When a Custom GPT is the wrong solution

Someone on your team spent a weekend building “Company Brain 9000,” a Custom GPT stuffed with half a drive of PDFs, a refuse list that says “be excellent,” and a public-ish share link. Monday they demo it in stand-up. By Wednesday three people treat it as the CRM. By Thursday Legal asks who approved customer emails going into a bot. By Friday the owner is on PTO and nobody can edit the instructions. The model did not create that mess. The wrong tool choice plus no ownership did.

This is Part 4 of the Custom GPTs tutorial, and it closes the series. Parts 1 through 3 covered building a simple GPT for a repeating task, instructions/knowledge/actions at a light depth, and sharing without chaos. Here we lock the chooser: when a Custom GPT is the wrong solution, which door to use instead, and how this four-part series fits the full ChatGPT track on AMS. This post also closes the ChatGPT deep curriculum on the site: Learn, product map, everyday, Work, Codex, and Custom GPTs. Optional next steps include the Gemini track or other Phase P work on the content plan. Foundations stay in Learn ChatGPT from scratch.

What you’ll learn

• A decision table for GPT vs plain Chat vs Work vs Codex vs API vs “not AI”
• Why one-offs, live systems, code, secrets, and official KPIs are poor GPT fits
• How to retire or downgrade a GPT that already exists without drama
• A recap of this four-part Custom GPTs series
• A recap of all six ChatGPT series on AMS and how they connect
• Sensible next steps: Gemini track, other Phase P topics, or stop and practice

Product menus change. The chooser logic should not. Re-check OpenAI Help the week you write team policy.

The short answer

A Custom GPT earns its keep when three things are true: the job repeats, the recipe is stable, and the data class is allowed. If any one of those fails, open a different door.

• One-off or still inventing the process → plain Chat
• Multi-step office work with tools, files, approvals → Work
• Repo, tests, software changes → Codex
• App for many users, keys, logs, automation without a human session → API / Platform
• Secrets, heavy PII, licensed advice as authority, official KPI truth → hard no or source systems + humans

Decision table: right door, wrong door

Use this before you open Create GPT. If two rows conflict, pick the safer surface and the smaller blast radius.

One-offs belong in Chat

Create GPT is a commitment. You get a name, instructions, optional files, maybe actions, and eventually a share story (Part 3). That overhead is silly for a problem you will solve once.

Signs you are still in one-off land:

• You change the role every time you open the chat
• You cannot write a one-sentence job that stays true for a month
• You have not run the recipe three times on real inputs
• You want “something smart” more than a fixed output format

Use plain Chat (and the everyday tutorial skills: job line, constraints, structure, verify). If the same shape shows up weekly, then freeze instructions and promote to a GPT. Building the bot first is how teams collect abandoned avatars.

Live systems lean Work, connectors, or API

A Custom GPT can hold knowledge files and, when the product allows, call actions. That still does not make it your integration layer. Multi-step office work with connected apps, intermediate artifacts, and approvals is the Work lane on AMS. Deep connector governance lives with workspace admins. Automated systems that run without a human ChatGPT session live on Platform/API or other backends.

If your pitch is “the GPT will keep our Salesforce and finance sheet in sync,” you are describing software. Pilot the human workflow in Chat or Work. Then decide whether you need an engineer, not a better avatar.

When Work beats a GPT

• Many steps and many files into one deliverable
• You need a supervised agent loop, not only a standing prompt
• Approvals and tool grants matter more than a shared recipe name

See the ChatGPT Work tutorial for the operating loop. A GPT can still help with a repeating instruction pattern after the workflow is stable. It does not replace Work’s job class.

Code belongs in Codex (with review)

If the artifact is a repository, tests, or a pull request, open Codex. A Custom GPT with “you are a senior engineer” in the instructions is not a coding agent with project context, diffs, and git habits. You practiced that in the Codex tutorial: explore, bound the change, review, small commits, never force-push blind.

Exception that is not really an exception: a GPT that helps humans write tickets or RFC outlines is fine. A GPT that “owns production deploys” is fantasy and a policy problem.

Secrets: hard no

Do not put production passwords, API keys, private keys, full customer extracts, payroll files, or unreleased M&A decks into GPT knowledge “so it knows our world.” Shared access multiplies exposure. Chat logs and team use are not a vault.

Same rule for instructions: “Always use this bearer token…” is a credential leak with extra steps. Keep secrets in approved secret stores and approved apps. Redact examples. Use synthetic data when you teach a recipe.

If the only way the bot works is by holding secrets, the bot is the wrong design. Build an internal tool with real auth, or do not automate that step.

