All Things AI · Season 2 · The Models · Episode 11
Two or three, not one, chosen by the job rather than by reputation. A free local model for anything with a client in it, one paid service for convenience and hard reasoning, and a fast cheap one if you produce at volume. That covers almost everything.
The question is wrong, slightly
Which AI is best is the question everyone asks and it has no answer, in the same way that which vehicle is best has no answer.
The useful question is which one for what, and once it is asked that way it stops being a matter of opinion and becomes a short list of obvious matches. That is what this episode is.
The reason it matters is that most agents pick one, use it for everything, hit its weakest capability, and conclude the whole category is disappointing. The tool was fine. It was being used for the thing it is worst at.
The three-way split that covers almost everything
Not a ranking. Three jobs with different requirements, and a natural fit for each.
Anything with a client in it goes local
A document, a name, an address, a situation. A free open model on your own machine handles reading and extraction well, and the question of where the file went never arises. This is the category most agents currently handle worst, because they either use a hosted service and try not to think about it, or avoid AI on this work entirely.
Convenience and hard reasoning go to one paid service
The phone, the voice conversation, the thing you need while walking, and the small number of genuinely difficult problems. One subscription, not four. Most agents need less of this than they expect once the other two categories are covered.
Volume goes to something fast and cheap
If you produce a lot, the fast inexpensive services change what is worth doing, because running something across fifty items stops needing justification. If you do not produce at volume, skip this entirely.
Matching the job to the tool
The practical version, as a list you could keep.
- Reading a long document, extracting, comparing: local model. Free and private.
- Drafting something long from your own outline: local model, or an adapter on top of one.
- Thinking out loud in the car, dictating after an appointment: the paid service with the best phone and voice experience.
- A genuinely hard analytical problem where the answer is not in the material: the most capable paid service.
- Producing forty variations of something: the fast cheap one.
- Work spanning a whole folder of files: an editor with project-wide access rather than a chat window.
- Anything already sitting in your mail and documents: whatever model is already built into that.
What almost nobody needs
Four subscriptions. This is the most common failure and it comes from evaluating tools by feature rather than by job.
The major hosted services are broadly comparable for most work, and the differences that get compared in detail are rarely the differences that affect an agent’s week. Having three of them means paying three times for the same capability while still missing the two categories they do not cover.
One paid service, chosen for the interface you actually like, plus the free local ones, is a better setup than four subscriptions and it costs a quarter as much.
What to do first if you have none of this
In order, and the whole thing is one evening plus a week of use.
- Install a local model application and download a small model. Free, one evening, and it covers the category you are currently handling worst.
- Use it for a week on document work only, so you learn what it does before adding anything.
- Add one paid service, whichever phone app you prefer, and use the voice capability rather than typing.
- Only after a month, look at whether volume work or folder access is a real need for you. For many agents neither is, and the first two cover the job.
What to do if you already pay for several
Cancel down to one, and add the free local option you almost certainly do not have.
The test is honest usage rather than intention. Look at what you actually opened in the last month. Most agents find one service accounts for the overwhelming majority and the others are paid for out of a sense that they might be better at something.
The money is beside the point. Multiple similar tools mean no habit forms with any of them, and the capability that comes from knowing one tool properly is larger than the capability difference between any two of them.
Where this goes wrong
Three failures.
Choosing by benchmark. Published comparisons measure difficult reasoning problems that bear little relation to reading a document or drafting an explanation. A model that wins those may be worse for your week.
Choosing by whichever is newest. The order changes every few months and none of it changes what an agent should be doing, which is why this episode is written around categories rather than names.
Using one for everything. The single most common mistake and the reason so many agents have decided this technology is overrated. They met one tool’s weakest capability and generalised.
How to actually run the comparison yourself
Nobody else’s comparison is worth much, because the thing that matters is how a model handles your work rather than how it handles a benchmark.
Take one real task you do regularly and run it through each candidate with an identical request. Not a test question and not something clever. The actual thing, with your actual material, phrased the way you would normally phrase it.
Read all the outputs side by side, a day later rather than immediately, because immediate reactions are dominated by whichever answered fastest or opened most confidently. A day’s distance makes the differences in usefulness visible in a way they are not in the moment.
Then repeat it once with a second kind of task. Two tasks across three tools is about an hour of work and it will tell you more about what fits your week than any amount of reading, including this page.
Why one tool used properly beats three used casually
The capability difference between the major services is smaller than the difference between knowing a tool and dabbling with it.
Everything useful here comes from accumulation: standing instructions that got refined, memory that filled up, prompts that got saved and reused, a sense of how it fails and what to check. None of that transfers between tools, and none of it forms if attention is split.
An agent who has used one service properly for six months gets substantially more out of it than someone switching between three chasing whichever is currently rated highest. The switching itself is the cost, and it is invisible because it looks like diligence.
So pick, and stay, and change only when a genuinely different capability appears rather than when a slightly better version of the same one does.
How to hold this as things change
Build around the job rather than the product, and switching costs an afternoon rather than a rebuild.
The categories are stable even though the names in them are not. There will always be a most capable expensive option, a fast cheap one, and a free local one, and the right work for each does not move even when the leader does.
So when something better arrives, the question is only which slot it fills. That is a five-minute decision, and it is the reason it is worth understanding the shape of this rather than following which model is currently ahead.
What it costs
The local models are free permanently. One paid subscription in the usual monthly range covers the convenience and reasoning category. A fast cheap service is usage-priced and only worth adding if you actually produce at volume. For most agents the correct total is one subscription.
Questions agents actually type
Which AI is best for real estate agents?
There is no single best one, and picking one for everything is why most agents conclude the category is disappointing. Use a free local model for anything with a client in it, one paid service for convenience and hard reasoning, and a fast cheap one only if you produce at volume.
How many AI subscriptions do I need?
Usually one. The major hosted services are broadly comparable for most work, so paying for several means paying repeatedly for the same capability while missing the free local option that covers the category you are handling worst.
Should I choose an AI based on benchmarks?
No. Published comparisons measure difficult reasoning problems that bear little relation to reading a document or drafting an explanation. A model that wins those may well be worse for the work an agent actually does.
What should I set up first?
A local model, because it is free, takes one evening, and covers client document work, which is the category most agents currently handle worst. Use it for a week before adding a paid service, and add anything else only after a month.
Season 1 is Claude Code for Real Estate Agents, filmed and on its own hub. All three seasons: AI for real estate agents.