Field note · Interface design

Not everything should be a chatbot.

We keep reaching for the same answer — wrap it in chat, hand it to an agent — no matter the question. But some work still wants a screen you can point at. Some should just run in the background. And some needs an interface built for one decision, in front of one team, then thrown away.

Direct GUI Headless agent Generative UI

3x3 journey matrix

Same columns. Different winners.

Each row is a workplace journey a CMO or executive team would recognize. The winner changes because the cognitive load changes.

Direct GUI Best when the object is clear and the action is small.
Agent / headless Best when the work is long, multi-step, and mostly procedural.
Generative UI Best when teams need to see tradeoffs from several angles.
Best fit Workable Poor fit
Direct GUI

Works when the user can point to the thing.

Strong at
Editing, approving, selecting, scanning, and acting on a known object.
Breaks when
The user has to stitch together many systems, policies, or time periods.
Executive test
If a person can say "change this line," give them the line.
Token profile
Near-zero AI spend per action once the screen exists.
Agent / headless

Works when the process matters less than the result.

Strong at
Running tool chains, checking data, retrying failed steps, and keeping evidence.
Breaks when
The user needs to steer tradeoffs visually or compare several futures at once.
Executive test
If the work is "do the analysis and show exceptions," hide the machinery.
Token profile
High spend is acceptable when it replaces hours of tool work.
Generative UI

Works when the view needs to change with the decision.

Strong at
Making a temporary planning surface around teams, constraints, and tradeoffs.
Breaks when
It burns tokens to create a panel but still makes people read and judge more output.
Executive test
If each function needs a different lens on the same plan, generate the lens.
Token profile
Expensive if every prompt rebuilds context, layout, and data bindings.

Token economics

Generated UI has a meter running.

It spends tokens before the user acts: to understand the request, summarize the context, choose components, bind data, and explain the result. That cost is worth paying only when the view saves more executive attention than it consumes.

Direct GUI Higher design cost up front, low AI cost during repeated use.
Agent / headless High token use can be justified when it replaces real process work.
Generative UI Token spend must buy cognitive compression, not just a prettier surface.
Known object

Edit the launch brief.

The document is already the workspace. The fastest path is to touch the thing directly.

Hero: Direct GUI
Direct GUI Best fit
Low overhead, high control
  1. IntentThe executive knows exactly which line needs to change.
  2. ActionEdit the line, leave a comment, approve the doc.
  3. ResultNo prompt, no summary, no extra interpretation layer.
Agent / headless Poor fit
AI adds reading work
  1. IntentThe user turns a visible edit into a conversation.
  2. ActionThe agent creates options the user must now inspect.
  3. ResultThe work gets slower unless the edit was genuinely unclear.
Generative UI Workable
Useful only if review is complex
  1. IntentThe system creates controls around a simple document action.
  2. ActionThe user reviews tone, risk, and policy checks.
  3. ResultHelpful for a sensitive launch, overbuilt for a line edit.
Analysis engine

Run the market analysis.

The value is in the loop: pull data, test assumptions, keep evidence, surface exceptions.

Hero: Headless agent
Direct GUI Poor fit
Too much process on screen
  1. IntentThe CMO wants a defensible answer, not another dashboard tour.
  2. ActionTeams export CRM, ads, web, revenue, and survey data by hand.
  3. ResultThe interface exposes every step but does not do the work.
Agent / headless Best fit
Human reviews assumptions, not clicks
  1. IntentAsk for the analysis outcome and the standard of proof.
  2. ActionThe agent runs tools, retries failed steps, and keeps an audit trail.
  3. ResultThe executive reviews the few calls that actually need judgment.
Generative UI Workable
Good for review, not for the whole run
  1. IntentThe system builds a useful surface for inspecting results.
  2. ActionThe user still has to drive the analysis loop.
  3. ResultGreat as a review layer, weaker as the engine.
Cross-team decision

Align roadmap and marketing.

Product, Marketing, Sales, and Success need different views of the same priorities.

Hero: Generative UI
Direct GUI Workable
The SaaS object is too rigid
  1. IntentThe team needs one plan across roadmap, standup notes, and launch calendar.
  2. ActionFixed filters show one department's model at a time.
  3. ResultExecutives do the cross-functional comparison in their heads.
Agent / headless Poor fit
The answer is too compressed
  1. IntentThe executive can describe the planning problem naturally.
  2. ActionThe agent picks a recommendation without showing enough of the map.
  3. ResultLeaders ask follow-up questions because they cannot steer the tradeoff.
Generative UI Best fit
The interface adapts to the planning question
  1. IntentBring weekly standup, roadmap, and marketing priorities into one decision.
  2. ActionThe app generates team lenses, conflicts, dependencies, and next best moves.
  3. ResultExecutives can steer the plan without forcing everyone into one SaaS shape.

Operating principle

AI should compress work, including its own cost.

Direct GUI wins when the user already knows what to touch. Agents win when the process is long and tool-heavy. Generative UI wins when leaders need a temporary shared surface for a messy decision, and when the token spend buys back attention.

Known object Do not wrap a simple edit in AI ceremony.
Analysis loop Let the agent run, then inspect the evidence.
Team planning Use generated UI when the changing view is worth the meter.