Official KPIs are not a GPT’s job

Models are fluent. They are not your warehouse. A shared GPT that “knows our metrics” will eventually invent a number, round wrong, or cite last quarter’s definition after finance changed the grain. Humans will still paste the answer into a board deck because the avatar looked official.

Healthy pattern:

• KPI truth lives in BI, warehouse, or finance systems of record
• AI may help write SQL, outline a brief, or rephrase a narrative you already verified
• GPT instructions should ban inventing metrics and require an assumptions list (Part 2)
• Team share (Part 3) does not convert a recipe into an audited metric store

Rule of thumb: If someone will make a money, headcount, or compliance decision from the number, the number must come from a system you trust, not from a chat completion.

Full product UX is software

A Custom GPT is still ChatGPT under the hood: conversation UI, plan limits, OpenAI’s product rules. It is a great packaging for internal recipes. It is a weak host for a customer-facing product with your branding, your auth, your SLAs, and your data residency story.

The product map’s API part already drew this line. If you need multi-tenant apps, batch jobs, or deep audit trails in your systems, you are in builder land. Prove the prompt and workflow with a GPT if you want. Do not stop at the GPT and call it “the product.”

A five-minute chooser script

Before Create GPT (or before you promote a private bot to the team), answer these out loud. If you stall on any line, stay in Chat or pick another door.

1. Job in one sentence:
2. How often? (daily / weekly / monthly / once)
3. Output format fixed? (yes/no)
4. Data class allowed in ChatGPT here? (yes / no / unknown)
5. Needs live systems or multi-file agent loop? (yes/no)
6. Needs repo changes? (yes/no)
7. Would a wrong answer look “official”? (yes/no)
8. Named owner + backup if shared? (names or n/a)

Gate:
- once, or format still changing → Chat
- weekly + fixed format + allowed data + no live systems → Custom GPT candidate
- multi-file office agent → Work
- repo → Codex
- unknown data class or secrets → stop; ask IT
- official KPI / licensed authority → source systems + humans

Print that block into a team wiki if you want. It is shorter than a policy novel and longer than “we should AI this.” Pair it with Part 3’s share kit only after the gate says GPT.

Worked example: three wrong builds, three fixes

Wrong 1: The one-off turned permanent

Alex built a GPT for a single board narrative in March. By September the instructions still mention Q1 themes. People open it because it is in the sidebar.

Fix: Retire it. For the next board cycle, use plain Chat with a fresh job line, or a new GPT only after the format stabilizes. Add a retire date next time (Part 3).

Wrong 2: The CRM brain

Sam uploaded a CRM export to knowledge so the GPT could “answer any account question.” The export included emails and deal notes. The share level was workspace-wide.

Fix: Unshare immediately. Delete the knowledge file. Move account questions back to the CRM or an approved internal tool. If AI assist is allowed, use redacted samples and admin-approved connectors, not a stale dump.

Wrong 3: The coding GPT that ships to main

A team GPT pastes “production-ready” code into Slack. Nobody opens Codex, nobody runs tests, nobody opens a PR.

Fix: Ban “paste and deploy” in team norms. Use Codex for repo work, require review and CI, keep any prompt helper limited to tickets and design notes. Green text in chat is not a release.

How to downgrade or retire without drama

You already shared something that should not be a GPT. Good. Fix it in public with a short note.

Retiring: Account Answer Bot (workspace)
Why: knowledge held CRM export; wrong door for account truth
Do instead: CRM search / your AE / approved BI for revenue
Owner action: knowledge removed, share set to only me, then unlisted
Questions: #ops-gpt-bugs

Optional downgrade path when the job is real but the packaging is wrong:

Common mistakes (chooser edition)

FAQ

Is a Custom GPT ever the right first tool?

Yes, when you already know the weekly job and the format, and you have run it in Chat enough times to trust the recipe. First exploration still belongs in plain Chat.

Can one GPT replace Work, Codex, and the API?

No. Different doors, different blast radius. The whole ChatGPT track on AMS exists so you stop forcing every job through one avatar.

We already have ten GPTs. Should we delete them?

Inventory first (Part 3). Keep the few with owners, jobs, and allowed data. Retire the rest. Ten thin bots is worse than two maintained ones.

Does this mean Custom GPTs were a waste of three posts?

No. Parts 1 through 3 teach a real skill: package a repeating chat recipe and share it without chaos. Part 4 keeps that skill from becoming a hammer for every nail.

What this Custom GPTs series covered

Custom GPTs tutorial was the deep track for packaged chat recipes:

If you keep one idea from four parts: a Custom GPT is a maintained recipe with a named owner, not a substitute for Work, Codex, source systems, or real software. Build small, share carefully, retire often, refuse secrets and fake KPIs.

