Case Study: How We Turned an Internal Feature Into a Standalone Paid Product

I run two products in parallel: nuph, the LinkedIn outreach platform whose blog you are reading, and AgonProp, an AI-first real estate platform in Argentina. Recently at AgonProp we did something I now think most SaaS founders should at least consider: we took an internal feature — the automated valuation model our real estate agencies used to price listings — and turned it into a standalone, self-serve, paid product with its own domain, its own checkout, and a completely different buyer.

Nothing technically new was invented in the process. The engine already existed. What changed was the packaging, the pricing, and a handful of small product decisions that made it sellable to strangers on the internet. That is what this post is about, because if your SaaS has been alive for more than a year, there is a decent chance you have a feature like this sitting in your codebase right now, quietly doing work that a much larger audience would pay for.

Quick context: what AgonProp is

AgonProp is a real estate platform built AI-first for the Argentine market. Agencies create listings from photos (the AI writes the copy and extracts the features), syndicate to multiple portals from one place, and manage leads in a CRM with automatic scoring. Buyers can search in natural language through an assistant we call Francesca. If you want the full technical breakdown of how the platform is built, I published a deep-dive on the architecture and product decisions on my personal blog — fair warning, that one is in Spanish.

The internal feature: an AVM for agencies

One of the platform features is an AVM — an automated valuation model. When an agency loads a property, the system looks at comparables in the same zone with similar typology and surface area, and suggests a price range with a confidence level. It exists because pricing is one of the hardest conversations in real estate: the owner has a number in their head, the agent has a market, and the two rarely agree. Giving the agent a data-backed range with real comparables turns an argument into a conversation.

For a long time, that was all it was: a helpful widget inside the listing flow, invisible to anyone who was not an AgonProp agency.

The signal that it could stand alone

Here is the thing about a valuation engine: the question it answers — what is this property actually worth? — is not an agency question. It is an everyone question. Every owner thinking about selling wants that answer before they talk to a single agency. In the US, that itch gets scratched for free by Zillow's Zestimate. In Argentina, there is no equivalent habit; people either ask an agent (who has an incentive baked into the answer) or guess from portal listings (which show asking prices, not market reality).

So we had an engine built for one audience, answering a question a much larger audience asks constantly, in a market with no default answer. Different buyer, different context, same code. That asymmetry — engine built for professionals, question owned by consumers — is exactly the shape to look for in your own product.

What we shipped

The result is a standalone online property valuation tool at tasador.agonprop.com. The flow is a four-step wizard: address, property characteristics, photos, checkout. The AI analyzes the photos to assess condition and finish quality, pulls real comparables from the zone, and generates a professional PDF report. Payment happens through MercadoPago — the dominant payment rail in Argentina — before you see the report, at a launch price of ARS $29,000, going to $75,000 once the promotional window closes in October 2026.

Simple on the surface. But almost every detail in that paragraph was a deliberate decision, and those decisions are the transferable part of this story.

The product decisions that mattered

1. Sell a report, not a number

The internal AVM outputs a range and a confidence score. If we had productized exactly that — type your address, see a number — nobody would pay, because a number on a screen feels like something that should be free. What we sell instead is a professional PDF report: the comparables, the photo-based condition analysis, the methodology, the range.

The PDF does three jobs the number cannot. It has perceived weight — a document feels like work was done. It is shareable — the owner forwards it to their spouse, their family, the agency they are negotiating with; it travels and argues on their behalf. And it justifies the price — you are not paying for an answer, you are paying for a document you can put on the table.

A feature answers a question. A product answers the question, charges for it, and hands you something you can show to someone else.

2. Make expensive inputs optional

Floor plans genuinely improve a valuation. They are also something most owners do not have at hand, and a required upload the user cannot complete is not friction — it is a wall, and the wizard dies right there. So floor plans are optional by design: the report gets better if you attach them, and the flow completes fine if you do not. The general rule: when a required field can kill the entire funnel, demote it. You can always ask for more later.

