Chat, deep research, agent or automation: which one does this job need?

The idea

All three big assistants now offer more than a chat box. Each one can research a question across many sources, and each one can take a goal and do the work in your apps. Google's agent is only for personal accounts on Google AI Pro or Ultra so far. That gives you four kinds of help:

• Chat. You ask, it answers, in one or two turns. Good for a quick question, a rewrite, a first draft from notes you give it.
• Deep research. It plans a set of searches, reads many sources (web pages, and your own files if you add them) and hands back a report with its sources. Its job is to read and report, not to do things in your apps.
• Agent. It takes a goal, works out the steps, uses your apps, files or a browser to do them, and hands back finished work: a deck, a sorted inbox, a booked meeting, usually with your approval for the steps that matter. In this Field Guide's words: an agent is the same kind of model you already chat with, given three things: a goal, a set of tools, and permission to keep going.
• Automation. A fixed routine that runs on a schedule or when something happens: every Monday at 8, or each time a certain email arrives. The same steps every time.

What changed in 2026: fewer buttons

Two of the three joined their chat and their agent this year.

• Claude, 16 Sep 2026. Cowork, Claude's agent for office work, became part of ordinary Claude chat. In Anthropic's words: "Users no longer have to choose which mode a task belongs in." You type a quick question or hand over a big job in the same message box, with your projects, connectors and skills carried over. It reaches Pro and Max first, then Team and Free; Enterprise admins get at least 30 days' notice. If you still see Chat and Cowork as two separate choices, the change has not reached your account yet, and you pick Cowork for a task.
• ChatGPT, 9 Jul 2026. One desktop app now holds Chat, Work and Codex. Chat and Work sit behind one toggle, and Work is the agent. Codex, for software, is a separate view.
• Gemini. Chat and Task sit side by side below the text box, and Task is the agent (Google's earlier pages call it Gemini Spark). At a company on Google Workspace, Gemini mostly lives inside Gmail, Docs, Sheets, Slides, Drive, Chat and Meet instead.

So in Claude, "which mode?" has mostly become "how big is the job?" The three-question test below still decides what you ask for, what it may touch and how much you check.

Why the choice still matters

Picking the wrong one costs you. Agent work is slower and uses more of your allowance: Anthropic says task work "consumes limits faster than Chat", and OpenAI's own split is "Chat for questions and conversation, Work for research and finished deliverables" (both as of Oct 2026). Go the other way and you get a chat answer from memory when the job needed twenty current sources.

The research firm Gartner put the rule simply in June 2025: use "AI agents when decisions are needed, automation for routine workflows and assistants for simple retrieval". The same release forecast that over 40 percent of agentic AI projects will be cancelled by the end of 2027, because of rising costs, unclear business value or inadequate risk controls. That is a forecast, not a measurement, but its point holds: many jobs do not need an agent.

The habit to build is the pause before you start. In Microsoft's 2026 Work Trend Index (a survey of 20,000 AI users in 10 countries, February to April 2026), the most advanced group of users, which the report calls Frontier Professionals, were more likely than others to say they "pause before starting work to decide what should be done by AI versus a human": 53 percent against 33 percent. It is a self-reported survey by a company that sells AI tools, so read it as a sign of a habit, not proof of a result.

The three-question test

Ask these three questions about the job, in this order:

1. Does it need decisions along the way? Will the next step depend on what it finds? Writing a reply from your notes: no. Sorting forty emails and deciding what each one needs: yes.
2. Does it touch your apps, files or the web? Only what you paste or attach: no. Reading many web pages, opening files in your drive, sending, booking or editing: yes.
3. Does it repeat on a schedule, or start on an event? Once: no. Every Friday, or whenever a message arrives: yes.

Then read the answers like this:

A file you attach to a chat is still chat. The question is whether it has to go out and get things, or do things, on its own.

Ten everyday jobs to sort

Cover the answers, run each job through the three questions, then check.

1. Explain what one term in a report means, from the sentence you paste in.
2. Make a paragraph of your own shorter and friendlier.
3. Compare how three named project tools handle guest access, from their own help pages, with links.
4. Summarise what industry reports published this year say about hybrid work, with links to each.
5. Turn your bullet notes into a first draft of a team update.
6. Sort this week's unread customer emails by topic and draft a reply to each, for you to approve.
7. Pull last quarter's figures from three spreadsheets in your drive into one deck with charts.
8. Find a free hour next week for six people and draft the invite, without sending it.
9. Every Monday at 8:00, summarise last week's messages in the team channel.
10. Each time an email with "Weekly numbers" in the subject arrives, write you a three-line summary.