ChatGPT track complete on AMS

With this post, the planned ChatGPT deep curriculum on Analytics Made Simple is complete. Six series, one path from “what is this product” to “which door for which job.”

You do not need to reread all six in order every year. Use them as shelves: foundations when you are new, map when names blur, everyday when Chat is the job, Work when deliverables need an agent, Codex when code is the job, GPTs when a recipe should be shared and maintained.

A practical “I have thirty minutes” map:

• Names and modes feel fuzzy → product map (especially Chat vs Work vs Codex, and GPTs vs plain chat)
• You avoid ChatGPT for real work because outputs feel messy → everyday tutorial
• You keep starting agents for jobs that need a human paste → Work tutorial decision tables
• You ship code from chat summaries → Codex review and git parts
• Your team has five half-dead bots → Custom GPTs Parts 3 and 4 (this series)

What to do next (optional)

Finishing a track is a fine place to stop and practice. If you want the next product depth on AMS, the plan points to other Phase K work and later phases:

• Gemini track (optional next product deep-dive): Learn Gemini, product map, everyday use, Workspace patterns, and coding surfaces as the plan ships them. Start from Learn when those series go live.
• Other Phase K / later plan items: Grok, open models, Hugging Face, cloud data platforms, and Phase P topics appear on the content plan when scheduled. Do not assume every logo needs a bot on day one.
• Sibling stack: If you work across vendors, Learn Claude from scratch and the Claude Code / Cowork tutorials are the Anthropic shelf.
• Stay on ChatGPT: Revisit one series with a real work problem this month. Inventory your GPTs. Delete two. That is a better win than opening a seventh product tab.

All paths stay listed on Learn.

How to practice this week

1. List every Custom GPT you use at work. Mark each as keep, downgrade, or retire using the decision table.
2. For one keep: confirm owner, version notes, and data class (Part 3).
3. For one wrong door: write a three-line retirement note and move the job to Chat, Work, Codex, or a source system.
4. Pick one real Monday task and name the door before you open ChatGPT (Chat / Work / Codex / GPT / not AI).
5. Skim the six-series table above and bookmark the one shelf you actually need this quarter.
6. Optional: if Gemini is next for your company stack, note what job you would test first without building a marketplace of bots.

Quick recap

• GPT when the job repeats, the recipe is stable, and data is allowed.
• One-offs → Chat. Live multi-step office systems → Work / approved integrations. Code → Codex. Products at scale → API/software.
• Secrets and official KPIs are not GPT knowledge problems to “solve with more files.”
• Retire and downgrade in public; inventory beats ghost bots.
• Custom GPTs series complete: build → layers → share → chooser.
• ChatGPT track complete on AMS: Learn, Map, Everyday, Work, Codex, GPTs.
• Optional next: Gemini track or other plan items; or practice what you already have.

Sources

Research and further reading used for this article:

• OpenAI: Introducing GPTs (what GPTs are for; personal, company, public scopes)
• OpenAI Help: ChatGPT Work and Codex (Chat vs Work vs Codex product split)
• OpenAI Help: Managing GPT access in Enterprise and Edu workspaces (create/share governance; verify current controls)
• OpenAI Help Center (search current Custom GPTs, sharing, connectors, and plan feature articles)
• OpenAI: ChatGPT product overview (plans and surfaces move; re-check before policy)
• OpenAI Platform docs overview (when app/API path differs from ChatGPT app seats)
• Analytics Made Simple: Custom GPTs tutorial (this series)
• Analytics Made Simple: Learn ChatGPT from scratch
• Analytics Made Simple: ChatGPT product map
• Analytics Made Simple: ChatGPT everyday tutorial
• Analytics Made Simple: ChatGPT Work tutorial
• Analytics Made Simple: ChatGPT Codex / coding tutorial
• Analytics Made Simple: Learn (related paths on this site)

More definitions in Key Terms →

Merch on Etsy

Open store →

Dashboard to Excel Ceramic Coffee Mug View on Etsy → You're On Mute White Ceramic Coffee Mug, Data Science Mug, Data Mug, Analytics Mug, Statistics Mug, Programming Mug, Gift Mug, Coffee Mug View on Etsy → You're On Mute White Ceramic Coffee Mug, Data Science Mug, Data Mug, Analytics Mug, Statistics Mug, Programming Mug, Gift Mug, Coffee Mug View on Etsy → Naive Bayes Mouse Pad, Data Science Mouse Pad, Analytics Mouse Pad, Bayesian Mouse Pad View on Etsy →