to the person reading them, which lowers the rate further, which produces more applications.

Lagos also has a specific shape she had not adjusted for. A large share of the good roles are remote contracts with foreign companies, which means her competition is global and her evidence has to be public, and the hiring channel that actually works there is community rather than job board.

The decision, and what it cost

The proposal from a friend in the same server was blunt: stop applying for six weeks. Spend the same eight hours a week writing to twenty named people who do the work, with something specific attached, and publishing one write-up a month.

The costs were real and none of them were about strategy.

Sending applications is visible effort. Her uncle asked every Sunday how many she had sent, and “none, I wrote four emails” is a difficult sentence in a house where money is tight. Approaches also fail silently and slowly: an application produces a rejection email, which is at least an event, while a message to a stranger produces nothing at all for a fortnight.

And she had a third option she kept not taking: a customer support role at a payments company, twenty per cent below what she wanted, sitting next to the transaction data. Taking it would have meant admitting, out loud and to her family, that eleven months of study had ended in a support job, while being, on the evidence of how most working analysts actually got in, the single highest-probability route available to her.

The brief was ordinary: find the drivers of chargebacks in the attached extract, produce a short write-up, five hours. The zip file was not ordinary. It contained merchant names, partial card numbers, timestamps, and a free-text complaint_note column in which about nine hundred rows contained a customer’s name and, in a few dozen, a phone number.

The brief said nothing at all about using AI.

The tempting ten minutes

Classifying nine hundred free-text notes into complaint categories is the kind of task a hosted model does well and a person does slowly. Pasting the column into a chat window would have taken ten minutes and produced a usable classification. Doing it by hand would take most of an evening he did not have, because he was doing this after a full shift.

He asked the three questions in order and the first one settled it. Is this data mine to share? No. It is not even the hiring company’s to share carelessly; it is their client’s. His own CV would have been fine to paste. This was not that.

The second and third questions (what the provider does with it, and whether he would mind seeing it in somebody else’s output) did not need answering, because the first had already decided the matter. That order is the point: most of the trouble in this area comes from skipping straight to the provider’s retention policy.

He checked the settings on his own account anyway, which took two minutes and which he had never done in a year of using it. The training toggle was on by default, as it usually is on a consumer tier. That would have mattered for his own CV. For somebody else’s customer file it was beside the point: “we do not train on your data” is not “we do not keep it”, retention for abuse monitoring is standard, and the disclosure has already happened at the moment of the paste.

The decision

Paste it. Ten minutes, a better classification than he could produce by hand, and a submission that would almost certainly never be traced. The cost is that a disclosure of a client’s customer data to a third-party service is a genuine breach rather than a bad look, and he would be starting a relationship with a company by doing to their client exactly what they would fire him for later.

Do it locally. A small open-weight model through Ollama on a laptop with sixteen gigabytes of memory. Free, private, and meaningfully worse: slower, weaker on long text, and an hour of setup out of a five-hour budget.

Skip the free-text column entirely and write in the README that he had left it out. Honest, fast, and a visibly smaller piece of work than the other candidates would submit.

He had four days, one of which was a Sunday, and a real chance that the extra hours would produce a worse submission than the shortcut.

The thing he could not resolve by choosing well

There was a fourth question underneath the other three, and it was about the company rather than about him. A take-home that ships a real client’s identifiable data to unvetted candidates is a data-handling failure by the employer, whatever the candidate then does with it. He was being asked to be careful with material that had already been handled carelessly.

He decided that was worth raising, once, in the submission rather than in an email: a single sentence noting that the extract contained customer names and numbers, and describing how he had contained it. Raising it as an accusation before an offer exists costs a candidate the process; raising it as a description of what he did costs nothing and is checkable.

The offer came by telephone at 16:40 on a Thursday: analytics consultancy, sixty people, ₹9,00,000 base, joining in three weeks. The recruiter added a sentence at the end. We need your answer by ten tomorrow morning: the budget is being reallocated on Friday.

The other process was a global capability centre. He had reached the final round, scheduled for Tuesday. At their recruiter screen four weeks earlier he had asked for the band and been told ₹9,50,000 to ₹12,00,000. He had not asked which level the role sat at, which turned out to matter.

How he had got here

The consultancy’s process had run over five weeks: a recruiter screen, a hiring manager call, a four-hour take-home he had time-boxed to four and a half and said so in the README, and a deep dive on the take-home in which he had been asked why he chose a left join three separate times. He had answered the third one badly and had emailed the following morning saying he had thought about it and got it wrong, with the version he would write now.

He assumed that email had cost him. It had not: panel notes travel forward, the interviewer in the later round had already read the earlier one, and revisiting your own answer reads as strength rather than as weakness precisely because the note is already there.

What was actually being decided

Not whether to accept. Whether to ask for time, and everything about that felt asymmetric at 16:40 on a Thursday. He had one offer. He had watched a friend push back on a deadline in March and be told the position had gone to somebody else, which may or may not have been true and had been extremely persuasive anyway.

