Quick answer (updated August 2026): Build a custom GPT or Claude Project when the same questions get answered from the same documents over and over; it takes one to three hours on the chat subscription you already pay for, not new software. Skip it when the task needs actions rather than answers (that is automation territory) or when a purpose-built category tool already exists. The gate: client or patient data requires a business tier with training turned off.
Why nobody answers the real question
Search “custom GPT for my business” and you get two kinds of pages: button-by-button tutorials that assume you already know you need one, and vendor pages selling a monthly chatbot platform that assume the answer is always yes. Nobody sits in the middle and prescribes: is this the right fix for your problem, or the wrong one?
That middle is where we work. Hourback sells no software and takes no commissions, so we have no reason to talk you into a build or a subscription. This page gives you the diagnostic we use, the honest setup effort, the real cost, and the two questions that disqualify the whole idea. If you want the wider decision logic first, start with build vs buy for a small business; this article is the “smallest possible build” branch of that tree.
What a custom GPT actually is, in plain English
A custom GPT is a saved version of ChatGPT that starts every conversation already knowing two things: your documents and your standing instructions. Per OpenAI’s help center, you upload up to 20 knowledge files and write instructions that define behavior and tone; anyone you share it with then chats with a tool that answers from your material instead of from the open internet. No code is involved, and the builder itself works conversationally.
A Claude Project is Anthropic’s version of the same idea: a workspace with its own knowledge base and project instructions, available even on free Claude accounts (capped at five projects), per the Claude help center. Claude skills serve the same repeat-workflow purpose for step-by-step procedures.
Strip the branding and the concept is one sentence: your documents plus your standing instructions, living inside the chat tool you already use. Everything else on this page follows from that.
The diagnostic: does the same question keep hitting the same documents?
Here is the clean test. A custom GPT or Claude Project is the right prescription when three things line up: the same questions arrive repeatedly, the answers live in documents you already have, and the output you need is an answer, not an action.
That pattern is everywhere in small businesses once you look. A brokerage office fielding the same listing questions from agents all day. A restaurant group answering allergen and sourcing questions from the same menu specs. A contractor quoting from the same price book. An HR-of-one answering PTO questions from the same handbook. In each case a person is acting as a human search engine over a fixed set of documents, and that is precisely the job this tool does well.
The payoff ceiling is real but modest: St. Louis Fed research (2025) found workers who use generative AI report saving about 2.2 hours per week. If your repeat-question workflow eats less than an hour a week, leave it alone and check what to automate first for a better target.
When a custom GPT is the wrong fix
Two disqualifiers, and they do most of the triage work.
The task needs actions, not answers. A custom GPT can tell you what the onboarding checklist says; it cannot create the folder, send the welcome email, and add the calendar events. Multi-step administrative work that moves data between systems belongs to automation tooling (the Zapier and Make class), which is a different prescription covered in what to automate first.
A category tool already exists. If the pain is meeting notes, missed phone calls, bookkeeping, or scheduling, purpose-built products already solve it better than anything you will assemble in a chat window. This matches the evidence on builds generally: in MIT Project NANDA’s preliminary 2025 research, externally purchased or partnered AI tools reached deployment about twice as often as internal builds (roughly 67% versus 33%; self-reported, and correlation rather than causation).
| The task in front of you | The right prescription |
|---|---|
| Same questions answered from the same documents repeatedly | Custom GPT or Claude Project |
| Drafting in a fixed house style (replies, descriptions, proposals) | Custom GPT or Claude Project |
| Multi-step administrative work that moves data between systems | Automation tool (Zapier/Make class) |
| Answering the phone, booking jobs | Category tool (AI receptionist) |
| Meeting notes and follow-ups | Category tool (recorder/notetaker class) |
| Bookkeeping, invoicing, scheduling | The AI already inside that category’s software |
What it costs and how long setup takes
The cost is the subscription you probably already have. Building a GPT requires a paid ChatGPT plan per OpenAI’s help center; ChatGPT Plus qualifies, at $20 a month per OpenAI’s published pricing as of July 2026. Claude Projects can be created even on free accounts. There is no per-build fee and no new vendor. Sharing one with your whole team is what pushes you toward a business tier, priced per seat.
Effort is measured in hours, not weeks. A realistic first version: 30 to 60 minutes collecting the source documents, 30 minutes writing a page of instructions, 30 minutes testing with the ten questions your team actually gets asked. What separates the useful ones from the abandoned ones is not prompt perfection on day one; it is two or three revision passes over the following weeks, fixing the wrong answers real use exposes.
That low floor matters because the pattern is already proven at the individual level. NANDA’s same 2025 report found workers from over 90% of the companies surveyed were regularly using personal AI tools for work, while only 40% of companies had bought an official subscription. General-purpose chat tools also stick: roughly 80% of organizations investigated them and about 40% implemented them, versus just 5% of task-specific enterprise tools reaching production. A custom GPT rides the tool your people already adopted instead of introducing a new one.
