पाठशाला Pathshala · उत्पाद Utpād, The product · Lesson 26 · Scale

AI features that earn their place in the product

An AI feature has a running cost traditional software never had. Keep the ones that move retention or revenue by more than they cost, and retire the ones that only move the pitch deck.

Pathshala, The Founder Library · 11 October 2026 · 7 min read

Most AI features are built because a demo looked good in a meeting. Some of them change what customers do and how long they stay. The rest add a running cost to every month and a slide to the next deck.

This lesson gives a way to tell the two apart before and after building: start from the metric, price the inference, prove the effect against a holdout, and review every AI feature as a line in the profit and loss. A figure computes whether a feature pays for itself.

Two kinds of AI feature

The first kind changes behaviour that the business is run on. An invoice that reads itself saves an accountant twenty minutes, so she stays. A search that understands misspelt Hinglish queries converts a browsing session into an order. These features show up in [retention curves](/library/retention-curves-and-flattening-test) and conversion rates. The second kind changes what the company can say about itself: a chat assistant in the corner of the screen, a summarise button nobody presses, an “AI-powered” label on a feature that worked before. These show up in the pitch deck and nowhere else.

The distinction matters because novelty fades fast. Sequoia’s Sonya Huang and Pat Grady reported in Generative AI’s Act Two in September 2023 that generative AI apps had a median ratio of daily to monthly active users of 14 per cent, against 60 to 65 per cent for some of the best consumer companies. Users try AI and leave. The same pattern happens inside a product: a feature gets a burst of trial in launch week and then a usage curve that falls to a few per cent. A feature has earned its place when the curve flattens at a level that changes a business metric.

Start from the metric, not the model

Google’s People + AI Guidebook puts the first question plainly: “Often a rule or heuristic-based solution will work just as well, if not better, than an AI version.” It lists where AI tends to be better, such as personalisation, natural language understanding and prediction, and where it is not: when the product must be predictable, when errors are costly, when the user needs full transparency, and when speed and low cost matter most. A GST rate lookup is a rule. A vendor name read from a crumpled photo of a bill is a model.

Write the case for an AI feature as a sentence with a number in it before anyone builds: this feature will cut monthly churn among small accounts from 4 per cent to 3.5 per cent, or raise checkout conversion on mobile by two points. Name the metric, the segment and the size. If the team cannot write that sentence, the feature is for the deck. Use the [one-question test](/library/prioritisation-rice-and-one-question) from prioritisation: if this works, what number moves?

The cost line software did not have

Traditional software costs almost nothing to run one more time. AI does not. Martin Casado and Matt Bornstein of a16z wrote in The New Business of AI in 2020 that the AI companies they saw often had gross margins in the 50 to 60 per cent range, against 60 to 80 per cent or more for comparable software businesses, partly because cloud and inference costs were so high. Model prices have fallen since, but the shape has not changed: every call has a cost, and a popular feature costs more. At the scale of the whole industry David Cahn of Sequoia framed the same gap in AI’s $600B Question in June 2024: spending on AI infrastructure running far ahead of the end-user revenue needed to pay for it. Every product team faces a small version of that question with each feature it ships.

Price each feature per call in rupees, including retries, the longer prompts that context adds and the hosting around the model. Multiply by calls per customer a month and divide by revenue per customer. That share is the [gross margin](/library/gross-margin-why-investors-fixate) the feature takes. A feature that costs 3 per cent of revenue and keeps customers longer is cheap. One that costs 12 per cent and keeps nobody is a subsidy.

At the defaults, 2,000 customers paying ₹4,000 a month with 3 per cent monthly churn, a feature that cuts churn by half a point and costs ₹0.40 a call at 300 calls a month keeps about ₹74 lakh of revenue over two years and spends about ₹44 lakh on inference. It turns net positive only in month 14. Retention gains compound slowly while inference is paid from the first day, which is why a year-one review often kills a feature that would have paid. Cut the cost per call to ₹0.15, by caching or routing simple calls to a smaller model, and the feature is positive from month six. Set the churn reduction to zero and no cost is low enough.

