The Disappearing Dashboard
A field guide for AI teams on why your logo, dashboard, and chat UI all need a redesign.
Before the future,
look at the present.
AI products are starting to look the same — and function the same. Almost everything gets built around a chat window. Worth pausing on that before scrolling further.
Sixteen companies.
One shape.
Gradient blobs, circular marks, the same soft curves — scroll through and see how little separates them.
Consumer hardware & software
Loop
Data science & MLOps platform
Radial
Creator of ChatGPT
Loop
Creator of Claude
Radial
AI assistant across Microsoft 365
Container
AI-powered answer engine
Modular
Google's multimodal AI model
Sparkle
AI-powered creative tools
Orbit
AI search and assistant
Container
Computer vision AI platform
Modular
Meta's AI assistant
Orbit
AI platform for legal teams
Radial
AI for healthcare teams
Container
xAI's conversational assistant
Loop
AI video generation platform
Modular
AI personal finance advisor
ContainerEvery chat window performs
the same thinking.
Sparkle icon, streaming text, a spinner that says "thinking" — the interaction pattern that's become shorthand for "this is AI."
Chat as the primary interface
A persistent input box, streaming responses, and turn-based history — borrowed from messaging apps but repurposed as a general-purpose command line.
History sidebar
A persistent list of past conversations, so work can be resumed, searched, or grouped into projects instead of starting from scratch each time.
Streaming / typewriter text
Tokens appear incrementally rather than all at once, signaling "thinking" and keeping latency feeling tolerable.
Generating your answer word by word...
Slash commands and @-mentions
Quick ways to invoke specific modes, tools, files, or people without leaving the text field (e.g. /image, @document).
Prompt suggestions / starter chips
Example queries shown on empty states to overcome blank-page paralysis.
Regenerate / branch responses
Letting users get an alternate answer or explore multiple response paths from the same prompt.
Visible reasoning / "thinking" traces
Collapsible sections showing chain-of-thought or intermediate steps, giving users a peek into the "why" without cluttering the main answer.
Tool-use transparency
Showing what the AI is doing — searching the web, running code, calling an API — as it happens, rather than just returning a final answer.
Citations and source grounding
Inline citations or footnotes tying claims back to retrieved sources, addressing hallucination concerns.
Revenue grew 34%1 driven by new markets2.
Confidence signals
Hedging language, or in some products, explicit confidence scores or "verify this" prompts.
Edit / interrupt controls
Stop-generation buttons, the ability to edit a previous prompt and regenerate downstream, or steer output mid-stream.
Thumbs up/down and flagging
Lightweight feedback baked into every response, used both for user satisfaction and model improvement.
Inline correction affordances
Highlight text and ask for a rewrite, rather than re-explaining the whole request from scratch.
The meeting is schedule for tomorrow.
Artifacts / canvas panels
A persistent side panel where generated content (code, documents, diagrams) lives separately from the conversational thread, so it can be iterated on without re-reading chat history.
Diff views
Showing before/after when the AI edits existing content (especially in code tools), so changes are auditable rather than opaque overwrites.
Progressive disclosure of complexity
Simple chat by default, with "advanced" panels (parameters, system prompts, model settings) tucked away for power users.
Task/plan checklists
The AI lays out a multi-step plan before executing, and checks off steps as it goes, so long-running tasks don't feel like a black box.
Human-in-the-loop checkpoints
Explicit confirmation steps before consequential actions (sending an email, spending money, deleting files).
Send email to team@company.com?
Background / async task cards
A task kicks off, the user can leave, and comes back to a completed result rather than watching a spinner.
Drag-and-drop / paste-anything input
Images, PDFs, and screenshots dropped directly into the chat input, blurring the line between "upload" and "converse."
Voice-first with visual fallback
Voice interaction paired with a transcript or visual summary, so users aren't stuck relying on audio memory.
AI isn’t just changing products. It’s changing what a product is.
The first signs are already here: screens that build themselves when you ask, agents you give a goal instead of a task, products designed for machines as well as people, and memory that matters more than the screen.
Two axes. Four futures.
AI products can go in four directions. The screen can stay fixed or get built for each request, and the product can feel like something specific or like every other AI product.
A real point of view, on a surface that never moves. Most good design still lives here — and this is the quadrant with an expiry date.
Surfaces that assemble around an intent and still carry a hand. Almost nobody is here yet, which is exactly why it is worth aiming at.
A sidebar, a sparkle, a chat window. Shipped in a fortnight, indistinguishable in a month, replaced in a year.
The interface builds itself and looks like nobody made it. Genuinely powerful, completely anonymous, trivially swappable.
Beyond the sparkle.
