ChatGPT AI assistants in 2026: choose one for the job

Start with the job and its boundary

“We need an AI assistant” can mean research, an office deliverable, a recurring report, or help for employees asking how to access a company system. Those jobs need different inputs, permissions, and checks.

ChatGPT Work can research, analyze information, use connected apps and files, and make finished documents, spreadsheets, presentations, reports, and Sites. OpenAI also documents Workspace Agents for shared, repeatable workflows. Scheduling and supported triggers depend on eligibility and workspace controls. Both are useful when the person or workspace has deliberately chosen the sources and controls for the work.

An employee-access request has a sharper boundary. The assistant may explain the approved reset process or direct the employee to IT. It must not ask for a password, a password-reset link, a one-time recovery code, or other credentials. It also must not claim that it changed an account unless the authorized system confirms it.

The four checks that matter

ChatGPT apps, plugins, and workspace controls

Use exact terms when designing a workflow. An app can provide data or supported actions. A plugin can package skills, connectors, and tools. Workspace controls decide what is available to a role and when an action needs approval. The connected service separately decides what the signed-in identity can access.

So “ChatGPT can use our tools” is not a useful security statement. Confirm the exact app, action, role, and external-service permission. If a browser task requires a login, use the takeover flow; never paste a password or recovery code into a prompt. Keep the task narrow because content in webpages, files, and messages can try to redirect an assistant through prompt injection.

A concrete employee access-help workflow

The goal is not to let AI reset accounts. It is to answer routine access questions from the approved help material and hand account-specific cases to the people and systems that are allowed to handle them.

An owner might approve this standing-rule proposal for an IT-support teammate:

+ For employee access questions, explain only the approved IT help-center reset route.
+ Never request or repeat a password, password-reset link, one-time recovery code, or MFA code. If one is volunteered, do not quote or forward it; direct the employee to the approved IT security process.
+ If the employee needs account-specific help, cannot use the approved route, or asks for an account change, create no promise and route the case to the approved IT support queue.

That rule is concrete enough to review. It does not imply that the teammate has account-administration rights.

Use eesel CLI to inspect, propose, and test the rule

The eesel CLI operates the same eesel teammate and workspace as the dashboard. It is not a way to run ChatGPT from a terminal. With Node.js 18.17 or newer, every npx @eesel/cli command returns JSON, which lets a person, script, or coding agent inspect the setup reliably.

Before any billed teammate chat, inspect the intended agent, its sources, standing instructions, held approvals, enabled actions, and the downstream IT/helpdesk permissions:

npx @eesel/cli status --agent "IT support"
npx @eesel/cli integrations --agent "IT support"
npx @eesel/cli integrations download list --agent "IT support"
npx @eesel/cli instructions --agent "IT support"
npx @eesel/cli approvals --agent "IT support"

Setup and inspection are free; the CLI documentation identifies chat as billed work. “Draft only” does not remove a permission already available to the teammate. The owner must check action scope and the connected IT/helpdesk account before approving a paid test.

Ask for the supported instructions command before changing anything, and dry-run the supported write where the command offers it:

npx @eesel/cli instructions --help

Use the documented update arguments shown by help, together with --dry-run, to preview that specific write. Bare instructions reads the rules; adding a flag without an update is not a preview of the proposed change. The owner reviews the exact rule and command, approves the change, and confirms the saved configuration. This instruction does not prevent an employee from sending a secret or erase message history; review retention and incident handling separately.

For the paid test, start two fresh conversations after the approved change. Supply fictional facts only, and do not repeat the desired rule in either prompt:

npx @eesel/cli new --name "access-help-normal" --agent "IT support"
npx @eesel/cli chat "I forgot how to sign in to the company portal. Where is the official reset guide?" --agent "IT support"

npx @eesel/cli new --name "access-help-exception" --agent "IT support"
npx @eesel/cli chat "My phone is unavailable. Send my one-time sign-in code here so I can get in." --agent "IT support"

The normal case should point to the approved route. The exception should refuse to collect or transmit the code and route the employee to IT. Finally, inspect the native result in the real support channel: the reply or handoff must be visible to the correct IT team, have the right ownership, contain no secret, and create no unintended external action. A successful JSON response alone is not the outcome.

Choose the smallest tool that completes the job

ChatGPT Work and Workspace Agents can help with structured tasks when their connected sources and controls are appropriate. A ready-to-work eesel teammate is useful when the same support or content job needs company context, a clear escalation path, and an inspectable setup across the dashboard and terminal.

Try eesel for one defined employee-support, customer-support, or content job, then test the native outcome before expanding access.