A small commercial real estate firm can safely build its own custom GPT for any workflow where a person reads the output before using it, and should commission custom software only when the workflow has to run without that person in the loop. That single line, human-in-the-loop or not, decides more than price, features, or how technical you are. A custom GPT is a chat assistant you configure in an afternoon with no code; custom software is engineered automation that plugs into your systems and runs on its own. Confusing the two costs firms money in both directions: paying to build what a saved prompt would do, or trusting a chat window with work that needed a real, tested pipeline. This piece draws the line task by task, flags where DIY quietly turns dangerous, and shows how to use the free tool to de-risk the paid one.
What a Custom GPT Actually Is
A custom GPT is a saved configuration on top of a chat assistant you already pay for. You give it standing instructions, upload a few reference files, and it behaves like a specialist version of ChatGPT for one job. “Custom GPT” is ChatGPT’s name for this; Claude calls the same idea a Project, Gemini calls it a Gem, and Microsoft Copilot lets you build agents in Copilot Studio. The mechanics are close enough that the decision in this article applies to all of them.
Three things define what a custom GPT is, and what it is not:
- It is a chat interface. You type, it answers, you read the answer. Nothing happens until you copy that answer somewhere useful. The person is the last step every single time.
- It has no memory of your systems. It knows what you paste in and what you uploaded, and nothing else. It cannot see your CRM, your inbox, or last month’s rent roll unless you hand it over in the moment.
- It is built with no code, in an afternoon. A managing broker with no technical background can stand one up. That accessibility is the whole appeal, and the source of most of the confusion about what it can do.
A custom GPT is, in plain terms, a well-briefed assistant who only works when you are in the room.
What Custom Software Actually Is
Custom software is engineered automation. Someone writes code that connects to your tools, moves data between them, applies your rules, and runs on a trigger or a schedule without anyone watching. It is the difference between an assistant who drafts a reply when you ask and a system that reads every inbound email, ranks the deals, and updates your pipeline overnight.
The distinguishing features are the mirror image of a custom GPT:
- It integrates. It reads from and writes to your systems of record directly, so data does not get copied by hand.
- It runs unattended. The point is to remove the person from the loop for a repetitive task, which is exactly why it has to be built and tested properly.
- It needs a developer and a budget. A single well-scoped automation for a small firm generally runs $25,000 to $150,000 depending on scope and how messy your source data is, plus an annual maintenance tail to keep it correct. The principle that decides whether that spend is justified is laid out in our buy-versus-build playbook for CRE firms.
Custom software earns its price when a workflow is both painful and repetitive enough that a human doing the final step is the bottleneck. Until then, it is an expensive way to do what a chat window already does.
The One Line That Decides It
Forget price and feature lists for a moment. The question that actually sorts your workflows is this: does a person read every output before it is used?
If yes, a custom GPT is safe. A first-draft lease summary you will read and correct cannot do much harm if it is occasionally wrong, because you are the check. The tool is fast, cheap, and the error rate is caught by the human who was always going to review the work anyway.
If no, you need real software. The moment an output goes to a tenant, an investor, or your books without a human reading it first, “usually right” is not good enough. You need error handling, tested rules, and an audit trail, none of which a chat window provides. That is not a knock on the assistant; it is the wrong tool for unattended work, the same way a calculator is the wrong tool to run payroll.
This is why the honest framing is not “GPT or software” but “which of my workflows still has a human in the loop.” Answer that for each task and most of the decision resolves itself. The same instinct, matching the tool to whether a person stays in the loop, runs through where a Zapier-and-ChatGPT setup stops holding together.
What You Can Safely DIY Today
For a 4-to-20-person firm, a surprising amount of daily work sits safely on the DIY side, because it is drafting and reading work a person reviews anyway. A custom GPT loaded with your templates and a few strong past examples handles these well:
- First-draft lease summaries. Paste a lease, get a structured abstract of the terms you care about, then verify it. You read every one, so the occasional miss is caught.
- LOI and email drafting in your house style. Upload your best past LOIs and the GPT will draft new ones that sound like your firm, not like a generic template.
- Market write-ups and OM narrative. Turn your bullet points and comps into clean prose you edit before it goes out.
- Inbox triage and reply drafts. Paste a thread, get a suggested reply and a summary. You still hit send.
- Cleaning and reformatting data you paste in. Standardizing a messy rent-roll export into a consistent format, for you to check, not to file.
- A house-knowledge assistant. Upload your process docs and templates so anyone on the team can ask “how do we handle X” and get your firm’s answer.
Each of these keeps you as the final reader, costs a business-tier subscription, and takes an afternoon to set up. Getting fluent at exactly this kind of prompting is the cheapest capability a small firm can add, and it is the substance of our training playbook for small CRE firms.