Custom GPT vs Claude Project at a glance
| Custom GPT (ChatGPT) | Claude Project | |
|---|---|---|
| Plain-English definition | Saved ChatGPT with your files + instructions | Workspace with a knowledge base + instructions |
| Plan needed to build | Any paid ChatGPT plan | Free account works (five-project cap) |
| Knowledge files | Up to 20 files, 512 MB each | Knowledge base; large sets need a paid plan |
| Team sharing | Link, workspace, or GPT Store | Team and Enterprise plans |
| Training on your data | Consumer plans may train unless you opt out; Business/Enterprise off by default | Commercial terms bar training on customer content |
| Best first use | Repeat Q&A over fixed documents | Same, plus longer working sessions in one workspace |
Vendor details from each company’s published documentation as of July 2026 (linked in Sources); both change features often, so verify before you standardize on one.
The data rule before anything confidential goes in
One rule outranks every productivity argument: know which tier of the tool your data is sitting in. On consumer plans, OpenAI’s published privacy documentation states that data from individual ChatGPT versions may be used for training (opt-out available); its business tiers state “by default, we do not use your business data for training our models.” Anthropic’s commercial terms bar training on customer content for business use. And per OpenAI’s published policy, consumer ChatGPT accounts have no BAA path at all, a BAA being the signed contract that makes a vendor legally responsible for patient data.
The practical version: general documents like menus, price books, and public listings are fine anywhere. The moment client files, patient information, or privileged material would enter the knowledge base, you need a business-grade tier with training off, and for health data a signed BAA. Our guide to AI for professional-services firms walks through which client data can touch which tier; if that reads as overhead, it is cheaper than explaining a disclosure to a client.
How these builds fail
The failure modes are boring and predictable, which is good news, because you can design against all three. Gartner predicted in 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing unclear business value among the causes; the small-business version of that abandonment usually looks like one of these.
Stale documents. The GPT answers from the files you uploaded, forever. Prices change, policies change, listings close, and the tool keeps serving last quarter’s truth with full confidence. Every knowledge file needs a refresh date.
No owner. A custom GPT is a small product. Someone must own the instructions, the files, and the fix when it answers wrong. Unowned builds decay in weeks, and the adoption data shows how common that quiet decay is: the Federal Reserve’s 2025 Small Business Credit Survey (6,525 employer firms) found 46% of small businesses use AI while just 7% of those users have fully integrated it.
Overtrust without review. These tools answer fluently even when wrong. Keep a human skim on anything that leaves the building, and say so in the instructions (“cite the source file for every answer” is the single highest-value line you can write).
A worked example, labeled as an illustration
This is an illustration with made-up round numbers, not a client result. Picture a 12-person residential brokerage. The office coordinator answers roughly 15 repeat questions a day from agents and buyers: square footage, HOA rules, commission-split policy, disclosure timelines. At four minutes each, that is an hour a day, about five hours a week, spent being a human index of the listing sheets and the office policy manual.
The prescription: a custom GPT on the office’s existing paid ChatGPT plan, loaded with the current listing sheets and the policy manual, with instructions to cite the source file in every answer and say “not in my documents” rather than guess. Setup: about three hours across a week, including two revision passes. Assume conservatively that it fields only half the questions well and that every answer still gets a ten-second human skim: the coordinator gets roughly two hours a week back, rounded down, in line with the St. Louis Fed’s 2.2-hours-per-week finding for generative AI users. Ongoing cost: zero new dollars, one monthly document refresh, one named owner.
Whether the same math holds in your business depends on your actual question volume, which is a counting exercise, not a guess. Mapping those workflows and ranking them by payback is what an AI opportunity assessment does; the wider tool landscape is in our owner’s field guide to AI.
Bottom line
A custom GPT or Claude Project is the right fix when the same questions keep hitting the same documents: one to three hours of setup on the subscription you already pay for, a realistic ceiling of about two hours a week back per person (St. Louis Fed benchmark: 2.2), and zero new software. It is the wrong fix when the task needs actions (use automation tooling) or when a category tool already exists (buy it; bought tools reached deployment about twice as often as builds in MIT NANDA’s self-reported 2025 sample, a correlation rather than proven causation). Before anything confidential goes in, move to a business tier with training off. And give the thing an owner, because the tool does not fail loudly; it just quietly goes stale.
Sources
- OpenAI Help Center, “GPTs in ChatGPT,” 2026: help.openai.com
- OpenAI, ChatGPT plans and pricing page, as of July 2026: chatgpt.com
- OpenAI Help Center, “Creating and editing GPTs,” 2026: help.openai.com
- Claude Help Center (Anthropic), “What are projects?”, 2026: support.claude.com
- OpenAI, “Enterprise privacy at OpenAI,” 2026: openai.com
- OpenAI Help Center, “How can I get a Business Associate Agreement (BAA) with OpenAI?”, 2026: help.openai.com
- Anthropic, Commercial Terms of Service, 2025: anthropic.com
- Federal Reserve Bank of St. Louis, “The Impact of Generative AI on Work Productivity,” 2025: stlouisfed.org
- Federal Reserve Banks, 2026 Report on Employer Firms (2025 Small Business Credit Survey), 2026: fedsmallbusiness.org
- MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 (preliminary report), 2025: report PDF
- Gartner, “Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025,” 2024: gartner.com