All Things AI · Season 3 · What You Can Build · Episode 14
Your lead sources, your response times, your conversion, and what happened to everyone who did not become a client. A model reading your own operating data will tell you where the business leaks, which is almost never where agents assume it is.
Everyone measures the wrong end
Ask an agent how business is and you get a number of closings. Ask what produced them and it gets vague quickly, usually resolving to a general sense that referrals are the main thing and the rest is a mix.
That vagueness is expensive. It means effort gets allocated by feeling, and feeling is systematically wrong about this, because the things that feel productive are the visible ones and the things that actually produce are often quiet.
The data to answer it properly already exists in your systems. What has never existed is a way to interrogate it without either a spreadsheet afternoon or paying somebody.
What is worth asking
These questions are answerable from records you already have, and the answers are frequently uncomfortable.
- How long does it actually take me to respond to a new enquiry, measured rather than remembered, and does that number change by day and by hour.
- Of everyone who enquired last year, what proportion did I contact more than twice.
- Which source produced enquiries that became clients, as opposed to which source produced the most enquiries.
- What is the typical number of contacts before somebody engages, and how many do I actually make before stopping.
- Who went quiet at the same stage, and is there anything they have in common.
- What did I spend money on that produced nothing traceable.
The one that always surprises people
Response time, measured honestly, is almost never what an agent believes it is.
The belief is based on the enquiries you remember, and you remember the ones you answered immediately because answering immediately is memorable. The ones that sat for eleven hours are not memorable, precisely because you were doing something else.
A model reading timestamps produces the real distribution, including the tail, and the tail is where the losses are. Most agents discover that their median is fine and their worst quarter is dreadful, and that the worst quarter correlates with specific days and specific hours. That is an actionable finding and it is invisible without measurement.
Why a model rather than a report
Your systems already produce reports. They answer the questions the software vendor decided to build, which are rarely the questions you want.
The difference is that you can ask this anything, in a sentence, including questions that span systems that do not talk to each other. Take the enquiry records from there, the contact history from there, the outcomes from there, and tell me what the ones that converted have in common.
That question has no report behind it and never will. It is also the most useful question in the list.
This is your data, not anyone else’s
Worth stating clearly because it sets the boundary for this whole capability.
This is about your operations: your response times, your sources, your conversion, your spending. It is analysis of how you work, and everything it produces is a fact about your own business that you are entitled to know.
It is not analysis of client circumstances, and it is not a tool for producing conclusions about property or anything requiring a licence or a qualification to state. Keeping that boundary clean is what makes this straightforwardly useful rather than complicated.
How you set it up
Exports and a question. It is genuinely this simple, which is why it is worth doing this month rather than eventually.
- Export what you have. Enquiries with dates, contact history, outcomes, and whatever spending records exist. Messy exports are fine.
- Put them in one folder and run a local model against it, because this is your entire business in one place and it should not leave your machine.
- Ask one question at a time and make it specific. Broad questions produce broad answers that tell you nothing you did not already suspect.
- Require it to show the numbers behind every claim, so you can check rather than believe.
- Rerun the same questions quarterly and keep the answers, because the trend matters more than any single number.
Where this goes wrong
Four failures, and the first is a genuine risk rather than an inconvenience.
Accepting a number without checking how it was calculated. A model will produce a confident figure from a messy export, and messy exports have duplicates, blanks, and records that mean something different from what the column heading suggests. Ask how it counted before you act on what it counted.
Asking questions too broad to answer. How is my business doing produces a paragraph of nothing. Specific questions produce specific answers.
Reading a pattern into a small number. Forty enquiries is not a dataset, and a model asked to find patterns will find them whether or not they exist. Be suspicious of any finding based on a handful of cases.
Doing it once. The value is in the trend, and one snapshot tells you where you are without telling you which direction you are going.
What to do with an uncomfortable answer
Most agents who run this properly find at least one thing they would rather not have found. Usually it is response time, sometimes it is how little follow-up actually happens, occasionally it is that a source everyone believes in produces nothing.
The instinct is to explain it. There were reasons for those particular months, that source was unusual this year, the numbers do not capture the relationships. Some of that is even true.
The more useful response is to fix the one that is easiest to fix and rerun it in a quarter. A single measured improvement is worth more than a good explanation, and the rerun tells you whether the fix worked, which is the thing no amount of reasoning about it can.
What it grows
Judgment, because you stop guessing which half of the effort is working.
Everything else on this site makes you more productive. Productivity aimed in the wrong direction is expensive rather than valuable, and most agents have no reliable way to know which direction is right, so they hedge across everything and dilute all of it.
Knowing where the business actually leaks changes what you do on Monday. That is a different kind of gain from doing Monday faster, and it is the one that compounds over years.
What it costs
Nothing, if it runs locally, which it should. This is your complete operating picture in one folder and there is no reason for it to go anywhere. The work is structured reading and arithmetic, which is well within what an open model on your own machine handles.
The only cost worth planning for is the hour it takes to get clean exports out of systems that were not designed to give them to you, and that hour is paid once because the next quarter’s export is the same export.
Questions agents actually type
Can AI analyse my real estate business data?
Yes. Export your enquiries with dates, contact history, outcomes, and spending into one folder and ask specific questions of it. It can also answer questions that span systems which do not talk to each other, which is where the most useful findings tend to be.
What should I ask about my own numbers first?
Measured response time, including the worst quarter rather than the average. Almost every agent believes theirs is better than it is, because the enquiries you remember answering quickly are the memorable ones and the ones that sat for eleven hours are not.
How do I know the numbers it gives me are right?
Require it to show how each figure was calculated. Exports contain duplicates, blanks, and columns that mean something different from their heading, and a model will produce a confident number from all of that. Check the method before acting on the result.
Should I run business analysis on a local model?
Yes. This puts your entire operating picture in one folder, and there is no reason for that to leave your machine. The work is structured reading and arithmetic, which open models running locally handle without difficulty.
How often should I run this?
Quarterly, using the same questions each time and keeping the answers. A single snapshot tells you where you are without telling you which direction you are moving, and the direction is the part that changes what you do.
Season 1 is Claude Code for Real Estate Agents, filmed and on its own hub. All three seasons: AI for real estate agents.