Article reader Listen + reading controls
Article reader
Preparing the reader…
Reading settings
GPTs turn prompting into configuration management¶
At its first developer conference, OpenAI has introduced generative pre-trained transformers (GPTs): custom versions of ChatGPT that can combine instructions, uploaded knowledge, and selected capabilities for a particular purpose. People can build them without conventional programming and share them inside an organization or, eventually, through a public store.
This will make useful experimentation easier. It will also turn a large number of informal prompts into organizational configurations that can affect real work.
A no-code assistant is still a system¶
These custom assistants may look like lightweight personal tools. Once colleagues rely on one for onboarding, support, drafting, or analysis, it has users, source material, expected behavior, and an owner—even if the organization has not named them.
The absence of code does not remove software-lifecycle concerns. Instructions can conflict. Uploaded documents become stale. External actions can change. A platform update can alter behavior. Users may apply the assistant to cases its creator never considered.
No-code lowers the barrier to creation, not the need for stewardship.
Instructions become executable policy¶
A custom assistant's instructions may encode tone, process, prohibited topics, escalation, and preferred sources. Those choices translate local practice into machine behavior.
That can help make tacit routines reusable. An experienced employee can capture the sequence of questions used to triage a request or the distinctions that make a briefing useful. It can also freeze an individual's habit into an apparent organizational rule.
Nonaka's theory of organizational knowledge creation helps frame the opportunity. Converting tacit knowledge into explicit instructions can expand access, but the explicit form should return to the community for testing and refinement. One expert's prompt is a hypothesis about practice, not the final process.
Create a registry before a sprawl¶
Organizations should make internal GPTs visible through a simple registry. For each one, record:
- purpose, intended users, and prohibited uses;
- accountable owner and review date;
- instruction and configuration version;
- knowledge sources, owners, and expiration;
- external actions or application programming interfaces (APIs);
- representative tests and known limitations; and
- usage, feedback, and retirement status.
The registry should not become a slow approval queue for every experiment. Use risk tiers. A private brainstorming assistant may need only basic data rules. A shared assistant that interprets policy or takes action needs stronger review, access control, and monitoring.
Test the boundaries, not only the happy path¶
Creators will naturally test whether a GPT performs its intended task. Review should also ask how it behaves when:
- the source material conflicts or is outdated;
- the user asks for something adjacent but unauthorized;
- uploaded content contains instructions;
- an external action has an irreversible consequence;
- the assistant lacks enough information; or
- a platform or model update changes familiar behavior.
The National Institute of Standards and Technology (NIST) Artificial Intelligence Risk Management Framework provides a useful lifecycle structure: govern the role, map the context, measure relevant behavior, and manage the risk over time.
Let successful experiments graduate¶
Most custom GPTs will remain small conveniences. A few will become important. Organizations need a graduation path from personal experiment to managed service.
Graduation should add domain review, formal source stewardship, a broader evaluation set, support ownership, incident reporting, and contingency plans. It should also ask whether the no-code configuration remains the right architecture or whether the workflow now deserves a purpose-built application.
OpenAI's GPTs make it possible for more people to shape artificial intelligence behavior around their own work. That is a meaningful form of democratization. The organizational response should not be to shut it down or pretend it is harmless.
Treat custom assistants as small systems whose importance can grow. Make their configuration visible, preserve what they know, test what they do, and give the useful ones a responsible path into the enterprise.
Sources and research trail¶
- OpenAI, “Introducing GPTs” (November 6, 2023).
- OpenAI, “New Models and Developer Products Announced at DevDay” (November 6, 2023).
- National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0) (2023).
- Nonaka, “A Dynamic Theory of Organizational Knowledge Creation” (1994).
- Walsh and Ungson, “Organizational Memory” (1991).