Anticipatory design in 2026: what AI predicts, and what humans still want chosen

Anticipatory design — the idea that interfaces should predict what users want before they ask — moved from concept to ubiquitous reality between 2023 and 2026. Most modern AI-augmented products now do some version of it: Notion AI predicting your next sentence, GitHub Copilot writing your function, Gmail’s Smart Reply choosing your response, Spotify generating your daily mix.

The hard question isn’t whether to anticipate. It’s WHEN to anticipate and when to give users the dignity of choice. Get this wrong and you build products that feel either annoying or paternalistic.

The case for anticipation

The version that works: anticipating low-stakes, high-friction decisions where the user’s preference is either obvious or doesn’t matter much.

Examples that work:

• Autocomplete in URL bars (you’ve typed it before, the prediction is almost always right)
• Calendar suggestion times based on attendee availability
• Smart Reply for short acknowledgments (yes, no, thanks, will do)
• Spotify’s auto-generated playlists (you can always skip a track)
• GitHub Copilot for boilerplate code (you can always reject the suggestion)

Common thread: the cost of a wrong prediction is low. Worst case is one extra second of work to override.

The case against anticipation

The version that doesn’t work: anticipating high-stakes or identity-signaling decisions.

Examples that fail:

• Email composer auto-completing complex messages on your behalf (the wrong tone with a client damages the relationship)
• Auto-generated social media posts (the homogenization of voice is the cost; readers can detect it)
• Auto-generated essay drafts in personal writing (the writing IS the thinking; outsourcing it ruins the value)
• Automatic financial allocations (Apple’s Savings tool: low-stakes; AI rebalancing your retirement: high-stakes)
• Auto-generated dating profile bios (the cost of a fake self is dating the wrong people)

Common thread: the prediction is doing thinking the human SHOULD be doing because the thinking matters more than the output.

The right test for “should this be anticipated by AI?” is: what’s the cost of being wrong? If the user’s recovery is one extra click, anticipate. If the user’s recovery is a damaged relationship, a misrepresented identity, or a wrong long-term decision, ask.

Three patterns that work in 2026

The anticipatory design patterns we see compounding across products:

1. Suggestions with one-tap acceptance. The product predicts; the user accepts or ignores. No automatic action. Notion’s text suggestions follow this — they appear as ghosted text and only activate on Tab.

2. Defaults that match recent behavior. The product picks the most likely option but exposes the alternative. Apple Calendar suggesting meeting times in your usual range, but with the full picker available.

3. Generated drafts with explicit editing prompts. The product generates a starting point and explicitly invites the user to edit. ChatGPT and Claude do this implicitly — the output is a draft, not a final answer.

What these have in common: the user retains agency. The product anticipates to accelerate, not to replace decision-making.

Three anti-patterns we see repeatedly

1. Hidden auto-actions. A product takes an action on the user’s behalf without surfacing the decision. The user finds out later when something is wrong. This violates trust and is increasingly common in AI-augmented products that prioritize speed over transparency.

2. Aggressive defaults that nudge. A “select all” that’s pre-checked. A “yes, send marketing emails” that’s the default. Anticipatory design becomes dark pattern when the prediction is biased toward the product’s preferences instead of the user’s.

3. Over-personalization that flattens choice. Recommendation algorithms that only show you what you’ve already seen. The filter bubble in product form. Users become passive consumers of an algorithm’s prediction rather than active discoverers.

How we apply this to client products

When we audit a product for our Brand + UX engagements, we explicitly look for:

• Where does the product make decisions FOR the user?
• Where does the product offer suggestions WITH the user?
• What’s the recovery cost when the prediction is wrong?

Most products err in one of two directions: either over-personalizing (filter bubble) or under-personalizing (every user gets the same experience). The right calibration is product-specific but the test above is universal.

We then map specific UX surfaces to one of:

• Predict + auto-apply (low-stakes, low-reversal cost)
• Predict + suggest (medium-stakes, easy to override)
• Don’t predict, let the user choose (high-stakes, identity-signaling)

The case for friction

Counterintuitive but important: some design friction is good.

Forms that require typing instead of auto-filling — when the cost of being wrong is high. Confirmation steps before money moves. Mandatory delays before irreversible deletions.

Anticipatory design without friction is paternalism. Friction is how products show users that THEIR decision matters.

We argue with clients about this regularly. The PM instinct is to remove every click. The right answer is to remove every click that doesn’t matter and keep the clicks that signal “this is your call.”

What’s next

The next wave of anticipatory design — already starting in 2026 — is multi-step agent behavior. An AI assistant that not only suggests your next reply but proactively schedules the follow-up meeting. The same calibration question applies, scaled up: when does the agent act autonomously, when does it pause to ask, and how does it surface the actions it took?

Anthropic’s Claude Code and OpenAI’s agentic features point at where this is going. The teams building these are explicitly thinking about agency calibration in a way most product teams aren’t.

What we do for clients

For product clients, we run UX engagements that explicitly include anticipatory design audits — what your product predicts, where it should and shouldn’t, and how it surfaces predictions to users.

Tell us what you’re working on.

Related reading:

• eCommerce UX in 2026 — adjacent: predicting in commerce
• Brand-driven web design — adjacent: design as signal