Skills vs. Plugins: Costs and Returns for Business AI
What you’ll learn
ROI measures added profit and avoided cost, minus implementation and operating cost, divided by that cost. P&L means a company’s profit and loss; gross profit is revenue minus the cost of providing the goods or services. Incremental gross profit is the additional gross profit attributable to the rollout.
- How procedure, connection, distribution, and permission boundaries distinguish skills from plugins
- The organizational controls in ChatGPT Business and Enterprise and Claude Team and Enterprise
- The shared model and product-specific differences in OpenAI and Anthropic skills and plugins
- How to connect time saved to incremental gross profit, avoided cost, and full operating cost
- When to choose a skill, a plugin, a staged rollout, or no rollout
The unit of reusable work separates skills from plugins
A skill gives AI a reusable procedure, decision criteria, and output format for a particular kind of work. A plugin is an installable business package that can bundle skills, connections to business data and capabilities, or both. External connections are optional; a plugin can distribute skills alone. OpenAI and Anthropic include different components in their packages, but this basic distinction is shared.[1][7]
A clear request to AI may be enough to save an individual time. Producing organizational impact on revenue, gross profit, throughput, or quality requires more. The procedure must remain reusable when the operator changes, data access must follow permissions, and administrators need a way to govern distribution and use.
A “basic versus advanced” framing does not distinguish skills from plugins. The decision turns on whether the reusable unit is the procedure alone or a distributable business workflow that can include several procedures, connected systems, and actions. By the end of this article, you will be able to select that unit from the P&L outcome and controls the organization needs, rather than from product names or feature counts.
A skill standardizes the method; a plugin packages the workflow
A standard operating procedure and a departmental work kit provide a useful analogy. The procedure aligns the inputs, sequence, decision rules, deliverable, and checks. That corresponds to a skill. The work kit packages the procedure for departmental distribution and can also include access to the required systems and tools. That corresponds to a plugin.
The analogy explains the roles, not every implementation detail. A real skill may include examples, sample outputs and programs that automate steps in addition to prose. Installing a plugin does not automatically grant access to a connected service. The workspace (the shared environment for company members) must allow the capability, and the account whose identity has been verified still needs permission in the external service.[2]
The useful comparison is therefore not how many features each option has. It is what the organization will manage as one change and distribution unit.
| Aspect | Skill | Plugin |
|---|---|---|
| Reusable unit | Procedure, judgment criteria, and output format | Skills, connections, executable capabilities, and distribution settings |
| Best-fit problem | Output variation, missed steps, and knowledge transfer | Cross-system workflows, departmental distribution, and connection management |
| External data | Can describe how to use data but needs a separate connection | Can include connections to external services |
| Permissions | Who can use, create, or share the procedure | Plugin availability, provider authentication, and allowed actions |
| Typical evaluation | Completion, correction rate, output variation, training time | Throughput, connection success, approvals, business outcomes, external failures |
A skill is the smaller change unit when the immediate need is a consistent method. A plugin becomes useful when external retrieval or updates, departmental distribution, and action permissions need to move together.
For the foundational concepts, see What Are Plugins? Skills, MCP, and Distribution. AI Agent Skills vs. Instruction Files covers the design of the skill itself in more detail.
Administrative scope in ChatGPT Business and Enterprise
The comparison covers company-account sign-in (SSO), identity checks beyond a password (MFA), automatic account creation and suspension as staff join or leave (SCIM), and conversation or execution-record retrieval (Compliance API). Governance here means managing access to information and actions, approvals, and records.
OpenAI positions ChatGPT Business as a shared workspace for growing organizations and Enterprise as the option for company-wide deployment with more advanced controls. On October 3, 2026, the official pricing page listed Business at US$20 per user per month with annual billing, or US$25 with monthly billing, for two or more users. Enterprise uses a sales contact process.[3]
| Administrative area | ChatGPT Business | ChatGPT Enterprise |
|---|---|---|
| Positioning | Shared workspace with essential administration | Business plus enterprise-grade security and control |
| Identity and access | Company sign-in (SAML SSO), additional identity checks (MFA), essential admin controls | Adds account synchronization for joining and departing staff (SCIM), company-domain verification, and permissions by role |
| Audit and analytics | Day-to-day workspace administration | Adds conversation and execution-record retrieval (Compliance API), audit logging, and broader usage governance |
| Data controls | Business data is not used for training by default | Adds controls for storage duration and the region where data is stored |
| Skill and plugin operations | Workspace-wide enablement and sharing | Better fit for permissions by role, public-plugin catalog review, and granular governance |
OpenAI’s Codex is an AI agent that writes and checks code, among other tasks. An OpenAI skill is a reusable workflow made from task instructions and supporting resources. A plugin can include skills and MCP servers (connections that let AI use data and functions in external services), with optional custom interfaces and hooks (automatic actions triggered at defined points in a Codex run). ChatGPT and Codex use a shared public plugin list (directory).[1][2]
Plugins work in Chat and Work on ChatGPT web, desktop, and mobile, in Codex inside the ChatGPT desktop app, and through the Codex CLI (Codex in a screen where you type instructions and commands) plugin browser. The current documentation says that the Codex IDE extension (Codex inside a code editor) does not support plugins.[2]
Making a plugin available does not grant access to Slack, Google Drive, or another provider. Plugin availability, MCP access, allowed actions, and the connected account’s permissions remain separate control layers.[2]
Administrative scope in Claude Team and Enterprise
A marketplace is a list for distributing plugins. Claude Cowork is an environment for assistance with computer-based work; Claude Code supports development tasks such as writing code. Sync here means applying distribution settings to a member’s environment.
