Which AI model should you use for which job?
Written for agents, brokers, lenders, contractors, home services.
The short answer
Two questions decide it. Is being wrong here expensive? That picks the model. Does the answer depend on weighing things, or on finding them? That picks how hard the model should work.
Finding is low intensity. Weighing is high intensity.
In practice: a first draft of a listing description is the middle model. A price reduction conversation with a seller is the top model. Renaming and filing 300 photos from a shoot is the smallest model.
One rule with no judgment in it: anything scheduled runs on the smallest model that can do the job. A scheduled task pays its cost every day, forever, and nobody reads the bill line by line.
There is no automatic router that reads the work and picks the model for you. That absence is deliberate, and the cost is that somebody has to decide out loud, once, and write the decision down.
Most people pick one model and use it for everything, usually the biggest one, because bigger sounds like it should be safer. So the same expensive model drafts a listing description, renames a folder of photos, checks that every document in a loan file has a date on it, and helps think through a price reduction with a seller who is already annoyed. Two of those four jobs were worth the money.
Choosing the model should be somebody's job. It should not be the business owner's job every time they open a chat window. What follows is the routing rule we settled on 2026-08-01, the lineup as we verified it against the vendor documentation on 2026-08-08, and one real bill that explains why the rule exists.
Two questions, and the second one is the one people skip
How do you decide which AI model a job needs?
Ask two questions. Is being wrong here expensive, which picks the model, and does the answer depend on weighing things or on finding them, which picks how hard that model should work. Finding is low intensity. Weighing is high intensity.
The first question is about the cost of an error. A listing description that reads flat costs you an afternoon and a rewrite. A price reduction conversation that lands wrong costs you the listing and possibly the referral behind it. Same week, same business, and only the second one needs the expensive model.
The second question is about what the work is made of. Finding is retrieval. Pull the six months of notes on this past client. List every job where we used that tile supplier. Tell me which files in this transaction folder have no date. A right answer is already sitting somewhere and the work is going and getting it.
Weighing is judgment. Should we take this listing at the number the seller is holding to. Is this subcontractor bid credible or is somebody about to eat a change order. Float or lock. Nothing is sitting anywhere waiting to be retrieved. Somebody has to trade one thing off against another and be accountable for the trade.
What each model is for
Verified against the vendor documentation on 2026-08-08. Model lineups change every few months, so check the date on any list like this one, including this one, before you build a habit on it.
| Model | Built for | Relative cost | A job in your week |
|---|---|---|---|
| Claude Opus 5 | Judgment work where being wrong is expensive. Strategy, pricing, anything a client will read. | Highest of the general models | Preparing the price reduction conversation. Deciding whether to take a listing at the number the seller wants. |
| Claude Sonnet 5 | Most work, and the sensible default. Research, drafting, building, a normal working session. | Moderate | A first draft of a listing description. What you know about a neighborhood before a listing appointment. A punch list template. |
| Claude Haiku 4.5 | Well specified work where the procedure is already clear. Filing, formatting, moving things, checking things. | Lowest | Renaming and filing 300 photos from a shoot. Checking every document in a loan folder has a date on it. |
| Claude Fable 5 | A specialist model. The rule we run is that it gets used when it is asked for by name and at no other time. | Above Opus 5 | No general recommendation. If nobody has named it for a specific job, it is not the answer to that job. |
The important line is the middle one. Sonnet 5 is the default and most of a working week belongs there. The top model is for the handful of decisions a month where an error is expensive.
The bill nobody was watching
On 2026-08-01 a file migration ran for several hundred turns on the largest model available. The work was file copies and checksums. A checksum is a short comparison confirming a copied file matches the original, so the job was copy this, confirm it arrived intact, repeat. The smallest model would have produced identical output.
In a brokerage that same job is moving four years of closed transaction files into a new folder structure. In a contracting business it is pulling every invoice out of one system and filing it under the right job. Nobody would pay a partner rate for that work if a person were doing it, and yet that is what happened, because the model was carried over from the last thing somebody did.
Nothing flagged it, because nothing was watching. There was no alert and no line item saying this is the wrong tool for this. The output was correct. The bill was several times what it should have been, and the only way anyone found out was by looking.
Why should a recurring task always run on the smallest model that can do it?
