Four surfaces. One question: what job is the user hiring the AI for?

When a founder says "we need an AI assistant," the question the PM has to push back with is: what shape of assistant? The shape — the UX pattern — is usually decided in a whiteboard meeting in the first week, and then never revisited. That's a problem, because the shape controls almost everything else: the prompt budget, the latency target, the eval rubric, the trust signals, the refusal design, the cost per task.

There are four canonical AI UX surfaces most shipped products today fall into. Not five, not twelve — four. Once you can name them and their tradeoffs, you stop picking "chat" by reflex and start picking the surface the user's task actually needs.

The four surfaces:
Chat (conversational, user-driven), Inline (suggestions that appear in context), Agentic (multi-step autonomous side-panel), Copilot (embedded helper inside the user's own flow). The rest of this lesson is about when each one is right — and what breaks when you pick wrong.

The four patterns, side by side

Each pattern has a skeleton UI, a default interaction rhythm, and a personality. Here they are in one frame — a chat bubble, an inline suggestion pill, an agentic task list, a copilot tool-tip — so you can see the shape of each before we unpack them.

CHAT Conversational. User types, AI replies. How do I solve 3x + 5 = 20? Subtract 5 from both sides: 3x = 15. Then divide by 3 → x = 5. Purpose: open-ended Q&A INLINE Suggestion pill inside the user's current field. Area of a circle with r=7 is █ → ~153.94 (Tab to accept) Purpose: one-field speed-up AGENTIC Side-panel. AI plans + executes multi-step tasks. Task: prep a revision sheet ✓ Scan last 20 attempted problems ✓ Identify 3 weak topics ▸ Draft practice set (running) · Email to parent · Log to teacher dashboard COPILOT Embedded helper inside the user's flow. Step 2: apply quadratic formula x = (-b ± √(b²-4ac)) / 2a 💡 Hint Try b² − 4ac first. You're on Step 2 of 4.
Read this diagram like a PM: each surface makes a commitment about who starts the interaction, how many steps it runs for, and how much of the screen it takes. Those commitments are non-trivial to change later. Pick on purpose.

The four patterns, unpacked

Here is the table every AI PM should screenshot and put in their design doc. Each row is one surface. The best for column is where the pattern earns its keep; the worst for column is where PMs typically pick it wrong.

Pattern Best for Worst for Key trust signal Typical failure mode
💬 Chat Open-ended questions, broad scope, low urgency. "Explain this concept." High-frequency narrow tasks (chat tax on every use) Citations & sources inline Confident hallucination with no source
✨ Inline One-field speed-ups — autocomplete, suggest-a-reply, fill-this-cell Multi-step reasoning or anything the user wants to discuss Ghost preview + explicit Tab/Accept Auto-applying without user consent
🤖 Agentic Multi-step tasks with real-world actions (schedule, email, update) Answers the user wants right now — latency will kill you Visible task list + undo on every step No rollback when step 4 of 6 goes wrong
🧭 Copilot User has a canvas they own (code, doc, problem) and wants help in context Cold-start problems with no existing artifact Confidence level + "why I suggested this" Hiding its reasoning, so users can't calibrate

The trust signal that defines each pattern

If there is one design decision per pattern that makes or breaks it, it's the trust signal. Four cards, one trust signal each. Miss one, and the surface collapses into distrust.

🔗
Chat → Citations
Every non-trivial claim gets a source, clickable. No citation, no trust. Users forgive a wrong answer with a source faster than a right answer without one.
↩️
Agentic → Undo
Every completed step must be reversible. If step 3 just emailed a parent, there's a "Revert last action" button — and it really works. Without undo, agents terrify users.
🎯
Copilot → Confidence
Show how sure the AI is: "Confident · 87%" or "Guess, check my work." Copilots live inside the user's workflow; they must not pretend to certainty they don't have.
🤝
Inline → Handoff
If the user rejects three suggestions in a row, stop suggesting for this session. Inline surfaces that don't know when to shut up become nags. A graceful handoff to the user's own typing is the trust signal.

Scenario — VidyaPath, 10 weeks, one doubts-helper

📚
Real-world scenario
VidyaPath doubts solver — Amara picks the surface

Amara is a PM at VidyaPath in Bangalore. She has 10 weeks to ship an AI doubts-helper for high-school math — Class 9 and 10, CBSE board. The brief from leadership is one line: "help students when they're stuck, without just giving the answer." Target: 60% of students who use it pass their next topic test.

