Internal assistants: from GPTs to plugins
On this page
What changes by December 11, 2026
For two years, internal GPTs were the natural way to encapsulate expertise and make it available across an organization. OpenAI is retiring them in favor of plugins, which combine skills—reusable instructions and reference files—with connected applications.
The timeline is set. Custom GPTs will be retired on December 11, 2026, across all plans. FedRAMP spaces and Enterprise spaces with an approved exemption have until February 11, 2027, though the exemption does not preserve the right to create new GPTs. In affected Enterprise spaces, the creation of new GPTs ended on October 26, 2026; on Business and Edu, it depends on product availability, space settings, and permissions. Until then, existing GPTs remain usable.
The good news: almost everything you know about building a GPT transfers directly. Instructions become a skill, connected applications follow suit, and only custom actions need to be rebuilt.
What makes up an internal assistant
A GPT consisted of five elements: instructions, which defined its mission, tone, limits, and output format; conversation starters; a knowledge base, the reference files it consulted; selected capabilities, such as web search or image generation; and either connected applications or actions to your own APIs, never both at once.
A plugin follows this logic but structures it differently. The skill holds the instructions and reference files, and ChatGPT applies it automatically when a request matches its description. Applications provide access to services, including your internal MCP servers. Application templates, reserved for administrators, standardize their configuration. Skills are available on Business, Enterprise, Healthcare, and Edu.
In both cases, quality depends first on the instructions. A disappointing assistant is almost always one whose instructions were written in three minutes. Take the time to specify the mission, tone, constraints, and expected format.
Ten assistants to create first
| Assistant | Mission | Instructions and sources | Expected benefit |
|---|---|---|---|
| Onboarding | Answer new hire questions | Welcome booklet, org chart, HR FAQ | A 24/7 entry point |
| Marketing writer | Write in the brand voice | Editorial charter, approved content, personas | Consistent first drafts |
| Internal support | Resolve common IT issues | IT FAQ, procedures, guides | Fewer simple tickets |
| Data analysis | Help analyze CSV or Excel files | Analysis guidelines, metric definitions | Simple analyses without the data team |
| Meeting prep | Propose agendas and prep notes | Meeting templates, Drive meeting minutes | Better-prepared meetings |
| Sales writing | Proposals, emails, RFP responses | Templates, pitch decks, case studies | Faster sales cycles |
| Compliance | Answer questions on internal policies | AI policy, guidelines, procedures | A consistent answer instead of a search |
| Translation | Translate using house terminology | Glossary, business terms | Consistent translations |
| Document summary | Condense long documents | Drive or SharePoint connectors | Better-prepared decisions |
| Prompting coach | Help improve prompts | Internal guide, good and bad examples | Higher-quality usage across the board |
The prompting coach is the only one with an indirect effect: it improves how your colleagues use all the others. Teams often create it last, but it should be among the first.
Migrating your existing GPTs
Start with an inventory: which GPTs are actually used, who created them, and which use custom actions. When migration becomes available for your account, the creator opens My GPTs and selects Migrate to plugin; in an Enterprise space, an administrator can also migrate when access is enabled. Migration starts from the last published version: finish major edits first.
The resulting plugin is private by default and does not install automatically for other members. The original GPT remains usable until retirement but becomes read-only. Test the plugin with typical requests, compare results, then organize its sharing. Allow extra time for GPTs with actions: these integrations do not transfer and must be rebuilt using an application or an MCP server.
Governing without stifling
An internal assistant follows a lifecycle you should formalize before you have fifty of them:
- Creation—a colleague identifies a need and builds a prototype.
- Testing—five to ten users try it out for a week.
- Validation—a space administrator and the IT department verify compliance.
- Sharing—the assistant is opened to relevant groups or the entire space.
- Maintenance—an identified owner keeps instructions and sources up to date.
- Archival—obsolete assistants are removed every quarter.
Four rules govern sharing. Every assistant has an identified owner; otherwise, no one corrects drift. Its instructions are documented and accessible to administrators. Its sources are kept current: a “Compliance” assistant citing a repealed policy gives outdated information the authority of an official system. Finally, those that write to your systems require additional technical validation.
On the measurement side, Admin Console statistics track usage of GPTs, projects, and skills, with the latter ranked by number of invocations. An assistant unused after a month should be reviewed or archived; a heavily used one deserves documentation and adaptation for other teams with the same unexpressed need.
Your turn
Take inventory of your GPTs. List the GPTs in your space with their creator, recent usage, and whether they use custom actions. Sort them into three piles: migrate first, rewrite directly as a skill, or let retire. Set a test date for each migration, well before December 11, 2026.
Official sources: GPTs in ChatGPT · Custom GPT retirement and migration FAQ · Plugins in ChatGPT · Skills in ChatGPT · Workspace analytics for ChatGPT Enterprise and Edu (OpenAI Help Center)
Key takeaways
- GPTs will be retired on December 11, 2026, or February 11, 2027, with an exemption; creation has ceased in affected Enterprise spaces.
- Plugins replace them: the skill takes over instructions and sources, applications take over connections.
- Migration goes through My GPTs > Migrate to plugin; custom actions must be rebuilt.
- Ten cross-functional assistants cover most needs, including a prompting coach to create early.
- Every shared assistant has an owner, up-to-date sources, and a lifecycle through to archival.