Skip to main content

CG Lab

Case study · Real estate

WhatsApp and Gmail listings, one searchable database.

Elite SP Realtors, a South Mumbai brokerage, hired CG Lab to automate part of their daily work. The Property Inventory Suite reads broker messages from WhatsApp groups and Gmail, including flyers and images, and turns them into one clean inventory, matched against client requirements.

See it

On a laptop and on a phone.

The team searches one shared inventory from a laptop or a phone browser. Click through the views.

CG Lab Property Inventory Suite home screen with summary tiles, a Sync Now button and the latest listings.
The Sync page: summary tiles, one-click sync and the latest listings.
Search Listings page with price, area, area-group and configuration filters and a results table.
Search by price, size, area group, configuration and more.
Requirements page listing open client requirements with budget, locations and requester.
Client requirements captured from the same messages.
Matches page pairing client requirements with listings, with strength, reasons and caveats.
Each match is ranked, with the reasons and any caveats.
Home screen on a phone with summary tiles, Sync Now and a bottom tab bar.
Home on a phone.
Search Listings filters on a phone.
Search filters on a phone.
Matches page on a phone.
Matches on a phone.

Screens use fictional demo data. No real brokers, phone numbers or clients.

≈8,500listings in the live inventory
≈1,260open client requirements tracked
≈7,500requirement-to-listing matches kept up to date
≈19 paiseAI cost per message read

Figures as of mid-November 2026. In production since October 2026 with a five-person team.

The brief

The broker who replies first wins the client.

The brief was simple. When a prospective client shares a requirement, say a 3 BHK, a minimum size and a few areas, get back to them with matching properties, fast.

Speed decides the deal. A prospective client has usually contacted more than one broker, and tends to go with whoever comes back first with properties that actually match. The old way meant going through several WhatsApp groups and many emails by hand. It took many hours, sometimes more than a day.

  • Listings buried in chatA property posted at 11 am was lost by 3 pm under hundreds of newer messages.
  • The same listing, again and againBrokers repost and forward, sometimes with a changed price.
  • Messy messagesOne message can hold ten flats. Prices appear as “3.5 cr”, “350 L” or “Rs 450 psf”. Configurations as “3bhk”, “3 + SR” or “2+2 jodi”.
  • No single inventoryEach team member kept their own notes, so nobody could see what the others had received.
How it works

One click. Six things happen.

  1. Reads. The WhatsApp groups the team chooses (read-only), the shared Gmail inbox, and images such as flyers and brochures.
  2. Extracts. AI pulls out property type, configuration, area, price (sale or monthly rent), floor, parking, condition, broker contact, location and notes.
  3. Cleans. Fixed rules normalise everything, so “3bhk + SRB”, “3 BHK” and “3 bed with servant room and toilet” end up as the same clean values.
  4. Removes duplicates. Reposts are caught, and a changed price is flagged instead of added as a new row.
  5. Files by area, or asks. If a message does not say where the property is, it goes to a short list for a person. The tool never guesses a location.
  6. Matches. Client requirements captured from the same messages are matched against current listings and ranked Strong, Good or Possible, with the reasons.

The AI proposes and fixed rules decide. People stay in charge. Everything lands in one shared Google Sheet.

By the numbers

Reads thousands of messages. Costs paise per message.

Every sync records the AI tokens it used and converts them to rupees. These are averages from the tool’s own log across its first month in production, for a five-person team.

≈19 paiseaverage AI cost per message read
≈₹0.53average AI cost per new listing captured
≈₹190to read 1,000 messages
≈750messages read on an average working day

What the usage looked like

MeasureValue
Messages read by the AI in the first monthmore than 15,000
Messages on an average working dayabout 750
New listings added to the inventorymore than 5,500
AI calls madeabout 9,500
Tokens processedabout 100 million

Estimates calculated from token counts at the AI provider’s published prices, at ₹95 to the US dollar. The first weeks include one-off catch-up runs, so they overstate the steady-state cost.

Why it stays cheap

Four decisions keep the bill small.

  • Right-sized AIA fast, low-cost model reads text and email. A stronger vision model is used only for images and flyers. Per 1,000 new listings, that is about ₹110 on the low-cost model against about ₹1,270 on the stronger one, roughly one-eleventh of the cost. Treat that as indicative, not a lab benchmark.
  • Skip what is already knownBefore the AI is called, the tool checks whether the same message already produced a row in the inventory.
  • Prompt cachingThe standing instructions are cached, so they are not paid for in full on every call.
  • No serversIt runs on the team’s own laptops, with the Google Sheets, Drive and Gmail they already use. There is no hosting bill, and token use is metered on every sync, so a cost surprise shows up straight away.
Challenges we solved

The hard parts, and what we did about them.

  • WhatsApp has no official way to read a team’s groupsSo the connection is strictly read-only. It never sends messages, read receipts or presence, and listens only to groups the team allow-lists. For chats it ignores, it logs counts, never content.
  • Broker messages are wildly inconsistentDigests with a dozen flats, “jodi” pairs, ranges, rates written as prices, Hindi mixed with English. The AI extracts, then fixed rules re-check every field. Impossible values, such as a rent read as a price per square foot, are caught. 206 automated tests cover the rules.
  • AI that guesses is worse than AI that asksAn invented area silently corrupts the inventory. So locations are filled only when the message states them. Otherwise the listing goes to a Missing Info list, with suggestions from the team’s own history. A person’s decision is remembered and applied from then on.
  • “Same place” is a judgement callLower Parel, Prabhadevi and Worli are different localities but one market for a buyer. Editable Area Groups treat nearby places as one in search, and the team maintains them in the sheet without a developer.
  • Five laptops, one inventory, constant repostsThe shared sheet is the single source of truth. A fingerprint check catches look-alike reposts and flags a changed price instead of adding a duplicate.
  • Updating non-technical users without a serverEach launch checks a shared Drive folder, verifies every file against a list of hashes, and installs only a complete, intact release.
Good to know

What it is, and is not.

It runs on each person’s Windows laptop and syncs when someone clicks Sync Now. It is not an always-on cloud service yet, and a hosted version is a natural next step. It depends on an unofficial WhatsApp connection, which is why it is deliberately read-only. The AI can misread an unusual format, which is why important fields are re-validated by rules, unclear locations go to a person, and the team can correct the sheet directly.

Node.jsSQLiteAnthropic Claude APIWhatsApp Web (read-only)Google Sheets, Gmail, Drive
Questions

Quick answers.

Does it only read WhatsApp?

No. It reads WhatsApp groups, a shared Gmail inbox and images such as flyers and brochures, and puts everything in one inventory.

Will it send messages from our WhatsApp?

No. The connection is read-only. It never sends messages, read receipts or presence, and it only listens to groups your team chooses.

How much does the AI cost to run?

On average about 19 paise per message read, and about half a rupee per listing captured. These are estimates from token counts; the AI provider’s own console is the source of truth for the actual bill.

Can the AI get a listing wrong?

Sometimes. That is why every important field is re-checked by fixed rules, unclear locations go to a person, and the team can correct the sheet directly.

Where does the data live?

In a Google Sheet your team already uses. There are no servers to pay for or maintain.

Can this work for our brokerage?

Yes. The idea is not specific to one firm. We adapt the fields, areas and rules to how your team works. A free AI audit is the place to start.

More from CG Lab Digital

Other tools we have built.

Want one like it?

Tell us what slows your week down.

For a brokerage, a clinic, a school or an office. We start with a free AI audit.