She has to choose one surface to ship by Week 10. Is it chat? Inline? Agentic side-panel? Copilot inside the problem-solving flow? The answer is not obvious, and the uncomfortable thing Amara discovers when she watches a dozen students use the prototype is that the surface she picks changes the pedagogy, not just the UX. A chat bubble begs to be given the answer. An inline auto-complete finishes the student's work for them. An agentic panel does too much, in the background, silently. A copilot — a hint-helper inside the problem canvas — is the only surface that gently nudges without taking over.

Amara picks copilot. Not because chat is wrong in the abstract, but because for this task (supervised learning, teacher-approved, anti-cheat) the shape must invite struggle, not replace it.

The 2×2 that made the call

Amara drew this on a whiteboard. Two axes: latency tolerance (how long the user will wait) and scope (one step vs many steps). The four patterns fall cleanly into four quadrants.

↑ Broad scope (many steps) Narrow scope (one step) ↓ Low latency High latency tolerance 🧭 COPILOT in-context, many steps, fast hints on demand 🤖 AGENTIC background work OK, show progress + undo ✨ INLINE one field, instant, user accepts or ignores 💬 CHAT one question, full reply, users wait if it's thinking
Why copilot won for Amara: a student solving a quadratic is in a many-step flow, but each individual hint needs to be fast — they're mid-thought. Top-left quadrant. Chat felt too slow; inline was too narrow; agentic did too much behind their back.

How it actually works — Amara's journey strip

Every AI UX decision eventually collapses to one sentence per column: what does the user want, which surface meets them, what does the AI do, what is the outcome? For the VidyaPath copilot, the strip looks like this.

User intent "Stuck on step 2" AI surface Copilot hint panel AI action Hint, not answer Outcome Student solves

Notice what the surface makes implicit: the copilot never returns a full worked answer — the AI action is "hint, not answer." That's not a prompt trick; it's a UX commitment. The surface shape enforces the pedagogy. Pick chat instead and the default rhythm is "user asks → AI replies with full solution." Pick inline and the default is "user writes → AI completes." Only copilot has "user struggles → AI nudges → user struggles more → AI nudges deeper" as the natural rhythm.

The PM insight:
surface choice is pedagogy choice. Marketing-copy tools default to inline because the user wants the sentence written. Tax assistants default to chat because the user wants to talk to a person. Ops automations default to agentic because the user wants the work done. Learning products should default to copilot, because the user needs to grow, not just finish.

The AI connection

How real AI teams reason about this

The four patterns aren't just labels — they map onto four very different engineering stacks. Chat is built on streaming responses and conversation memory. Inline is built on low-latency small-model inference and ghost-text UI. Agentic is built on tool-calling, planners and rollback logs. Copilot is built on context-window packing (pulling the user's current artifact into the prompt) and confidence scoring. When you pick the surface, you pick the stack — and the stack decides your cost, latency and failure modes. streaming tool-calling context-packing rollback-log ghost-text

Mini challenge — pick the surface for three features

20 minutes, no coding. Three product briefs, one call each.

1. BoraPay transaction dispute helper

A Kenyan user taps "I don't recognize this charge." The AI needs to explain the transaction and offer next steps. Which surface? Write one line naming the primary trust signal. (Hint: low latency, narrow scope, user wants to talk.)

2. NaijaCart seller listing improver

A Lagos seller is writing a product title. The AI can suggest a better title with SEO keywords. Which surface? What's the failure mode you have to design against? (Hint: one field, one suggestion at a time.)

3. TaxSetu year-end filer

The user uploads 12 months of statements and says "file my taxes." The AI needs to classify entries, fill forms, and surface edge cases for review. Which surface? What must be reversible? (Hint: the work takes 4 minutes; the user won't watch a spinner.)