Because a scheduled task pays its cost every day, forever, and nobody notices. A one off session on the wrong model costs you once and you see the output the same afternoon. A daily routine on the wrong model bills quietly until somebody reads a statement carefully.
This is the one place where model choice stops being judgment and becomes a rule. Everything else here is advice you weigh against your own situation. This one we run as a rule: anything scheduled runs on the smallest model that can do the job.
Look at what scheduled work usually is. A morning summary of what closes this week. A weekly pass over which past clients have gone quiet. A nightly check that every active listing folder holds the disclosure. All finding and checking, procedure already settled, no weighing anywhere in it.
Two ways to get this wrong
- High intensity on settled work. A session set to think hard about something already decided will re-argue it. You ask for the onboarding document to be tidied and it comes back questioning the commission split you closed in March. Nothing it says is stupid. It is reopening a decision that already cost you three meetings, and now it costs a fourth.
- Low intensity on a judgment call. The more expensive failure. Ask a fast, cheap setting whether to price a listing at the seller number in a market with four comparable sales and it returns the first plausible answer with complete confidence. No hedge, no second option, no sign anywhere that the question was hard.
The second one is worse because confidence reads the same whether or not it was earned. A thin answer and a considered answer arrive in the same font.
A week in a real estate business, routed
| The job | Model | Why |
|---|---|---|
| Draft a listing description | Sonnet 5 | Drafting. A human edits it before anyone else sees it. |
| Prepare for a price reduction conversation | Opus 5 | A seller reads the outcome. Being wrong costs the listing. |
| Rename and file 300 photos from a shoot | Haiku 4.5 | The procedure is already written. This is filing. |
| Pull every promise you made to a past client out of a year of notes | Sonnet 5 | Finding, across a lot of material, and missing one matters. |
| Daily summary of what closes this week | Haiku 4.5 | Scheduled. The rule applies and no judgment is involved. |
| Hire the second crew or subcontract the overflow | Opus 5 | Weighing, expensive to be wrong, hard to reverse for a year. |
One line on that table is worth pulling apart, because most real work is mixed. Preparing for the price reduction conversation has a finding half and a weighing half. Gathering six months of showing feedback, the comparable sales and the days on market is retrieval, and the smallest model does it well. Deciding what to recommend to a seller who has already told you the number is not moving is the other half, and it is the only part that needs the top model.
Split a job in two and run each half where it belongs. That one habit takes more off the bill than any other change here, and it improves the output, because the expensive model arrives at the hard question with the facts already in front of it.
Where the rule lives so nobody has to remember it
A routing rule is worth almost nothing if it lives in one person's head. We keep ours in the operating file of a PAD vault, which is worth explaining because you can build the same thing yourself in an afternoon.
PAD is People, Assets, Deals. Three folders of plain text files on your own computer, one file per person, one per asset, one per deal, plus a single operating file telling an AI assistant how to work inside them. That operating file holds the standing instructions: how to file things, who a promise belongs to, and which model to use for which job.
Ours says three things. Most work is the middle model. The largest model only when being wrong is expensive. Scheduled work runs on the smallest model that can do the job. It adds one requirement: before any substantial piece of work, the session states which model it is using and why, in one sentence, before starting.
That sentence is most of the value. It moves the decision from invisible to visible and takes two seconds to read. You do not have to be the one choosing. You have to be the one who can see what was chosen.
The four boxes
The four boxes applied to model choice in a small real estate or trades business. The fourth box is the one worth arguing about here.
There, and should be
A single default model that handles the ordinary working week, and a habit of naming it.
- Most businesses already have a default. Usually it is whatever the person opened first and never changed.
- A default is correct, and the middle model should be it. Make it a decision somebody took once rather than an accident of history.
- Review it when the lineup changes, which is roughly every few months.
Missing, and should be there
A written routing rule, in the place the assistant reads before it starts working.
- Two questions and three sentences is the whole rule. This is not a project.
- The part people skip is the requirement that the session names the model it picked and why, in one sentence, before starting.
- Without it, the model in use is whatever was in use last time, which is how a file migration ends up on the most expensive model available.
There, and should not be
Any recurring or scheduled job running on a model heavier than the work needs.
- The expensive entry, because it charges every day and produces output nobody compares against a cheaper version.
- Go and look at what your scheduled tasks are set to. Most people have never checked, and the setting was inherited from whatever session created the task.