Where DIY Quietly Turns Dangerous
DIY stops being safe when the workflow stops running through you. The danger is rarely obvious, because a custom GPT will cheerfully attempt anything you ask; it just fails silently at scale. Move to custom software when the task involves any of these:
- It has to run unattended. Anything triggered automatically, such as reading every inbound broker blast and building a ranked pipeline, has no human reading each output. That is software, not a chat.
- It writes back into your systems of record. The moment output has to update your CRM, your accounting system, or a tenant statement without a person transcribing it, you need a tested integration, not copy-paste from a chat.
- The volume is real. A GPT is fine for the ten leases on your desk. Three hundred documents on a deadline is a pipeline job, because no one is going to paste and check them one at a time.
- An error leaves the building unseen. If a wrong number or a bad clause can reach a tenant or investor without a human catching it, the human-in-the-loop safety net is gone and you need engineered controls.
- It has to be reproducible and auditable. When you must show your work later, a chat transcript is not an audit trail.
The tell is simple: if your plan for the workflow removes you as the reader, you have crossed from DIY into build territory, and the honest scope of that build belongs in a written statement of work, as covered in what a custom automation statement of work should actually contain.
The Number Problem
One boundary deserves its own warning: custom GPTs should not own arithmetic that has to be exactly right. Large language models generate the most plausible next token, which means they approximate math and will produce a confident, wrong total without flagging it.
For CRE that rules out a whole category of DIY. Do not let a custom GPT compute a CAM reconciliation, allocate expenses across tenants, total an underwriting model, or reconcile a rent roll as the system of record. It may get it right often, and that is the trap, because the one time it is off by a cell, the number goes to a tenant with audit rights or into a deal decision.
The safe pattern splits the work. Let the assistant do the reading, pulling caps, exclusions, and pro-rata bases out of a lease PDF so a person does not hunt through it, and keep the arithmetic in a spreadsheet or a rules-based tool where the math is deterministic. Reading is where a GPT shines; calculating is where it should never be trusted to run unchecked. If your workflow is mostly the math, that is a signal a purpose-built tool or a real build is the right buy, not a chat assistant.
The Data-Safety Checklist for DIY
Confidential deal data is the second kind of safety, and it is the one that makes principals nervous with good reason. DIY is safe here too, but only on the right plan and with the right settings. Before you paste a lease or a rent roll into a custom GPT, confirm all of these:
- You are on a business or enterprise tier, not free. ChatGPT Business and Enterprise and Claude Team and Enterprise both contractually exclude your inputs from model training by default and carry SOC 2 Type II certification. Verify the terms of the exact plan you buy rather than trusting the marketing page.
- The GPT is private, not published. A custom GPT you keep inside your workspace is very different from one shared to the public GPT Store. Keep anything touching client data private.
- You know what you are pasting. Redact what does not need to be there. The tool cannot leak a name or a number you never gave it.
- Uploaded knowledge files are business material, not a client’s crown jewels. Templates and process docs are fine; a signed, confidential PSA usually does not belong in a shared assistant’s permanent knowledge base.
Run that check once when you set the GPT up and it protects every session after. Skipping it is the most common way a well-meant DIY experiment turns into a data problem, and it is the kind of due-diligence gap flagged in the red flags CRE decision makers should watch for in AI proposals.
Use the GPT to Spec the Build
The two options are not rivals; the free one is how you de-risk the paid one. A custom GPT is the cheapest pilot you will ever run, and building the workflow by hand in one first tells you three things no proposal can:
- Whether the workflow is worth automating at all. If doing it by hand in a GPT already captures most of the value, you may not need a build. Many firms discover the bottleneck was never where they thought.
- What the real volume is. A month of doing the task in a GPT shows you how often it actually runs, which is the number that decides whether unattended software pays back.
- What the spec should say. You cannot write a good automation spec for a workflow you have never done step by step. Running it in a GPT hands your developer a precise description of inputs, rules, and edge cases, which shortens and de-risks the build.
Sequenced this way, DIY and custom software stop competing. The custom GPT captures value immediately and doubles as the requirements document for the build you commission only if the volume earns it. That fluency-before-spend order is the backbone of how a lean shop out-operates a bigger one, argued in full in the small-firm CRE manifesto.
A 30-Second Decision Test
Run any workflow through four questions in order. The first “no” that turns into a “must” is your answer.
- Does a person read every output before it is used? If yes, DIY a custom GPT. If no, keep going.
- Does it have to write into your systems or run on a trigger? If yes, that is custom software.
- Is the core of the task exact arithmetic? If yes, keep the math in a rules-based tool, not a chat.
- Is the volume more than a person can review one at a time? If yes, you have a build.