Interview questions

Easy Name the four canonical AI UX patterns.
+
Chat, Inline, Agentic, Copilot. Chat is conversational. Inline is an in-field suggestion. Agentic is a side-panel that plans and executes multi-step tasks. Copilot is an embedded helper inside the user's own artifact (doc, canvas, problem). The mnemonic I use: C-I-A-C — "chat is a conversation, inline is a nudge, agentic is a worker, copilot is a companion."
Easy What's the difference between chat and agentic?
+
Chat is one-turn-at-a-time, user-driven, and replies with text. Agentic plans multiple steps, often takes real-world actions (sending email, updating a record, calling APIs), and runs in the background. Chat is "I answered your question." Agentic is "I did your task." The trust signals are different too: chat needs citations, agentic needs a visible task list and undo on every step.
Medium When does an inline suggestion suggest vs auto-apply?
+
Default is always suggest — ghost text with an explicit Tab/Accept. Auto-apply is only safe when three conditions hold: (1) the action is trivially reversible (undo stack works), (2) the user has opted in explicitly, and (3) the confidence is very high from offline evals (e.g. >95% on the golden set for this narrow case). Miss any of those and auto-apply becomes the surface users hate most — "it typed for me, I didn't ask." Grammar correction is a good auto-apply case. Full sentence rewrite almost never is.
Medium Your team wants to default to chat for every new AI feature. What's your pushback?
+
Three arguments. One: chat taxes every task with a typing cost — users who do the same thing 20 times a day will resent it. Two: chat forces the user to do surface translation — they have to re-describe context the system could have known (which file, which customer, which step). Copilot and inline remove that cost. Three: chat has the loosest trust signal — without citations, it looks like confident magic, which is exactly the hallucination trap. The right question is "what is the user's task shape?" — and only a minority of tasks are genuinely conversational.
Hard Design the handoff from copilot → agentic when the task scope grows mid-flow.
+
This is the interesting one. A student opens a copilot on one algebra problem. Mid-session they say "actually, build me a revision sheet covering all the topics I've struggled with this month." The task just jumped from single-artifact (copilot) to multi-step (agentic). Here's the clean handoff:

(1) Detect scope growth. Either explicitly (user clicks a "Do this for me" CTA) or implicitly (the model's planner predicts >3 steps). Do not silently escalate.

(2) Explicit consent screen. "This will take about 4 minutes and will read your last 30 problems and draft a PDF. Proceed / Cancel." No autonomy without consent.

(3) Switch surfaces visibly. The copilot hint panel collapses; an agentic task list slides in with each planned step labeled. The user now sees progress, not a spinner.

(4) Undo and stop always available. Per-step rollback, plus a big "Stop" button. Agentic surfaces that can't be stopped are the number-one source of user distrust.

(5) Fall back gracefully. If any step's confidence drops below threshold, pause and flip back to copilot mode: "I started step 3 but I'm not sure — want to do this one yourself?" The surfaces should breathe in and out as the task demands, with the user always in the driver's seat.

Test your knowledge

Question 1 of 5
A user types in a form field and sees ghost text suggesting the completion. Which pattern is this?
Question 2 of 5
Which pattern most needs an undo on every completed step as its primary trust signal?
Question 3 of 5
Amara at VidyaPath picks copilot over chat for the doubts-helper. Her strongest single reason is:
Question 4 of 5
On the latency × scope 2×2, which pattern sits in high-latency-tolerance + broad-scope?
Question 5 of 5
When is it safe for an inline suggestion to auto-apply without an explicit Accept?
0
/5
You can now pick the right AI surface on purpose.

Common mistakes

Chat as default for everything
"We built a chatbot" is the most over-picked answer in AI product. Chat taxes every interaction with typing, adds translation cost, and has the loosest trust signal. Default to the surface the task actually needs, not the one that's easiest to prototype.
Agent with no undo
An agentic panel that finishes 4 steps and has no "Revert last action" is terrifying. Users will stop starting it. Undo on every step is not a polish feature — it's the price of entry.
Inline suggestion that auto-applies
Writing into the user's field without consent is the fastest way to turn a helpful feature into "why is it typing for me?" Default to ghost-preview + explicit Tab to accept. Auto-apply is an opt-in privilege, not a default.
Copilot that hides its reasoning
A copilot that just says "Try this" without a "why" denies the user the one thing that makes copilots useful — the ability to calibrate trust. Always show at least a one-line justification: "Because your last 3 attempts missed the discriminant."

Key takeaways

  • Four canonical AI UX surfaces: Chat, Inline, Agentic, Copilot. Learn all four before you pick one.
  • Pick the surface by mapping the task onto the latency × scope 2×2 — not by what's easiest to prototype.
  • Each pattern has a defining trust signal: chat → citations, inline → explicit accept, agentic → undo, copilot → confidence.
  • Surface choice is pedagogy choice for learning products, and authority choice for everything else — it defines who owns the next action.
  • Design the handoff between surfaces when task scope grows or shrinks. Explicit consent, visible switch, undo always available.
  • "We'll just build a chatbot" is almost never the right answer. Make the team defend chat the way they'd defend any other PRD decision.
Africa & India's Best AI Academy
🤖 AI-first 🎓 Placement Assistance

You can now pick the right AI surface.
Now design the trust signals around it.

Join the Best AI Program: The AI-Era Agentic Engineer — 5 months, fully online, with placement assistance. Week 1 is free.

Explore The AI-Era Agentic Engineer →