Decide: What to Delegate, Which Tool, Which Mode
This content was developed with AI assistance and is regularly reviewed for accuracy.
Snapshot: October 2026. Model names, prices, and product features change monthly; What's New in AI is the maintained chronology.
The most skipped step in working with AI happens before you type anything: deciding whether the assistant should do this task, how much of it, with which tool, and in which mode. Microsoft's 2026 Work Trend Index found that the most effective AI users were far more likely than everyone else to pause and make that call up front (53% versus 33%), and also more likely to deliberately do some work without AI to keep their own skills sharp. This lesson gives you a fast version of that pause.
Learning focus
- Ask "should AI do this?" rather than "can AI do this?" and answer it in under a minute
- Choose how much to delegate, from a first draft to an unattended scheduled task
- Pick a tool by where your data lives and what tier you are on, not by brand loyalty
- Match the task to a mode: fast, thinking, research, or agent, and know when a plain automation beats all four
Should AI do this at all?
Anthropic's AI Fluency framework puts the question plainly: you are not asking whether AI can do the task, but whether it should, and if so how much. Four quick tests settle most cases.
Can you check the result? If you could not tell a good answer from a plausible wrong one, delegating the task means trusting blind. Either narrow it to something you can verify, or keep it.
Does the work need to be yours? Learning a skill, building a relationship, and exercising judgment about people are the clearest examples. Research summarized by Ethan Mollick in 2026 found that students who used AI as support learned, while those who handed the whole task over did not. If the point of the task is that you come out knowing something, do it yourself and use the assistant to quiz or critique you afterward.
What does a wrong answer cost? A draft tweet costs a minute. A fabricated citation in a court filing has cost lawyers their licenses. Scale your verification to the stakes, and keep anything irreversible or customer-facing on the human side until you have a checking routine you trust.
Is it repeatable? A task you do every week is worth a brief, a saved project, or a scheduled task. A one-off may be faster to just do.
How much to delegate
Delegation is a dial, not a switch. Pick the lowest setting that gets the job done, then move up as your checking routine proves out.
- Draft for me. The assistant produces a first version; you edit and own it. Lowest risk, works on any tier.
- Do it and show me. The assistant completes the task, such as building the spreadsheet or booking the meeting, and waits for your approval before anything leaves your account.
- Do it and tell me. The assistant acts inside limits you set and reports afterward. Appropriate once the same task has gone right several times under approval.
- Do it on a schedule. A recurring task runs unattended. Reserve this for reversible work with a narrow set of permissions; the Protect lesson explains why.
Vendors' own guidance follows the same ladder: actions that are hard to reverse, affect shared systems, or could be destructive should ask you first.
Which tool
Choosing between ChatGPT, Claude, Gemini, and Copilot by raw capability is mostly a waste of time in 2026; the frontier models trade places monthly, and the AI Model Comparison page tracks that. Four questions matter more than the leaderboard.
Where does your data already live? Copilot inside Microsoft 365, Gemini inside Google Workspace, and each assistant's connectors into Gmail, Drive, Slack, and the rest decide how much context you can supply without copying and pasting.
What has your workplace approved? The same sentence can be safe in an approved business tier and a policy violation in a personal free account. If your employer has an approved tool, that answer overrides the others.
What output do you need? Image and video generation lives in ChatGPT and Gemini, not Claude. Repository-scale coding lives in dedicated coding agents. The AI Tools & Platforms page maps output types to tools.
Which tier are you on? This decides which modes you can use at all, so it gets its own section below.
Use two assistants if you like, but settle on one as your main workspace so its memory, projects, and connectors build up. Switching every week resets all of that.
