Putting AI to work in a brokerage, honestly

Every CRM now claims artificial intelligence. Most of the time, under the badge, there is a formula somebody wrote in an afternoon: add fifteen if the budget is large, add twenty if the timeframe is immediate, call it a score, call the score AI.

We shipped exactly that ourselves for a while, and we have stopped calling it what it was not.

Two different things, both useful

There is real value in arithmetic. A lead score built from budget, urgency, temperature and reachability is transparent, instant, free, and good enough to sort a calling list. It has one enormous virtue: you can read the rule and argue with it. When it ranks a lead highly you know exactly why.

There is also real value in a language model, but it is a different value. A model is good at the things arithmetic cannot touch:

• Reading six call notes and telling you what the client is actually worried about.
• Noticing that a lead has been in negotiation for nine days and that the last three calls all ended in "call back later".
• Writing the follow-up message you have been putting off since Tuesday.
• Turning a page of property facts into a description a person would read.

Those are judgement and language tasks. They are not scoring tasks. Dressing up the score as intelligence gets you neither.

Where we use it

BluPlots asks a model for three things, and each one is a click, not a background process:

A pipeline briefing. The administrator asks, and the model reads today's real numbers — stage counts, unassigned leads, stalled leads, open deals — and answers with what to do today, naming actual leads and actual people. It is a second opinion from something that has read every record and has no opinion about who works hard.

A lead brief. One lead's history, summarised, with the risks of losing it and a suggested next action, plus a draft message the agent can edit and send.

Listing copy. The facts on the form, turned into a description and the SEO fields, using only what was entered.

The rules we set ourselves

Nothing is sent that the answer does not need. A brief needs to know a lead wants a shop on Residency Road at sixty lakh and has sat in negotiation for nine days. It does not need their phone number, so their phone number does not leave the building.

Every call is logged. Who asked, which model, how many tokens, what it cost, how long it took, and the error if it failed. An administrator can see the bill and the failures on the same screen. AI that you cannot audit is AI you cannot run in a business.

It never writes to the database on its own. The model suggests; a person decides. A draft message sits in a box until an agent sends it. This is not timidity — it is that the cost of a confidently wrong action in a property transaction is measured in lakhs.

It is allowed to say it does not know. Every structured answer carries a confidence field, and the prompt tells the model to lower it rather than invent detail when a record is thin. A brief that admits the record is empty is more useful than a fluent paragraph about a lead nobody has spoken to.

What we will not do

We will not have a model set prices. We will not have it decide who gets which lead. We will not let it send anything to a client unread. Not because the technology cannot produce a plausible answer — because plausible is the exact failure mode, and a brokerage runs on being right about numbers.

The honest summary: a language model is an excellent reader and a decent writer, working over records it can see. Used for that, it saves an administrator the first hour of every morning. Sold as an oracle, it will eventually cost somebody a deal.