What is a custom GPT or Gem?
Tools, assistants and prompting
A custom GPT or Gem is a user-configured version of an existing AI assistant, set up for a specific task, role or workflow. You combine written instructions with optional reference files, knowledge and tools, so the assistant behaves consistently for that job. For most business users this is configuration of an existing assistant, not training a new model. It packages repeat instructions, approved reference material and sensible defaults into something reusable. Because it is configuration rather than training, quality still depends on clear instructions, good source material and testing against the conditions you will actually use it in.
Reviewed by Jackie, Head of Learning & Development, Levellers · Last reviewed 8 June 2026
What this means
If you find yourself typing the same long instructions into an AI assistant again and again, a custom GPT or Gem lets you save that setup once and reuse it. You write the role and rules, optionally attach a few approved documents, and choose any tools the assistant may use. From then on, anyone with access gets the same configured behaviour without re-explaining the task.
The key point is what it is not. For most users it is not a brand-new model and it is not training. It is a saved configuration sitting on top of an existing assistant. That makes it quick to create and easy to change, but it also means its quality rests on the instructions and material you give it.
Why it matters
Custom assistants turn ad hoc prompting into something repeatable. That improves consistency across a team, reduces the time spent re-explaining tasks, and lets you bake in approved wording, reference material and guardrails. For a busy operation, packaging a good prompt pattern once and sharing it is a practical gain.
It matters too because the limits are easy to miss. A configured assistant can still give confident, wrong answers, and its behaviour can drift as the underlying assistant updates. Recognised risk frameworks stress that AI systems should be tested before deployment and regularly while in operation, measured against the conditions they will actually run in, not just a one-off demonstration.
How it works
Define the role and rules
You write clear instructions: what the assistant is for, how it should respond, what to avoid, and the format you want. This is the heart of the configuration.
Add context and knowledge
You can attach a small set of approved reference files so the assistant draws on your material rather than guessing. Keep this focused; a few authoritative documents beat a large unsorted pile.
Choose capabilities and tools
Depending on the platform, you can enable tools or connections, such as web access or links to other services. Each capability you add widens what the assistant can do and also what it can expose, so enable only what the task needs.
Preview, test and save
Guided creation flows let you preview behaviour, test it on realistic inputs, then save and share. Sharing and workspace controls govern who can use it, and data protections differ by plan, workspace and account, so check them rather than assume.
Examples
A proposal drafter that writes first drafts in your house style from a short brief and approved templates.
A policy explainer that answers staff questions using attached, approved policy documents.
A recurring analyst helper that applies the same structure and checks to a weekly report.
A task-specific connector that pulls from an approved source to answer a narrow, repeated question.
Common misunderstandings
"It is a brand-new model." For most users it is configuration of an existing assistant, not a new trained model.
"It stays good on its own." Behaviour can drift as the underlying assistant changes, so it needs occasional re-testing.
"Privacy is automatically handled." Data handling depends on the plan, workspace and account, and must be checked.
"The builder can read every chat." What a creator can see depends on the platform and settings, and should not be assumed either way; confirm it before relying on it.
Risks and boundaries
Quality drift: a configured assistant can degrade as the base assistant updates. Recognised guidance recommends evaluating against the actual conditions of use and monitoring over time rather than testing once.
Data handling varies: protections differ by plan, workspace and account. Confirm where inputs and files go before adding anything sensitive. Generative AI risk guidance highlights data privacy and the way humans interact with the system as specific risk areas to manage.
External exposure: enabling apps, APIs or connected services can send data outside your control. Add capabilities only where the task requires them.
No factual guarantee: like any assistant, it can produce fluent but wrong answers, so keep human review for anything consequential.
What to do next
Start with one repeated prompt pattern your team already uses. Write clear instructions, attach a small set of approved files, and build a short test pack of realistic inputs with known good answers. Run that test pack, then check three things: consistency across repeated runs, currency of the information, and privacy of the data involved. Re-run the test pack periodically, because behaviour can shift as the base assistant updates. Keep sensitive data out until you have confirmed where it goes, and limit connected tools to what the task genuinely needs.
Have a question or a suggestion, or want to understand how we research and review these guides? Read about our editorial standards and how to reach us.
FAQs
Is a custom GPT or Gem a new AI model?
No. For most business users it is a saved configuration on top of an existing assistant, not a newly trained model.
Can it use my company files?
Yes, you can attach a focused set of approved reference files. Confirm the platform's data handling before adding anything sensitive.
Can it connect to other systems?
Depending on the platform, you can enable tools or connections. Each one widens both capability and exposure, so enable only what is needed.
Is my data private when I use one?
It depends on the plan, workspace and account. Check the data protections that apply to yours rather than assuming.
Will it stay accurate over time?
Not automatically. Behaviour can drift as the base assistant updates, so re-test it periodically against realistic inputs.
Can the person who built it see my conversations?
That depends on the platform and settings. Confirm it before relying on either answer, especially for sensitive use.
Do I need technical skills to make one?
Usually not. Guided flows let you write instructions, add files and test, without coding.
How do I know it is working well?
Build a short test pack of realistic inputs with known good answers, run it on creation and again over time, and check consistency, currency and privacy.
Sources
Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (NIST). The MEASURE function's point that AI systems should be tested before deployment and regularly while in operation, against deployment conditions.
Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 (NIST). Generative AI risk categories including data privacy and human-AI configuration.
Challenges to the Monitoring of Deployed AI Systems (NIST (Center for AI Standards and Innovation)). Why post-deployment monitoring matters for tools that may drift in real-world use.