A useful AI assistant needs more than a clever answer. It needs to understand the work: which information matters, what a good result looks like, and where your permission ends.
For a creative business, that might mean preparing a client response without losing its voice. For an independent creator, it might mean turning an approved project brief into a consistent set of materials without explaining the same preferences every time.
Plugins and skills offer a practical way to organise that knowledge.
What the numbers actually show
In its 12 August 2026 report, OpenAI reports that 21% of weekly active users at frontier firms use plugins, compared with 9% at typical firms. For skills, the figures are 19% and 3%. It also reports weekly plugin use among 95% of its own active users.
These categories have specific meanings. Enterprise Signals ranks enterprise customers each month by output tokens per active user. “Frontier” means the top 10%; “typical” means the 45th–55th percentiles. These are usage categories, not rankings of creativity, profitability or business quality.
The analysis uses aggregated, de-identified enterprise usage data. The public report does not specify the exact measurement week or sample size for the plugin-and-skill comparison. Its separate analysis of work across functions draws on more than 10 million messages; that is not an identified sample size for these adoption percentages.
The figures describe adoption, not proof that plugins cause better performance. OpenAI’s internal 95% is also an AI-company example, not a target every organisation should copy.
A skill, a tool and a plugin are different things
OpenAI describes plugins as packages that can combine reusable skill instructions with apps connecting an agent to company information and actions. Its example pairs a sales playbook with CRM access so an agent can prepare a customer-specific response for review. Read OpenAI’s explanation.
A simple working distinction is:
- The skill describes the method: how to approach a task, which standards to follow and what to check.
- The tool performs a particular operation: for example, finding an account record or retrieving a document.
- The plugin brings useful capabilities together: instructions and connections arranged around work someone actually wants done.
Knowing the method does not grant access to private records. Having access does not grant permission to send a message.
One practical workflow for a creative studio
Imagine a small studio receiving a request for a short promotional film. This is an illustrative workflow, not a claim about a deployed Agentica IX service.
The studio creates a reusable skill containing its preferred response structure, tone, approved service descriptions and questions needed before quoting. It instructs the agent to flag missing details and avoid inventing availability, prices or production promises.
With authorised, limited access, the agent retrieves the enquiry, relevant client history and an approved proposal example. It prepares a draft that reflects the brief, identifies unanswered questions and links the records behind its factual claims.
The studio owner then checks the creative interpretation, scope, rights and commercial details. Sending the reply requires approval. Updating the CRM or sharing files is treated as a separate action with its own permission.
This keeps the first experiment small: one enquiry, one draft, one reviewer. The creator retains control over the relationship and the work’s direction.
Build the boundaries alongside the capability
Start with read-only access where possible. Connect only the information the task needs. Define when the agent must stop, ask or present a draft.
OpenAI’s plugin guidance says availability depends on the plugin, plan and workspace settings. Where supported, administrators can restrict connected apps to reading, permit selected actions and require approvals. Check the actual connection’s controls before relying on them.
Then test the result against your own standard. Did the draft preserve your voice? Were the facts traceable? How much correction did it need? Did it respect every approval boundary? Count review time as part of the work.
The Agentica IX view
The opportunity is to make your particular way of working easier to carry forward: your taste, your priorities and the details that make your work recognisable.
Choose one recurring task. Write down what good looks like. Connect only what is necessary. Keep the final judgement yours.
That is a practical place to begin extending your individuality with AI.