- If a scheduled job seems to need the top model, the real finding may be that the job needs a person.
Missing, and correctly missing
An automatic router that reads the work and picks the model for you without being asked.
- It would have to guess the cost of being wrong, and only you know that. The same request, tidy up this document, is trivial on a template and expensive on a settlement statement. Nothing in the words tells them apart.
- A router that guessed well most of the time would be worse than none, because you would stop checking. The failure you would stop catching is the low intensity answer to the judgment call, which arrives sounding certain.
- Name the cost honestly: somebody has to decide, write the decision down, and have it said out loud each session. That is a real burden and it is the correct one to carry.
Can you stop a team member from using the expensive model?
Not on a personal or professional plan. A hard block on a specific model, or a spend cap per user, needs a team or enterprise seat. Anyone describing either one on a personal plan is describing something that does not exist.
Three things hold without a team seat, and they are worth doing even if you later buy one. Choose deliberately and say the choice out loud in the session before the work starts. Default to the smallest model on anything that repeats. Say the weekly usage number in plain language: roughly this much got used this week, against this plan.
The last one matters more than it sounds. A percentage read off a settings screen and left there is not a number anybody acts on. A sentence in a Monday message is.
What this rule does not do
The whole rule fits on an index card. Is being wrong here expensive, and is the work weighing or finding. The reason to write it down is that on a Tuesday afternoon with a closing falling apart, nobody stops to think about it, and the model in use will be whatever was in use last time.
Common questions
- Which AI model should I use for writing a listing description?
- The middle model, which as of 2026-08-08 is Claude Sonnet 5. Drafting is normal working session material. A human reads and edits the description before a seller or buyer ever sees it, so the cost of a weak first draft is one round of edits.
- When is the most expensive model worth it?
- When being wrong is expensive and the answer depends on weighing things rather than finding them. Pricing, strategy, a hard conversation with a client, and anything a client will read under your name. In a normal week that is a handful of tasks out of everything you do.
- What model should a scheduled or recurring task use?
- The smallest model that can do the job. This is the one place where model choice stops being judgment and becomes a rule, because a scheduled task pays its cost every day, forever, and nobody notices. Most scheduled work is finding and checking, where the procedure is already settled.
- What is intensity and how is it different from model choice?
- Model choice answers whether being wrong here is expensive. Intensity answers whether the work is weighing or finding. Finding is low intensity. Weighing is high. High intensity on settled work makes a session re-argue decisions already made. Low intensity on a judgment call returns the first plausible answer with complete confidence.
- Can I set a spending cap per person on my team?
- A hard block on a model, or a spend cap per user, requires a team or enterprise seat. On a personal or professional plan neither exists. What works without one is choosing the model deliberately each session, defaulting to the smallest model on anything recurring, and saying the weekly usage number out loud in plain language.
- Is there a tool that picks the model automatically?
- There is no automatic router that reads the work and selects the model based on how expensive being wrong would be, and that is correct, because only you know what a given request is worth. The same words, tidy up this document, are trivial on a template and costly on a settlement statement.
Worth passing along
This one serves the brokerage owner and the lender branch manager in the same sphere, for opposite reasons. The broker is paying for sessions across a team and cannot see what any of them selected. The lender is running scheduled checks that fire whether or not anybody reads them, which is exactly the case where a model chosen once quietly bills every day. Both problems have the same three sentence fix.
Written by Brett K Moore. We build the PAD System, a records structure for people, assets and deals that lives as plain text files on your own computer. Read what it is.
Read next
How do you decide whether a tool belongs in your business?
A four box audit for a software stack: what is there and should be, what is missing and should be there, what is there and should go, and what is missing for good reason. Written for real estate agents, brokers, mortgage lenders and the trades who work alongside them.
What does it mean to connect your AI to your email and calendar?
What you are agreeing to when you click Connect, why a connector never tells you which account it is holding, and the one send rule adopted after a private client brief reached eighteen unintended recipients. Written for real estate agents, brokers, mortgage lenders and the trades who work alongside them.
Should your CRM be where your business remembers things?
A CRM is built to move people through a pipeline and it is very good at that. It is a poor place to hold what you know about a person, and the two jobs pull against each other. Where the line sits, and the one rule that stops the two systems disagreeing. Written for real estate agents, brokers, mortgage lenders and the trades who work alongside them.