How to Build a Custom GPT or AI Assistant for Your Team

Your team asks the same questions every day. Where is the latest pricing sheet? What is our refund policy? How do we format a client proposal? A general chatbot cannot answer these well because it does not know your business.

A custom GPT, or more broadly a custom AI assistant, fixes that. You give an AI model standing instructions, your own reference documents and a few guardrails, and it becomes a focused helper that answers in your voice using your material. No coding is required.

This guide walks you through the build step by step. It also covers an important 2026 change: OpenAI is retiring custom GPTs in favor of plugins, and Google is moving Gems to skills. The design work you do here carries over to any platform.

What a custom GPT actually is

“Custom GPT” started as the name of OpenAI’s feature for building tailored versions of ChatGPT. People now use the phrase for any no-code AI assistant built on top of a large language model.

Every version has the same 4 building blocks:

1. Instructions. Standing guidance the assistant follows in every chat: its role, its rules, its tone and its output format.
2. Knowledge. Files you upload, such as policies, product sheets or style guides, that the assistant can search and quote.
3. Tools. Optional connections to other apps, like a calendar, CRM or ticketing system, so it can look things up or take actions.
4. Sharing settings. Who can use it, who can edit it and whether it is visible across your organization.

A custom assistant does not retrain the model. It wraps the model in context. That is why it is cheap and fast to build, and also why it can still make mistakes. If you want to understand how knowledge files get searched behind the scenes, our explainer on retrieval-augmented generation covers the mechanics.

What changed in 2026

If you built custom GPTs in ChatGPT, plan for a migration. OpenAI’s help center says custom GPTs are scheduled to stop running on December 11, 2026, with timing that can vary by plan and workspace. The replacement is plugins, which package skills (instructions and resources for repeatable workflows) with optional connections to tools through the Model Context Protocol.

According to OpenAI’s migration guidance, GPT instructions can become a skill and knowledge files can be copied into plugin reference files. Some things do not carry over automatically, including existing conversations, custom actions and sharing settings, and migrated plugins start out private.

Google is making a similar move. Its Gemini help center says Gems will transition to skills, starting with personal Google accounts in November 2026.

The lesson: treat your instructions, files and test questions as the real asset, and keep copies outside any single platform.

Pick a platform for your team assistant

Use the AI platform your team already pays for and signs in to. Switching platforms just for this feature rarely pays off.

Two quick facts help you choose. Claude’s help center says free accounts can create up to 5 projects, and Team and Enterprise plans can share projects with “Can view” or “Can edit” permissions. Google’s Gems support page says you can add knowledge files from your device, Google Drive or a NotebookLM notebook.

Step by step: build your team assistant

Step 1: Give it one job

The most common failure is an assistant that tries to do everything. Pick a single, repeatable job with a clear owner, such as:

• Answering HR policy questions for employees
• Drafting first-pass replies to common support tickets
• Turning meeting notes into a client recap in your house format
• Checking proposals against your brand and pricing rules

Write 1 sentence that defines success, for example: “New hires get a correct, cited answer to benefits questions in under a minute.”

Step 2: Write the instructions

Instructions are where most of the quality comes from. Use a simple structure:

Role: You are the internal HR policy assistant for [Company]. Audience: Employees in the US offices. Sources: Answer only from the attached handbook and benefits guide. Rules: - If the answer is not in the files, say so and point to hr@[company]. - Never give legal, tax or medical advice. - Do not ask for or repeat personal data such as SSNs or salaries. Format: Short answer first, then the policy section name and a 1-line quote. Tone: Friendly, plain English, no jargon.

Be specific about what the assistant should do when it does not know. “Say you don’t know” is one of the most useful lines you can add. Our guide on what prompt engineering is has more before-and-after examples you can adapt.

Step 3: Add a small, clean set of knowledge files

More files are not better. Upload the current, approved version of each document and remove drafts and duplicates. Some practical rules:

• Use text-based files like PDFs exported from documents, Word files or spreadsheets with clear headers. Scanned images of text are harder to search.
• Name files clearly, such as “Employee-Handbook-2026.pdf”.
• Put a date and owner on every file so you know when it is stale.
• Never upload data the whole audience should not see. If an assistant can read it, assume any user can get it out.

Step 4: Connect tools only when you need them

Tools let an assistant check a calendar, search a ticket system or create a record. They also expand what can go wrong. Start with no tools. Add one only when users keep asking for something the files cannot answer.

When you do connect tools, prefer read-only access first, require confirmation before any action that sends, deletes or spends, and use official connectors from the software vendor where available. Many platforms now use the Model Context Protocol for these connections.

Step 5: Test with real questions

Before you share, collect 15 to 20 real questions from the people who will use it, including a few tricky ones. Run each one and score the answers.

Fix the instructions and files, then rerun the full set. Save these questions; you will reuse them after every change or platform migration.

Step 6: Share, own and maintain

Share it with a small pilot group first, ideally 5 to 10 people. Name an owner who reviews feedback weekly, updates files when policies change and reruns the test set monthly. Add a line at the top of the assistant’s description saying who owns it and how to report a wrong answer.

Key takeaway: The instructions, approved files and test questions are your real asset, not the assistant itself. Keep them in a shared folder so you can rebuild on any platform in an afternoon.

Guardrails that keep your assistant safe

A custom assistant is only as safe as what it can see and do. Build these guardrails in from day one:

• Least access. Only upload files and connect tools the entire audience is allowed to use.
• No secrets. Never put passwords, API keys or customer records in instructions or knowledge files.
• Human review. Require a person to approve anything that leaves the company, like customer emails or published content.
• Clear limits. Tell users in the description what the assistant is for and what it should not be used for.
• Admin settings. On business plans, ask IT to review data retention, sharing and connector settings.

Remember that instructions are not a security boundary. A determined user may get an assistant to reveal its instructions or file contents, so design as if they will.

Common mistakes to avoid

• Too broad a scope. “Company knowledge assistant” usually performs worse than 3 focused assistants.
• Stale files. Outdated policies produce confident wrong answers. Review files on a schedule.
• No “I don’t know” rule. Without it, the assistant fills gaps with guesses.
• Skipping testing. A handful of happy-path questions will not reveal the failures your users will hit.
• No owner. Unowned assistants drift out of date within months.

If you are deciding whether you need a simple assistant or something that acts on its own, our comparison of AI agents vs chatbots vs copilots can help. For broader ideas on where AI fits a small team, see our guide to generative AI for small business.

The bottom line

Building a custom GPT or team assistant takes an afternoon, not a project plan. Choose one job, write tight instructions, upload a clean set of approved files, test with real questions and give it an owner. Platforms will keep renaming and reshaping these features, as OpenAI’s move to plugins shows, but a well-documented assistant moves with you.

Sources

• OpenAI Help Center: Custom GPT retirement and migration FAQ — scheduled December 11, 2026 retirement and migration details
• OpenAI Developers: Plugins concepts — what plugins, skills and MCP servers are in ChatGPT and Codex
• Google Gemini Help: Use Gems in Gemini Apps — creating Gems, adding knowledge files and transition to skills
• Claude Help Center: What are projects? — project instructions, knowledge, plan limits and sharing permissions
• Model Context Protocol specification — how assistants connect to tools and the consent principles for tool use