3. Charge before revealing — but de-risk aggressively

Checkout happens before the user sees their report. That is a scary pattern to ship: you are asking for money for something the buyer has not seen. We made it survivable in two ways: a full example PDF anyone can inspect before paying, and a plain-language "how it works" section explaining what the AI does with the photos and where the comparables come from. The buyer never has to wonder what they are getting; the only mystery left is what their own numbers will be. That is the correct thing to leave as the mystery.

4. Boring engineering: idempotent payments

MercadoPago, like every payment provider with webhooks, will happily deliver the same payment notification more than once. If your fulfillment code assumes one webhook per payment, you will eventually process something twice. We made payment handling idempotent from day one: every notification is checked against what has already been processed, and a duplicate is acknowledged and dropped. This is not glamorous work. It is also the difference between a product that quietly works and one that generates confused support tickets from your very first paying customers.

5. Close the loop after the money

After purchase, the buyer gets an email with a link to their report — and a button to report a problem. That button matters more than it looks. When you charge consumers upfront, your support surface is part of the product. Someone who paid and hit a snag needs an obvious, one-click path to tell you, before frustration curdles into a chargeback or a public complaint. I would rather hear about every single problem immediately than discover them later in a payment dispute.

6. Honest urgency

The launch price is real and so is its deadline: October 2026, after which it goes up. No fake countdown timers that reset at midnight, no eternal "sale ends soon". A promotional price with a genuine expiry does two things at once: it gives early users a real reason to act now, and it lets us launch at a price that lowers the risk of trying an unproven product — without anchoring the long-term value of the report at the promo number.

The spin-off cost almost nothing in new infrastructure

The stack is the same one the main platform runs on: React with TypeScript and Vite on the frontend, Supabase for Postgres and Deno edge functions, OpenRouter for the AI calls (vision for the photo analysis, text for the report), MercadoPago Checkout for payments. We did not evaluate new vendors, provision new infrastructure, or learn new tools. The marginal cost of the spin-off was the wizard, the report template, and the payment plumbing. That is the economic case for extracting a feature instead of building a second product from scratch: the hard part is already amortized.

How to spot your own extractable feature

The pattern generalizes well beyond proptech. Walk through your product and ask, for each meaningful feature:

• Does it answer a question people outside your user base ask? Our AVM answered a consumer question from inside a B2B product. That mismatch is the opportunity.
• Can it produce a self-contained deliverable? A PDF, an audit, a score, an analysis. Deliverables can be priced on their own; features cannot.
• Can a stranger use it without onboarding into your whole platform? If it needs your full data model and three weeks of setup, it is not extractable yet.
• Is there a payment rail your new audience already trusts? In Argentina that is MercadoPago. Whatever it is in your market, meet buyers where their money already lives.
• Does it run on inputs the user already has? Ours runs on photos everyone carries on their phone. If the input is something users must go create, the friction compounds.

One honest caveat: the tool is young and the promotional window is still open, so I am deliberately not quoting conversion rates or revenue here. When there is a meaningful dataset, I will write the follow-up with real numbers — the same way I documented how we landed AgonProp's founding agencies through cold LinkedIn outreach, with the actual messages and what worked. Publishing numbers before they mean anything is how case studies turn into fiction.

The part that does not change

Extracting the feature was the easy half. What gave us the conviction to do it was everything upstream: building the platform side by side with real agencies, and understanding from those conversations where the valuation question actually lives. Products spun off in a vacuum fail in a vacuum. Products spun off from observed demand at least start with a hypothesis someone else wrote for you.

If you are validating your own product spin-off, you need conversations with real users — the founders, operators, and buyers who will tell you whether your internal feature is actually somebody else's missing product. That is exactly what nuph is for: finding the right people on LinkedIn and starting those conversations at scale, without the manual grind. Start free.