Accepting on the call was available and safe. The cost was that the moment he said yes, the conversation was over: no negotiation, no start date, no six-month review, and a Tuesday final round he would have to cancel while wondering for a year what the band would have produced.

Asking for time risked the offer. It also tested it. A deadline that survives a polite request for the weekend is a real constraint; one that evaporates within an hour was manufactured, and a company that manufactures pressure at the offer stage does it at every subsequent stage too.

The second thing he had not asked

The consultancy’s title was Senior Analyst. Three years of experience, sixty-person firm, and a title that was free to give. He had no idea what level that mapped to internally, or whether “senior” there meant anything a larger employer would recognise later.

And he had no competing offer, so any negotiation would be one sentence with nothing behind it.

The list he wrote on the Thursday evening

Before replying to anything he wrote down what was actually negotiable, because the base salary is only the loudest item and often the stickiest:

• The start date, which was worth real money against a notice period.
• A joining amount, which in most companies comes from a different budget than salary and is therefore easier to say yes to.
• A written salary review at six months with a defined band, rather than a vague promise about the annual cycle.
• The title, which costs the employer nothing today and shapes his next application.
• Whether the two remote days were written into the contract or existed by custom.

He also wrote down the number he would genuinely accept, because a range you quote is a range you may be offered the bottom of, and the bottom of it has to be a number you can live with.

She had thought of the grant as “ten thousand shares” for four years. Nobody had lied to her. The document said everything, and she read it properly for the first time in the week she resigned.

The numbers:

• Vested options: 10,000, fully vested after four years with a one-year cliff.
• Strike price: ₹250 per share.
• Most recent valuation used for tax purposes: ₹410 per share.
• Post-termination exercise window: 90 days.

So the paper gain was ₹16,00,000 and the cash required to realise any of it was ₹25,00,000. In India an ESOP is taxed as a perquisite at exercise, on the difference between fair market value and strike, so on top of the ₹25 lakh she would owe tax on ₹16 lakh of gain, in a year in which she could not sell a single share, because the company is private and there is no buyer.

She had ₹6,00,000 saved.

What she thought she had been paid

Four years earlier she had accepted ₹4,00,000 a year less than a competing offer from a bank, on the reasoning that the options made up the difference. That was the decision, and it was made before she knew what a strike price was.

The arithmetic of an option is not complicated. It is worth the current share value minus the strike, multiplied by the number of options, which means a grant issued at a strike equal to the then-current fair value is worth exactly zero on the day it is granted, and is worth something only if the value rises. Her grant had risen from ₹250 to ₹410 over four years, which sounds like a good outcome and, once the exercise cost and the tax are placed beside it, is not obviously one.

There had also been two funding rounds in those four years. Her number of options never changed and her percentage roughly halved, because the denominator grew. Nobody had hidden this; dilution is what a funding round is.

The questions she asked too late

Four of them, and she asked them in week one of her notice period rather than at the offer stage four years earlier.

1. How many shares are outstanding on a fully diluted basis? Answered.
2. What is the strike and the latest valuation per share? Answered.
3. What is the total liquidation preference, and is any of it participating? Declined, on confidentiality grounds.
4. How long do I have to exercise after leaving? Ninety days, as written.

The third refusal was the important one. Without the preference stack she could not value the grant at all. Investors are paid before common shareholders, so a company that has raised heavily can sell for a headline number that leaves nothing for employees, and the press release will still say a large figure.

The decision

Borrow and exercise everything. ₹25 lakh of borrowed money and a tax bill, converted into an illiquid holding in a private company whose preference stack she was not permitted to see, with no timeline to any sale.

Let all ten thousand lapse. Four years of vesting worth nothing, and the certainty of watching a headline in three years and doing the arithmetic she chose not to do.

Something in between, sized to what she could lose without it changing her life.

She asked for the window to be extended: some plans allow seven or ten years, and asking is normal. The company declined, politely, and the fact that it declined told her something about how it thought about people who leave.

She also, for the first time, ran the comparison she should have run four years earlier: guaranteed annual cash against guaranteed annual cash, with the equity entered at zero and the upside noted separately in a line of its own. On that arithmetic the bank had offered more every year for four years, and the difference compounded through every raise she had received since, because raises are percentages of a base.

The engagement was ₹4,20,000 across three phases: thirty per cent on start, forty on delivery of the build, thirty on acceptance. The deposit had been paid on time, in eleven days, by his contact’s own corporate card because “finance is slow”.

That sentence was the warning and he did not hear it.

Phase two was delivered on 14 March. Invoice 0042, for ₹1,68,000, went out the same day. On 14 May it had not been paid, his contact was answering with “it’s with finance”, and Vinod had already done most of phase three.