Most daily drafting and reading work stays on question one and is safely DIY today. The tasks that fail it are the ones worth paying to automate properly, and paying for the right one at the right time is the entire buy-versus-build discipline.
Frequently Asked Questions
What is the difference between a custom GPT and custom software?
A custom GPT is a no-code chat assistant you configure with instructions and a few uploaded files; it answers when you ask and a person reads every output. Custom software is engineered automation that integrates with your systems, applies tested rules, and runs unattended on a trigger or schedule. The practical difference is the human: a custom GPT keeps a person in the loop by design, while custom software exists to take the person out of the loop for a repetitive task. That is why one is safe to build yourself and the other needs a developer.
Can a small CRE firm build a custom GPT itself?
Yes, easily. Standing up a custom GPT takes no code and about an afternoon, which is the whole point of the format. A managing broker or ops director can configure one with the firm’s templates and reference docs and have a working lease-summary or LOI-drafting assistant the same day. The skill that matters is not technical, it is knowing how to prompt and how to check the output, which is exactly what a short hands-on workshop teaches. The limit is not your ability to build one; it is knowing which tasks are safe to hand it.
Is it safe to put lease or deal data into a custom GPT?
It can be, on the right plan. Use a business or enterprise tier, ChatGPT Business or Enterprise and Claude Team or Enterprise both exclude your inputs from training by default and carry SOC 2 Type II, keep the GPT private rather than published to the store, and redact anything that does not need to be there. Verify the data terms of the specific plan you buy, not the marketing page. On a free consumer tier, or with a publicly shared GPT, treat confidential deal data as off-limits.
Should a custom GPT do my CAM reconciliation or rent-roll math?
No, not as the system of record. Language models approximate arithmetic and can produce a confident, wrong total, which is unacceptable for numbers that go to a tenant with audit rights or into a deal decision. Use the assistant for the reading part, pulling caps, exclusions, and pro-rata terms out of lease PDFs, and keep the calculation in a spreadsheet or a rules-based tool where the math is deterministic. If a workflow is mostly arithmetic, that is a sign you want a purpose-built tool or a real build, not a chat assistant.
When is custom software worth the money over a custom GPT?
When the workflow has to run without a person reading each output. That means it triggers automatically, writes into your systems of record, handles volume no one can review one at a time, or produces results that reach clients unchecked. A single well-scoped automation generally runs $25,000 to $150,000 plus a maintenance tail, so it pays back only when a human doing the final step is a genuine bottleneck. If you are still reading every output, a custom GPT does the same job for the price of a subscription.
What can go wrong if I DIY something that needed real software?
The failure is usually silent, which is what makes it dangerous. A custom GPT will attempt any task and fail quietly at scale: a number off by a cell, a clause missed in document three hundred, an output that looks right and is not. When a person reads every result, those misses get caught. When you have wired the GPT into an unattended workflow it was never built for, the error leaves the building unseen. Notice the moment your plan removes you as the reader, and treat that as the line where DIY ends.
Do custom GPTs replace proptech subscriptions?
Rarely, and not the standard ones. Market data, e-signature, listing syndication, and accounting are workflows thousands of firms run the same way, so a vendor has already built the best version and you will not out-build it in a chat. A custom GPT replaces the drafting and reading work specific to your firm and poorly served by any tool. Think of it as adding a capable assistant for your bespoke tasks, not as retiring the stack.
Is a custom GPT a good way to test whether to build custom software?
It is the best way. Running the workflow by hand in a custom GPT for a few weeks is the cheapest pilot available and tells you three things a proposal cannot: whether the task is worth automating, how often it really runs, and what the build’s spec should say. If the GPT already captures most of the value, you may not need the build. If volume justifies it, you hand your developer a precise, road-tested description of the workflow, which shortens the project and cuts its risk.
Which is better for a firm with no IT department, a custom GPT or custom software?
Start with the custom GPT, always. With no IT department, a tool you can build and maintain yourself, and that keeps a human checking every output, carries far less operational risk than a build you must fund to keep alive. Custom software is still worth it for the right unattended, high-volume workflow, but only with a maintenance plan and a named owner in place first. Fluency and DIY first, then a targeted build once a specific workflow has earned it.
Where to Start
The first move is not to build a GPT or approve a project. It is to list your workflows and mark each one with a single letter: does a person read every output, yes or no. Everything marked yes is safe to DIY today with a custom GPT and your templates. Everything marked no is a candidate for custom software, to be scoped only after you have run it by hand in a GPT and confirmed the volume is real. A free AI-readiness assessment gives you that read fast: a short working session that sorts your workflows onto the right side of that line and returns an honest recommendation on what to build yourself and what, if anything, is worth paying to automate. Book a free AI-readiness assessment before you commission a build a saved prompt could have done, or trust a chat window with work that needed real software.
Arthur Wandzel