Anthropic positions Team for organizations with 2 to 150 seats (user places) and Enterprise for larger deployments. On October 3, 2026, the official pricing page listed Team Standard seats at US$20 per seat per month with annual billing or US$25 monthly. Premium seats were US$100 per seat per month with annual billing or US$125 per seat per month with monthly billing. Enterprise was listed at US$20 per seat per month billed annually, plus usage priced by model and task.[5]
| Administrative area | Claude Team | Claude Enterprise |
|---|---|---|
| Positioning | Team use for 2–150 seats, with Standard and Premium seats | Large-scale deployment with finer usage and risk control |
| Identity and access | Central billing, company sign-in (SSO), and administration | Adds account synchronization, permissions by role, and organization or user spend limits |
| Audit and data controls | Team administration | Adds activity logs, audit-record retrieval, storage-duration controls, and restrictions on network access |
| Skill and plugin operations | Organization-wide skills and plugin marketplaces | Adds stronger group targeting, custom permissions by role, and scans for malicious content |
Enterprise scanning covers newly added or edited third-party skills and plugins. MCP servers and hooks are outside its scope, so connections and automated actions require separate safety checks.[11]
Claude Skills are available on Free, Pro, Max, Team, and Enterprise, and require code execution and file creation to be enabled. Team and Enterprise owners (people responsible for organization settings) can provision skills across the organization, while members can turn permitted skills on or off.[6]
A Claude plugin packages skills, connectors (connections to external services), and subagents (helper AIs assigned separate work units from the primary AI). Plugins are available on all paid plans and can be used across web chat, the Chat tab in Claude Desktop, Claude Cowork, and Claude Code. Bundled skills and commands (short instructions that invoke a specific task) work across those surfaces, while hooks and subagents run in Claude Cowork and Claude Code rather than chat.[7]
Team and Enterprise owners (people responsible for organization settings) can create plugin distribution lists (marketplaces) and distribute them across the organization. Enterprise adds group overrides for Required, Installed by default, Available to install, and Not available.[8]
When a required plugin syncs to Claude Code, its hooks, subagents, and MCP servers may run on a member’s computer. The administrative review therefore extends beyond the skill text to the executable components and connected systems in the package.[8]
OpenAI and Anthropic differ in distribution surfaces and execution scope
Both providers treat a skill as a reusable procedure and a plugin as a package for distributing several capabilities. The practical selection question is not which company uses the better word. It is which product fits the organization’s existing contracts, work surfaces, connected systems, and control requirements.
| Comparison axis | OpenAI | Anthropic |
|---|---|---|
| Skill focus | Instructions and resources used when relevant in ChatGPT and Codex | Specialized knowledge and workflows used when relevant in Claude |
| Typical plugin contents | Skills, MCP servers, and optional interfaces or hooks | Skills, connectors, subagents, commands, and hooks |
| Main distribution surface | Shared plugin list for ChatGPT and Codex | Claude’s feature setup page (Customize) and organizational distribution lists |
| Organizational control | Separate layers for plugins, MCP access, actions, and provider authorization | Unified management list, group distribution, and sync controls for skills and plugins |
| Development-surface caveat | Plugins are unavailable in the Codex IDE extension | Hooks and subagents do not run in chat |
These differences make OpenAI a natural comparison for work spanning ChatGPT, Codex, and existing service connections, while Anthropic provides a clear comparison for work spanning Claude, Claude Code, and role-specific plugins. This is not a ranking by feature count. The decision depends on which product can complete the target workflow with access limited to the data and actions needed for the task.
The 2026 updates shifted attention toward organizational distribution
Updates in 2026 moved both providers from isolated features toward organizational distribution. OpenAI introduced Codex plugins on March 25 and now provides a shared ChatGPT and Codex plugin list, an Enterprise export of the public plugin list in a spreadsheet-readable format, and layered controls for roles and connections.[2][4]
Anthropic expanded private marketplaces and enterprise controls on February 24, then announced the September 16 merger of Claude and the former Claude Cowork experience. Skills and connectors are intended to carry into the unified experience, but Team and Enterprise rollout timing remains staged.[9][10]
The implication is that skills and plugins are becoming units through which organizations distribute, update, and suspend business capabilities, not just personal convenience features. What Is SkillOps? covers maintenance after distribution. What Is Capability CI/CD? explains CI (automated validation) alongside approval, distribution, and reevaluation.