Which mode
Every mainstream assistant now offers roughly the same four ways of working. Picking the right one matters more than any wording choice.
| Mode | What it does | Use it when | Cost and time |
|---|---|---|---|
| Fast (default or low effort) | Answers from memory with light reasoning; may skip search | Quick drafts, rewording, questions you can check instantly | Seconds, cheapest |
| Thinking (high effort) | Reasons longer, more likely to search and read files | Analysis, tricky code, anything with several constraints | Longer waits, more usage |
| Research (deep research) | Plans and runs many searches, reads sources, writes a cited report | Comparing options, literature reviews, market or vendor scans | Minutes to tens of minutes; usually paid tiers |
| Agent (work or agent mode) | Plans a multi-step task and acts through your files and connected apps, pausing for approval | Building a deck from a folder, triaging an inbox, filling a tracker | Minutes to hours; paid tiers; uses limits fast |
Two rules of thumb. First, effort is the primary control for trading off quality against time and cost, and the middle setting is right for most tasks; raise it when the task has several moving parts or when you need the model to search rather than recall. Second, for a large job, ask for a plan before the work: have the assistant interview you, write down what it intends to do, and only then start. Anthropic's Claude Code guidance has a useful cut-off: if you could describe the change in one sentence, skip the plan.
The names differ by product. In October 2026, ChatGPT's agent mode is ChatGPT Work, available on paid plans but not the free or Go tiers; Anthropic's agent, formerly Claude Cowork, is now simply part of Claude and rolling out to Pro and Max; Google lists Agent Mode as an early-access feature on paid Gemini tiers. The Agents in the Apps You Already Pay For page keeps the current list.
Free versus paid
A free account is enough for the whole core of this track, with one correction to the old advice: you are not getting the best model. In October 2026 free tiers run the small, fast model of each family: GPT-6 Luna in ChatGPT, Sonnet or Haiku in Claude, and Flash-Lite only in Gemini since October 9 (reported). ChatGPT's free and Go tiers have shown ads since February 2026. These models are capable, but when a free-tier answer disappoints, the fix is sometimes the tier rather than your brief.
The $20 tier has become the standard across all the major apps. Beyond the larger model, it typically unlocks the research mode, the agent mode, scheduled tasks, and higher usage limits, which is why the research and agent exercises later in this track note when a paid plan is required. The Free AI Tools Directory lists what each free tier currently includes, and the AI Model Comparison page has the paid ladders.
When a plain workflow beats an agent
Agents are the newest option, which does not make them the right one. If a task is the same every time, must leave an audit trail, or needs a named person to approve each step, a deterministic automation (Zapier, Make, n8n, Power Automate, or a simple scheduled script) is cheaper, more predictable, and easier to explain to a compliance officer. Reviewers in 2026 consistently report that consumer agents break on website logins, long unattended runs, and anything that must be repeatable and auditable.
Use an agent when the steps vary, the inputs are messy, and judgment is needed along the way. Use a workflow when you can draw the steps as a fixed diagram. Many real systems combine the two: a workflow that triggers on a schedule and hands one messy step to a model. The Build track's Workflows vs Agents page and the Loops recipe When a Workflow Beats a Loop go deeper.
A one-minute decision checklist
Run through this before any task you have not delegated before:
Task: [what I want done]
Checkable? [yes / only partly / no]
Needs to be mine? [learning, relationship, judgment about people: yes / no]
Cost if wrong: [trivial / annoying / serious / irreversible]
Delegation level: [draft / do and show / do and tell / scheduled]
Tool: [the one with my data and my employer's approval]
Mode: [fast / thinking / research / agent] at effort [default / high]
Or should this be a plain automation? [yes / no]
Most of the time this takes thirty seconds. The value is the cases where it stops you: the irreversible task you were about to schedule, or the research question you were about to ask a fast model that would have answered from memory.
Key takeaways
- Ask "should AI do this, and how much?" before asking how to phrase it
- Delegate on a ladder: draft, do and show, do and tell, scheduled, and climb only as your checking proves out
- Choose a tool by where your data lives, what your workplace approved, and which tier you are on
- Match the task to fast, thinking, research, or agent mode, and raise effort when the task needs search or many constraints
- A fixed, auditable, repeatable task is a job for a workflow, not an agent
Next steps
Continue to: Brief: How to Ask AI for What You Want
- See what each free tier includes right now in the Free AI Tools Directory
- Learn what the agent modes inside ChatGPT, Claude, Gemini, and Copilot can do in Agents in the Apps You Already Pay For