Newer AI brands are trading sparkles and orbs for mascots, pixels, volumes and utility glyphs.
Open-source AI lab
Mascot
Open machine learning platform
Mascot
European open-weight models
Pixel
Run AI models via API
Pixel
Spatial intelligence lab
Volume
Deep-tech research lab
Volume
AI infrastructure platform
Linked
AI for contact centers
Linked
Security-first AI
Glyph
Secure access for AI agents
Glyph
Generative image research lab
Glyph
AI audio generation
Wave
AI commerce platform
Wave
Work search and assistant
Letterform
Agent coordination
LetterformFourteen more, too new to call.
These carry no redesign-opportunity tag, because nobody has shipped a settled-enough version to call them generic or distinctive. Worth tracking rather than critiquing prematurely.
Transient, context-triggered UI
Interface elements that appear only when contextually relevant and disappear after, rather than living in persistent chrome.
Central to this report's own thesis, but too few working examples exist to say what “generic” even looks like here yet.
Per-response generative UI
A one-off interactive component generated for a single answer, not a reusable persistent surface — distinct from Artifacts-style panels.
Still mostly experimental and research-stage rather than shipped in mainstream products.
Multi-agent orchestration visualisation
Showing several agents working in parallel, or handing tasks off to one another.
No settled visual convention exists yet — every product doing this is inventing its own approach.
Proactive / ambient surfacing
Unprompted suggestions appearing based on inferred context, distinct from voice-first interaction.
Early and inconsistent; raises unresolved questions about interruption and trust.
Cost / compute transparency
Real-time visibility into what an autonomous action is costing, as agents take more expensive and more consequential actions.
Barely exists yet — most products still hide this entirely.
Cross-surface handoff continuity
A task starting in one app and continuing in another — browser to email to calendar — with a coherent UI thread running through it.
Requires interoperability standards that are still being worked out industry-wide.
Trust calibration over time
UI showing how reliable an agent has historically been at a specific kind of task, rather than a generic one-time confidence score.
Conceptually discussed, essentially unshipped in any mainstream product so far.
Emotion-aware interfaces
Interfaces that read tone, frustration or confusion and quietly adjust: offering a human handoff, slowing the pace, softening the visuals.
Only credible as opt-in and reversible, and no one has settled how to ask consent without breaking the moment, or where sensing tips into surveillance.
Risk-tiered conversation flows
One chat that switches register by stakes: freeform for exploring, rigid buttons and explicit confirmation for transfers, account changes and other hard-to-undo steps.
Teams are still guessing where the line sits and how to switch modes without the seam showing.
Capability disclosure up front
The assistant states what it can and can't do the moment it opens, through tags and starter actions, instead of a friendly persona that overpromises.
Mostly ad-hoc onboarding copy today, with no shared pattern and no answer for keeping it accurate as capabilities change.
Inspectable, editable memory
Memory made visible in layers, what's held within a chat versus across chats, with somewhere to see, correct and delete what the AI has kept.
Controls are appearing, but there's no shared model for how much to surface without making every conversation feel audited.
Ambient status lines
A small pulsing line inside the app’s own chrome narrates background work (“reading threads… drafting reply”) with no pop-up or panel.
Passive assurance is easy to overlook and easy to overdo, and there's no settled convention for how much narration reassures versus distracts.
Actionable uncertainty
Uncertainty that says why the AI is unsure and what to do about it (verify, compare, pick an alternative) rather than a colour or a percentage.
Coarse buckets beat false precision in principle, but few products tie the signal to a next step; most still show decoration.
Phase-aware waiting
Loading states that say whether the model is thinking, retrieving or generating, using skeletons instead of spinners and keeping results in the order they were requested.
When a later task finishes before an earlier one, trust collapses, and no convention exists yet for showing order and state honestly.
By 2027, the screen stops being the product.
Screens stop being fixed layouts and start being built for each request. The products that design the parts and rules behind what gets generated keep an identity. The rest generate the same thing as everyone else.
The end state isn’t every app with a chat window bolted on. It’s the app grid itself dissolving into one persistent, personal agent drawing on a marketplace of capabilities — and products becoming ingredients rather than destinations.
Hi, I am Agnieszka.
Staff Product Designer, 13+ years in B2B and regulated domains (HR/payroll at Personio, quantum cloud at IonQ, enterprise SaaS at BCG X). I also mentor designers on the side and write and make videos about the tech.
Connect with me on LinkedInIs your AI product part of
the sameness?
Subscribe to updates and get the free PDF as a welcome gift. The free test: 10 questions, a fix for every “yes,” and 5 predictions for 2027 with what to do about each.