What was actually wrong

He telephoned on day fourteen of the chase and asked one question (on what date will this be paid) and got the answer that mattered. There was no purchase order. In the retailer’s procurement software an invoice without a matching PO does not get rejected; it sits in an unmatched queue until somebody notices. His contact assumed it had been paid. Vinod assumed he was being ignored.

Worse: he had never been onboarded as a vendor at all. That process (a supplier form, bank verification, a tax declaration and proof of professional indemnity insurance) takes two to six weeks at a company that size and is nobody’s priority. It could have been started on day one of the project. It was started on day sixty-two.

The cash-flow arithmetic underneath it

He was profitable on paper and could not pay his own bills. Work delivered on 14 March, invoiced the same day on thirty-day terms, would have been paid in mid-April by a prompt client. This client was not prompt and was not refusing; it was simply a two-thousand-person company with a procurement department, and those pay in six to ten weeks as a matter of course rather than as a matter of intent.

That is the difference between profit and cash flow, and it is why the reserve exists. He had two months of expenses rather than the three to six the arithmetic asks for, which is what turned an administrative delay into a decision made under pressure.

The squeeze

Meanwhile his contact asked, pleasantly, for repository access and the warehouse credentials, because a new analyst was starting on Monday and “it would be good for her to get going”.

That was the decision.

Hand over. The client is the largest name on his site, a retainer had been mentioned twice, and refusing over money would be remembered. Once the repository, the credentials and the deployment are transferred, his only remaining instruments are a contract and the law, and both are slow enough to be theoretical for ₹2,94,000.

Hold the handover and pause the work. Protects the ₹1,68,000 outstanding and the ₹1,26,000 to come. Risks the retainer, the reference, and a reputation among the four people at that company who would recommend him.

There was also a smaller ledger he had been quietly absorbing: three “while you’re in there, could you also” requests, worth about a week between them, none of which he had priced or logged. Each had been small and reasonable on the day it arrived, which is how scope creep always arrives; nobody had demanded anything. The proposal had said what was included and had not said clearly enough what was not, and it did not number the revision rounds.

He was also aware, uncomfortably, that the client had not behaved badly at any point. His contact had chased finance twice. Nothing here required a villain.

Default rates had risen for two quarters. On Thursday afternoon the head of product asked for three features to be added to the credit model before Monday’s demonstration: the applicant’s postcode, the price band of the handset they applied from, and the applicant’s first name, which somebody had noticed “seems to carry signal”.

Nobody in the room was being malicious. Everybody was under pressure and the request took eleven seconds to make.

This is the shape these requests actually take. Nobody asks an analyst to build something obviously wrong. They ask, pleasantly, under a deadline, for something that sounds reasonable when said quickly, and the person being asked has to produce an objection in real time, in front of people they like, with no preparation. That is why the advice is to decide the line before the meeting, because the meeting is not where anybody thinks clearly.

Why it was hard, and it was hard

She knew that postcode in Nairobi is close to a proxy for both income and ethnic group; that handset price band is a proxy for income directly; and that a first name in Kenya carries ethnic information as plainly as it does anywhere in the world. She also knew that the system declines people with no explanation and no appeal route, and that nobody had ever measured the decline rate by ward.

Against that: she had been there two years, she had no savings to speak of, the demonstration was Monday, and refusing on Thursday makes you the person who blocked an investor meeting. “This is unethical” is a sentence that moves nobody in a room where everyone believes themselves to be reasonable.

What she did, in order

A clarifying question that surfaced the risk without accusing anyone. ”If a declined applicant asks us why, what do we tell them?” The room went quiet for a moment, because nobody had an answer and everybody understood the question.

The version she could build. Structured application fields and repayment history, with the accuracy cost measured rather than asserted, delivered by Sunday.

Three sentences in writing to her manager, naming the specific risk (a decline decision the company cannot explain, on features that stand in for ethnicity) rather than writing a manifesto. A verbal objection did not happen; a written one is visible to somebody accountable.

Escalation to the one person with authority, a compliance lead hired two months earlier whom most of the team regarded as an obstacle and who turned out to be delighted to hear about a problem before it shipped rather than afterwards.

The measurement that ended it

She did not win the argument by arguing. Over Friday she measured both models and produced one table: adding the three features moved the ranking metric by nine tenths of a point, and the decline rate in two wards went from near-parity to nearly double.

That table ended an eleven-minute meeting.

It is worth noticing what the table cost her. It was a day of work she had not been asked to do, on a weekend before a demonstration, to produce evidence for a position she already held. The alternative (asserting the position without the numbers) was free and would not have worked, because a room under deadline pressure discounts an objection it cannot check and cannot discount a decline-rate breakdown by ward.

There was also a rule underneath all of this that nobody in the meeting had mentioned. Several jurisdictions require that a person be reachable when a significant decision about somebody is automated, and the company’s product declined applicants with no explanation and no appeal route at all. That was not a feature request; it was a gap that existed before Thursday and would have existed if the meeting had never happened.