Modes of AI work

Modes of AI work is a companion framing to the AI scale. The scale — chat, context, automation, colleague — describes the natural progression most people follow as they mature with AI: starting with conversational use, then building custom assistants, then designing automations, eventually delegating work to autonomous loops. That progression is real and worth naming, because it tells someone with no prior framing where they currently sit and where they could go next. But the progression is additive, not substitutive: someone operating at the colleague level still uses chat constantly, still has custom assistants for ongoing work, still runs automations. Each level you reach gets added to your repertoire; you don't climb away from earlier ones.

The one-line reconciliation: the scale is for your skills, not your workflows. Your skills climb the scale once — each mode teaches what the next one needs. Your workflows never climb it — each workflow gets the mode that fits the work.

The transport metaphor

The enduring metaphor for the modes is transport: chat is walking, context is your bike, automation is the train, colleague is the plane.

The metaphor earns its place on three counts:

It dignifies the basic mode. Nobody apologises for walking. The most common failure pattern around the scale is shame — practitioners apologising for "still just using ChatGPT," as if chat were something to outgrow. Walking is not something you outgrow. You learned it first, you use it every day, and no amount of driving makes it obsolete.
The trip picks the mode. Nobody drives to their kitchen; nobody walks to Berlin. Mode choice follows from the trip — distance, frequency, cost of getting it wrong — not from the traveller's ambition or sophistication. Same for workflows.
The control and investment gradients map. Walking: all you, zero setup. Bike: your own vehicle, maintained by you, faster on your regular routes. Train: someone else drives, it runs on a schedule whether you remembered or not. Plane: full delegation, highest trust, most vetting before you board. Investment and delegation both rise across the modes exactly as they do across the scale.

The metaphor rejected on the way here: gears. "Stuck in first gear" and "kick into high gear" pejorate the low end — the idiom reintroduces exactly the shame the framing exists to remove.

The four modes (also the four levels)

Chat — walking

You ask, AI answers in real time. Conversational, ephemeral. You bring the question, the AI brings the response, and you stop when the answer is good enough or your question changes. Chat is the right mode for one-off questions, exploratory thinking, debugging unfamiliar problems, and any work where the next move depends on what AI just said. The investment per use is near zero; the same is true of value retained between uses, because nothing persists.

People at the chat level use AI mostly here. People at higher levels keep using chat for the work that fits this mode — and there's a lot of it.

Context — your bike

Your context — role, style, reference material, constraints — gets loaded once. The assistant arrives knowing the things you'd otherwise have to re-explain. You drive each conversation, but the assistant is a persistent collaborator you keep returning to for a category of work: drafting emails in your voice, editing against your style guide, planning content against your brand. Investment goes up because you maintain the loaded context; value also goes up because the assistant gets sharper with use — like a bike fitted to its rider.

People at the context level have built one or more custom assistants. They still use chat for ad-hoc questions; the custom assistants handle their recurring categories of work.

Automation — the train

The AI runs without you initiating each execution. The trigger is a schedule (every Friday morning), an event (when an email matching a filter arrives), or a manual on-demand kick (you say "run the weekly digest" and walk away). You review the output afterwards. The mode is a saved instruction plus a recurrence plus an output destination. Investment is in the design and the source bounding; value compounds through the schedule itself, because — like the train — it runs whether you remembered or not.

People at the automation level have one or more workflows running without their initiation. They still use chat and custom assistants for everything that doesn't fit the automation mode.

Colleague — the plane

The AI evaluates its own output against the requirements before involving you, regenerates if anything's missing, and only presents the result when it's confident you'll accept it. The mode is generation plus self-evaluation plus refinement. Investment is in articulating the success criteria the AI checks itself against; value comes from delegating both execution and quality control. This is the highest-leverage mode but also the most demanding to design — the AI needs to know what "done" looks like before it starts, the way a flight is vetted before anyone boards.

People at the colleague level have at least one autonomous loop running. They keep using chat, custom assistants, and automations for the rest of their work, because the colleague mode is overkill for most tasks.

The scale is for your skills, not your workflows

The most common misreading of the scale is that you climb it and leave the earlier modes behind. The truth is that maturity is measured by how many modes you have working in concert, not by which one you've reached. A founder who built three custom assistants a year ago and a scheduled automation last month has accumulated capability — but they still answer ad-hoc questions in chat all day, because that's what chat is for.

The reason the modes get named in scale order is that the learning progression is real. Chat doesn't require any setup; building a custom assistant requires having context worth loading; designing an automation requires the human-review patterns developed at the context level; designing an autonomous loop requires understanding the verification standards that make automation safe. Each mode builds on capabilities developed at the previous one. Skipping ahead is possible but rarely advisable for work the user hasn't already done at lower levels. You learned to walk before you learned to drive.

But once accumulated, all four modes coexist. The right design question for any new workflow isn't "what's my current level" — it's "which mode fits this work." The trip picks the mode.

Choosing a mode for a specific workflow

After a workflow has been identified and decomposed, the mode for it follows from a few questions about the work itself.

How often does this happen, and on what trigger? If the trigger is "when something occurs to me," chat is the right mode — there's no recurring pattern to automate. If the trigger is "every Friday" or "when a new ticket arrives," automation fits. If the trigger is "I notice this category of work and want to do it well each time," context (a custom assistant) fits.

Who initiates each execution? If the human starts each run, the mode is chat or context. If the trigger is external (time, event), the mode is automation. If the AI is supposed to start, evaluate, and finish before showing you anything, the mode is colleague.

Where does the output land? Chat output stays in the conversation. Context output is the conversation plus whatever the human copies out. Automation output goes to a destination the human checks — a folder, a channel, a doc. Colleague output is the final artefact, with the iteration history usually discarded. Choosing the destination is part of choosing the mode.

What's the failure cost of one bad run? High-cost single runs — legal advice, client communication, irreversible actions — push toward chat or context, where the human reviews each step. Low-cost single runs — a digest, a draft, a list of suggestions — make automation or colleague modes safe to design. The cost is what determines whether the human stays in the loop per-run or per-batch.

How this changes the design conversation

Naming the mode early focuses the design questions on what actually matters. Each mode has different design considerations:

Chat: how do I phrase the question so the answer is useful? What context do I provide? Where does the AI fall short on this kind of question, and how do I verify?

Context: what reference material does the assistant need loaded? What context stays the same across all conversations, and what gets supplied per-conversation? How does the assistant get better with use?

Automation: what sources does it read, and are they bounded? Where does the output land? What triggers a notification? What happens if a run fails silently?

Colleague: what are the success criteria the AI checks against? How many iterations before it gives up? What does it show the human, and what does it discard?

Naming the mode after a workflow has been identified produces concrete follow-up questions. Naming a level without a workflow attached produces vaguer reflection, which is useful for orientation but not for design.

Terminology note

This framing was called "shapes of AI work" until 2026-07-31, when it was renamed to modes (plain, non-hierarchical, needs no explanation) and the transport metaphor was adopted as the standing imagery. When drafting new content, use "modes" — "shapes" only survives in already-published material.

ai-workflow-redesign — Methodology for identifying which workflows in someone's life are good candidates for which mode
scheduled-automation — Detailed pattern for the automation mode, including cross-tool implementation
example-image-generation — Concrete example of a colleague mode with self-evaluation before human review
example-email-drafter — Concrete example of a context mode with loaded style guidelines
agent-design-principles — How mode choice affects agent design, especially the "use the dumbest agent that can do the job" principle for each mode