ROI needs a causal path from time saved to the P&L
Usage counts and time saved do not, by themselves, establish financial impact. Time saved is a useful early measure of change before financial results, but its P&L value is limited unless the released capacity becomes more selling activity, higher throughput, avoided outsourcing, or another measurable outcome.
The evaluation framework here aligns the measurement period and avoids counting the same benefit twice. The formula cannot be used if implementation and operating cost is zero:
ROI = (incremental gross profit + avoided cost - implementation and operating cost)
/ implementation and operating costDo not treat all additional revenue as benefit. Use incremental gross profit after discounts, the cost of goods, and service delivery costs. Implementation and operating cost should include more than AI subscriptions: skill and plugin design, connection development, permission review, evaluation, training, and maintenance also belong in the denominator.
The following sequence keeps early workflow measures separate from financial outcomes.
| Layer | Question | Example measures |
|---|---|---|
| Business outcome | Which P&L item should change? | Incremental gross profit, outsourcing cost, losses from refunds, customer cancellations, or missed deals |
| Business process | Which decision or step drives the outcome? | Proposals completed, response time, throughput, rework, approval delay |
| AI capability | Which part of the skill or plugin creates the change? | Procedure adherence, connection success, human correction, prohibited actions |
| Control | Who can use which data and actions? | Permission exceptions, approval rate, share of executions with records that can be checked later |
| Cost | What does continued operation require? | Seats, usage, development, evaluation, administration, training, maintenance |
Consider a sales proposal workflow. A skill can standardize how discovery notes become a proposal, how pricing is checked, which brand rules apply, and when a person reviews the result. A plugin can add permitted access to the CRM (customer and sales-opportunity management system), file store, and email system so the workflow can gather account context before preparing the draft.
Early workflow measures include proposal preparation time, correction count, and missing required fields. P&L measures include win rate (the share of proposals that become contracts) for a comparable set of opportunities, average discount, and incremental gross profit. Other initiatives also affect the result, so define the baseline, comparison group, measurement period, and exclusions before rollout.
Managing Enterprise AI Coding Costs applies the same separation to seat costs, variable usage, and business outcomes.
Conditions for choosing a skill, a plugin, a staged rollout, or no rollout
The first decision follows the missing capability and the operational risk.
| Condition | Choice | Reason |
|---|---|---|
| People can supply the inputs, and the main need is a consistent procedure and quality standard | Skill | The organization can improve a smaller unit without adding external connections |
| Work must retrieve data from several systems and be distributed to a department | Plugin | Connections, actions, distribution, and permissions move as one business unit |
| Search and drafting show value, but a wrong update would have material impact | Stage from read-only use | Quality, approval, and recovery can be evaluated before write actions |
| The business owner, baseline, success criteria, or data permissions are undefined | No rollout yet | ROI and accountability for failure cannot be evaluated |
Business or Team may be sufficient for a focused pilot. Account synchronization, auditability, storage-duration controls, and permissions by role can make Enterprise governance a selection requirement. Conversely, distributing a large plugin for work that needs no external connection can increase review and maintenance scope without improving the outcome.
Product selection should compare end-to-end completion, access limited to the data and actions needed for the task, evaluation evidence, departmental distribution, and change management rather than raw skill or plugin counts. With either provider, an organization still needs the option to limit rollout when the business outcome and control conditions are undefined.
Summary: stabilize the method with a skill and connect the business outcome with a plugin
Skills make organizational procedures and judgment criteria reusable. Plugins extend that unit to data connections, executable capabilities, and managed distribution. The target workflow determines how much reusable work can change a specific P&L item and what controls that scope introduces; feature count does not.
A skill is the appropriate first unit while people can supply the input and the priority is to stabilize the method and evaluation. The expansion trigger is when an installable package for distribution or connected systems becomes necessary for the workflow. This sequence separates individual time savings from organizational value by keeping incremental gross profit, avoided cost, and full operating cost in the investment decision. Pricing, supported surfaces, and administrative controls can change, so the contract review needs current official information and the organization’s permission requirements from the same review date.
References
- OpenAI, Skills & Plugins
- OpenAI, Plugin controls
- OpenAI, Pricing
- OpenAI, ChatGPT & Codex changelog, March 25, 2026
- Anthropic, Claude pricing
- Anthropic, Use skills in Claude
- Anthropic, Use plugins in Claude
- Anthropic, Manage plugins for your organization
- Anthropic, Claude Cowork and chat are now one Claude, September 16, 2026
- Anthropic, Cowork and plugins for teams across the enterprise, February 24, 2026
- Anthropic, Get started with skill and plugin scanning
For the latest release and update details, refer to the official product sites and documentation.