An AI feature earns its place when it moves a number the business is run on by more than its inference costs. Everything else is a slide.

Proving it moved the metric

Launch every AI feature with a holdout: a random 10 per cent of eligible customers who do not get it for a quarter. Compare churn, conversion or revenue between the two groups by cohort, not the users who chose to try the feature against those who did not, because the keen ones would have stayed anyway. Read the result only when the test has enough customers to detect the effect claimed; the [experimentation lesson](/library/experimentation-culture-real-statistical-power) shows how to check. Watch the usage curve weekly for the first two months and ignore launch week entirely.

Instrument quality as well as usage. Track how often the user edits, rejects or retries the output, as an [event in the tracking plan](/library/instrumenting-product-twelve-events). A feature with high usage and a high correction rate is being tolerated, not valued, and its retention effect will fade.

Watch the cost line as closely as the benefit. Inference cost per customer rises quietly as prompts grow, as the product adds context to each call and as heavy users find new uses. Set a budget per customer per month for each AI feature, alert when the top 5 per cent of users exceed it, and decide in advance whether heavy use is capped, priced as an add-on or absorbed. A feature that pays at average use can lose money on the customers who love it most.

A worked example: an accounting app in Jaipur

A Jaipur company sells bookkeeping software to 6,000 small businesses at ₹1,500 a month, with monthly churn of 4 per cent. It ships two AI features. The first reads purchase invoices from phone photos and fills the entry, GSTIN and tax lines included. The second is a chat assistant that answers questions about the books.

After a quarter against a holdout, invoice reading cuts monthly churn to 3.4 per cent; the chat assistant moves nothing measurable and is used by one customer in twenty. Invoice reading is valuable but expensive: 200 invoices a customer a month at ₹0.60 each is ₹120, 8 per cent of revenue, and over two years it would keep about ₹88 lakh of revenue while costing about ₹1.2 crore. The team routes invoices from repeat vendors through a template extractor and sends only new layouts to the model, which brings the blended cost to ₹0.25. The feature now nets about ₹38 lakh over two years and turns positive in month 13. The chat assistant is retired with a month’s notice and its budget goes to the extractor.

Data, trust and the law

An AI feature often sends customer data to a model provider. Under the Digital Personal Data Protection Act that provider is usually a data processor, and the company stays responsible for what it does with the data; the [DPDP lesson](/library/dpdp-act-what-it-requires-of-your-product) covers the obligations. Put the provider in the privacy notice, keep personal data out of prompts where the feature does not need it, and check whether the provider retains or trains on inputs. Trust is part of the metric: a customer who suspects her ledger is training someone else’s model churns for reasons no holdout will explain.

The quarterly AI review

Once a quarter, list every AI feature in production on one page. For each: the metric it was built to move and the holdout result; weekly usage and correction rate; cost per call, calls a month and gross margin taken; cumulative net from the ledger. Sort into three groups. Keep what pays back inside two years. Fix what would pay at a lower cost per call or a narrower audience, and give it one quarter. Retire what moves no metric, with notice to the customers who use it. Any new AI feature joins the list only with its metric sentence written.

The figure is a simplified model and the Jaipur company is illustrative. Published figures are as stated by their sources, checked 11 October 2026.

Sources

1. Martin Casado and Matt Bornstein, The New Business of AI (and How It’s Different From Traditional Software), a16z, February 2020 — Gross margins often 50–60 per cent against 60–80 per cent or more for comparable SaaS.
2. Sonya Huang and Pat Grady, Generative AI’s Act Two, Sequoia Capital, September 2023 — Median DAU/MAU of 14 per cent for generative AI apps against 60–65 per cent for the best consumer companies.
3. Google PAIR, People + AI Guidebook: User Needs and Defining Success — When AI is probably better, when it is not, and why a rule often works as well.
4. David Cahn, AI’s $600B Question, Sequoia Capital, June 2024 — The gap between AI infrastructure spending and end-user revenue.