Answers

1. Chat. One answer, nothing to fetch.
2. Chat. You supply the text; no decisions along the way.
3. Deep research. Many sources to read and cite, nothing to change.
4. Deep research. Same reason. Then open the sources yourself.
5. Chat. Your notes are the source.
6. Agent. It decides what each email needs and works in your inbox. Keep sending in your hands.
7. Agent. Several files, steps and judgement about what goes in.
8. Agent. It reads calendars and makes choices. "Without sending it" is the brake: set it in the tool's approval setting too.
9. Automation. Same steps, on a schedule. Review the first few runs.
10. Automation, started by an event. It reads email that may come from outside, so keep it to read and report: it may summarise for you, but not send, pay or delete. In ChatGPT, a Scheduled Task can start on a new Gmail message.

Common mistakes

• Sending a quick question to an agent. It is slower and uses more of your allowance, and a quick answer never needed it.
• Asking chat for something that needs current sources, then trusting an answer it gave from memory. If it needs many sources, use deep research and open them.
• Thinking an agent is smarter. An agent isn't a smarter AI. It's the same kind of AI, on a longer leash. And holding that leash is your job.
• Looking for the Cowork button. Once Claude's September change reaches your account there is none: hand the task over in an ordinary Claude conversation.
• Opening a code tool for office work. Claude Code, Codex and Antigravity are for building software. Documents, research and spreadsheets belong in ordinary Claude, ChatGPT Chat or Work, or Gemini.
• Picking the heaviest model every time. Fable 5.1 and GPT-6 at Pro are for the hardest work and use your limits fastest. Start with the default.
• Scheduling a task that sends, buys or deletes. Nobody is watching when it runs. Keep scheduled work to read, report and draft.
• Assuming every tool has every mode. As of Oct 2026, Claude's Research is on Pro, Max, Team and Enterprise; ChatGPT's Scheduled Tasks are not on Free, and tasks that start on an event are not on Free or Go; Gemini's Task needs a personal account on Google AI Pro or Ultra. Your admin may also have turned a mode off.
• Automating a routine nobody has written down. If you cannot list the steps, it is not ready to be an automation.

What to check before you trust it

• Before you start: that the mode is on your plan and allowed by your workplace, and that you are in the right account (work, not personal).
• Before an agent runs: what it is connected to, whether it can change things or only read, and that the approval setting makes it ask before anything you cannot undo. Set that in the tool, not in your message.
• After deep research: open the sources yourself and check that each one says what the report claims. The next lessons show how.
• After every scheduled run: read the output. A task that ran last week can go wrong this week.
• Whatever the mode or the model: you stay responsible for what goes out.

As of: product steps as they were on the vendors' own pages on 16 Oct 2026.

Every product name, plan and step above comes from Neosaipien's platform reference, checked on these vendor pages on 16 Oct 2026:

• Anthropic, Claude Cowork and chat are now one Claude (16 Sep 2026)
• Anthropic, Claude Cowork product page and FAQ
• Anthropic, Cowork overview
• Anthropic, Run tasks in the background with Dispatch
• Anthropic, Claude Cowork and chat are one Claude (help page; Research without a web search switch; Dispatch not offered to new users)
• Anthropic, Unified Claude in Claude Desktop (admin docs, 8 Oct 2026)
• Anthropic, Build live dashboards and animate explainers with Claude (8 Oct 2026)
• Anthropic, Use research on Claude
• Anthropic, Change the model, effort, and thinking settings
• Anthropic, Choosing the right Claude model (Claude Academy)
• Anthropic, Claude plans
• OpenAI, ChatGPT Work and Codex
• OpenAI, ChatGPT release notes (9 Jul 2026, 25 Aug 2026)
• OpenAI, Deep research
• OpenAI, ChatGPT agent
• OpenAI, GPT-6 and other models in ChatGPT
• OpenAI, ChatGPT plans
• Google, Manage your tasks and workflows in Gemini Apps
• Google, Use Deep Research in Gemini Apps
• Google, About Gemini models
• Google, Google AI plans
• Google, Get started with Google Workspace with Gemini
• Google, Gemini Apps release notes (13 Oct 2026, Antigravity in the Mac app)
• Gartner, press release, 25 Jun 2025
• Microsoft, 2026 Work Trend Index (5 May 2026)