# What is agentic AI, and when is it worth using?

From answering a question to carrying out several steps: examples, benefits, drawbacks and the costs to watch.

Evidence: source-verified | Updated: 2026-10-25 | Edition: codex

Tool focus: work, codex, dot | General applicability: copilot

## Give it an outcome, not just a question

“Which emails need a reply?” asks for an answer. “Review this week’s messages, make a reply list and draft the three most urgent replies” asks for several steps toward a result.

**Agentic AI** means AI can choose steps, use available tools and adjust as it works. An **AI agent** is the system doing that work. It might search, read files, edit a document or use a connected app. A chatbot can contain agent features too; the labels are not a clean dividing line.

A fixed automation follows a set recipe. An agent has more freedom to decide what to do next. That can help when the work changes, but it also makes results less predictable.

## What would I use it for?

| Task | Work you could delegate | Result to review |
| --- | --- | --- |
| Back from holiday | Find relevant mail, identify open requests and draft responses | A catch-up list and reply drafts |
| Research a purchase | Find options against your requirements and compare source information | A shortlist with links and trade-offs |
| Prepare a team update | Read approved notes and assemble the current status | A draft update with unresolved questions |
| Maintain a website | Edit content or code, run checks and prepare a release | Changed files and a reviewable result |

These are task designs. The app still needs the appropriate tools and permission. A general Copilot chat can help plan these jobs; direct inbox actions require the relevant Outlook experience. [Our holiday email example](/codex/en/copilot-back-from-holiday/) explains that distinction.

## Pros and cons

**The upside:** less copying information between steps, a deliverable you can inspect, and the ability to adjust the plan when new information appears. It is especially useful when you can describe a result clearly but the steps take a while.

**The downside:** a wrong assumption can carry into later steps. Tools can fail, access can be missing, and retries can take time or consume usage. Checking the result is still work. For a quick rewrite, an ordinary chat may be easier than setting up an agent.

## What does it cost?

| Cost | What to look for |
| --- | --- |
| App access | Your existing plan may include some agent work; other features need a subscription. |
| Usage | Credits or included allowance can be consumed by a long task and its retries. API billing is separate from consumer subscriptions. |
| Connected services | Hosting, paid data or another app may have its own charge. |
| Your time | Connecting accounts, explaining the job and reviewing output are part of the cost. |

A concrete price reference: OpenAI lists Plus at **US$20/month**, with Work and Codex sharing usage, and Pro from **US$100/month** as checked on 25 October 2026. That is a subscription example, not a per-task agent price or an assurance of every feature. A custom agent has no universal monthly price: it depends on its model, tools and how often it runs.

For a first go, use an app you already have and a small task. If a tool offers a spend limit, set one before an open-ended job. “Research these three options once and stop at a draft” is easier to budget than “keep finding better options.”

## An example brief you can adapt

“Use these approved meeting notes to make a one-page team update. Include decisions, owners, dates and anything still unresolved. Check each item against the notes. Save a draft for me; don’t send it. If a needed detail is missing, ask me. Stop when the draft is ready.”

Here the result is clear, the source is limited, and there is a useful point to hand it back. This brief is an example, not a recorded run.

## Start with one responsibility

Try one bounded job, inspect what happened, then decide what else to hand over. If you want help choosing an OpenAI experience, see [ChatGPT, Work, Codex and Dot](/codex/en/chatgpt-work-codex-dot/). For a task you can try with ordinary chat, [organize your meeting notes](/codex/en/meeting-notes-to-actions/).

## References

- [Anthropic: building effective agents](https://www.anthropic.com/research/building-effective-agents)
- [OpenAI: choose Chat, Work or Codex](https://learn.chatgpt.com/docs/use-chatgpt)
- [OpenAI: meet dots](https://learn.chatgpt.com/docs/dots)
- [OpenAI: Work and Codex pricing](https://learn.chatgpt.com/docs/pricing)
- [Microsoft: chat with Copilot in Outlook](https://support.microsoft.com/en-us/outlook/copilot-outlook/chat-with-copilot-in-outlook)
- [Microsoft: triage email with Copilot](https://support.microsoft.com/en-au/outlook/triage-email-with-microsoft-365-copilot-in-outlook)