All Things AI · Season 3 · What You Can Build · Episode 11
There is a difference between a model that writes you a note to paste in and one that opens the record and updates it. The second exists now. It operates software the way a person does, which means it works with the systems you already have rather than needing them replaced.
Drafting was never the bottleneck
Ask most agents what they want AI for and they say writing. Ask them where the hours actually go and it is not writing.
It is the twenty minutes after every appointment spent putting what happened into the system. It is the Sunday evening reconciling what you remember against what is recorded. It is the tags that are wrong, the stages that are stale, and the notes that live in your head because entering them was never worth the friction.
A model that drafts a follow-up email has helped with the easy part. The part that actually consumed the week is data entry, and data entry was untouched until very recently.
What operating software actually means
Models can now see a screen and use it. They read what is displayed, work out where things are, click, type, and navigate, the same way a person does.
That sounds like a small technical distinction and it is an enormous practical one. It means the model does not need permission from your software vendor. It does not need an integration to exist. It works with whatever you already have, including the systems that have no useful way to connect to anything.
No integration required
This is the part that matters most in this industry, where a lot of the software is closed and the available connections are limited. A model operating the interface bypasses that question entirely.
It is slower than a real connection
Working through a screen is not as fast or as reliable as a proper connection would be. Where a real one exists it is better. This approach is for everything else, which is most things.
It can be watched
You can see it work, which is unusual and useful. The steps are visible rather than happening invisibly, and that makes it far easier to trust and to correct.
The jobs worth handing it first
Repetitive, rule-shaped, and low-stakes if it goes wrong. That is the profile to start with.
- Logging what happened after an appointment, from a note you dictated in the car.
- Updating stages and tags according to rules you can write down plainly.
- Finding records that are incomplete or contradictory and listing them for you.
- Moving information between two systems that do not talk to each other.
- Pulling the same report from the same three places every week.
The rule that makes this safe
It proposes, you approve, then it acts. At least at first.
Something operating your systems can make a mess, and the mess is in the database your business runs on. The correct sequence is a stage where it says here is what I am about to change, you look, and it proceeds. That step costs seconds and it removes the entire category of failure where you discover a hundred records got altered wrongly.
After a few weeks you will know which jobs it does perfectly and those can run unattended. The ones involving judgment about a person keep the approval step permanently, because the cost of getting one of those wrong is not measured in records.
Start with reading before writing
There is an entire useful phase before you let it change anything.
A model that only reads your systems can already tell you who has not been contacted in ninety days, which records contradict each other, where the data is missing, and what the last note on someone actually said before you walk into a meeting. None of that can damage anything.
Most agents find the reading phase valuable enough on its own that they run it for a month before turning on any writing. That is a sensible order, and it also teaches you exactly how the model interprets your data before it can act on that interpretation.
Where this goes wrong
Four failures, and the second is the expensive one.
Turning it loose on the whole database on day one. Start with ten records. Watch what it does. The instinct to run it across everything immediately is the instinct that produces a bad afternoon of cleanup.
No backup before a bulk change. Anything touching many records at once gets an export taken first, without exception. This costs a minute and it is the difference between a mistake and a disaster.
Rules that are not actually rules. If you cannot write the rule down in a sentence, it is judgment, and judgment does not get automated. Move this contact to the next stage when they have replied twice is a rule. Move contacts who seem serious is not.
Assuming it stays correct. Software interfaces change. A process that operates a screen breaks when the screen moves, silently and without warning, so anything running unattended needs a check that it is still doing what you think.
What it grows
Hours, and accuracy, and those are more connected than they look.
The reason CRM data decays is not that agents do not value it. It is that entering it well is tedious at exactly the moment you are least inclined to be tedious, which is immediately after an appointment when you are already thinking about the next thing.
Move that work to something that does not mind doing it and the records get updated at the moment they change rather than reconstructed on Sunday from memory. Everything downstream that depends on the database being accurate becomes worth relying on, which for most agents is the difference between having a database and having a list.
What to write in the brief
The brief is the whole difference between something useful and something that makes a mess, and it is worth more care than the setup.
Write it the way you would write instructions for someone competent who has never seen your business. What the system is for, what each stage actually means in your practice rather than what the software calls it, and what a complete record looks like when it is done properly.
Then write the exclusions, which matter more. What it must never change. Which fields are yours alone. Which records are not to be touched under any circumstances. A model with clear boundaries behaves well inside them, and a model with no stated boundaries will find edges you did not know existed.
Keep it short enough that you would actually reread it in three months, because you will need to when something behaves unexpectedly and you have forgotten what you told it.
How this fits with the rest
This is the capability that makes several others actually usable.
Matching new inventory against your database only works if the database is current. Follow-up systems only work if the stages are right. Anything that reads your numbers only tells you the truth if the underlying records are true.
So while this is not the most exciting capability on the list, it is close to a prerequisite for the ones that are. An agent choosing where to start would do well to start here even though it is the least interesting to describe.
What it costs
It runs against a model you are already paying for, so there is no separate cost beyond the processing, and screen-based work uses more of it than plain text does. Against twenty minutes after every appointment and a Sunday evening every week, the comparison is not close.
Questions agents actually type
Can AI update my CRM for me?
Yes. Models can now operate software through the screen, reading what is displayed and clicking and typing the way a person does. That means it works with systems you already have, including closed ones with no useful way to connect to anything else.
Is it safe to let AI change my database?
With two habits, yes. Have it propose changes for approval before acting, at least for the first weeks, and take an export before anything that touches many records at once. Those two steps remove almost every way this causes real damage.
What should I automate in my CRM first?
Start with reading rather than writing. A model that only reads can tell you who has gone cold, which records contradict each other, and what the last note said before a meeting, and none of that can damage anything.
Why does AI drafting not save much time?
Because drafting was never where the hours went. The time goes into entering what happened, keeping stages current, and reconciling records from memory. Drafting helps with the easy part, and the tedious part is what actually consumed the week.
Season 1 is Claude Code for Real Estate Agents, filmed and on its own hub. All three seasons: AI for real estate agents.