Contextual Assistance
What is Contextual Assistance?
Contextual Assistance is an AI design pattern where systems proactively offer help based on user context and behavior, without waiting to be asked. Instead of interrupting workflows with generic tips, this pattern analyzes what users are doing right now and suggests relevant actions at the perfect moment. It's most effective for repetitive tasks, complex applications, and situations where AI can learn from patterns to predict needs. Examples include Gmail's Smart Compose finishing your sentences, search autocomplete guessing your query, and Notion suggesting relevant pages as you type.
Problem
Users need guidance but often don't know what or when to ask. Traditional help interrupts workflows.
Solution
Design intelligent assistance that proactively offers relevant help, suggestions, or information based on user context and behavior. Anticipate needs rather than waiting for explicit requests.
Real-World Contextual Assistance Examples
Implementation
When to use Contextual Assistance, and when it backfires
Use it when
- The user is mid-task and a suggestion saves a step they were already about to take: finishing a sentence, completing a path, surfacing the file they're clearly reaching for.
- The signal is strong and specific. You can name exactly what the user is doing and what they need next, not just that they're 'active'.
- The help is glanceable and ignorable: ghost text, an inline chip, a quiet panel the user opts into with one key, never a thing that grabs focus.
Don't, or minimize, when
- The signal is weak. 'The user opened a document' is not a reason to offer help; it's a reason to get out of the way.
- The suggestion interrupts: a popup that steals focus, a modal mid-keystroke, a tooltip that covers the thing the user is looking at.
- You're firing to look helpful, not because you're confident. Coverage metrics ('we suggested something on 90% of sessions') are how this pattern rots.
The trap
The Clippy reflex: help that fires on near-zero signal to prove it's working. 'It looks like you're writing a letter' interrupted everyone who opened a document, confident and useless. The math is asymmetric and most teams get it backwards: a suggestion you withhold costs nothing because the user just keeps working, but a wrong or ill-timed one costs attention even when it's dismissed, because the user still had to stop, read it, and decide it was noise. Optimizing for suggestion volume optimizes for the one cost that compounds. The discipline of the pattern is staying silent until you're sure, and a system that can't stay silent has betrayed it.
Take it into your own product
- 1
Silence is the default, not the failure state.
A suggestion you withhold costs nothing: the user just keeps working. A wrong one costs attention even when it's dismissed, because they had to stop, read it, and decide it was noise. When in doubt, say nothing. The best contextual assistance is invisible most of the time.
- 2
Fire on a strong signal, not on presence.
'The user opened a document' is not a reason to help; it's a reason to get out of the way. You earn the right to suggest by naming exactly what the user is doing and what they need next. If the trigger is just 'they're active', you don't have a signal, you have an excuse.
- 3
Glanceable and ignorable, or it isn't assistance.
Ghost text, an inline chip, a quiet panel: help that sits beside the work and waits. The moment a suggestion steals focus or covers what the user is reading, it stops being help and becomes an interruption you dressed up as one.
- 4
Measure what you withhold, not just what you show.
Accept-rate per surface is the honest metric. Coverage ('we suggested something 90% of the time') is how this pattern rots: it rewards volume, which is exactly the cost that compounds. A surface with many shows and few accepts is noise to cut, not a win to report.
- 5
Let the user turn it down.
Someone who dismisses the same suggestion three times is telling you to stop. Give every stream a volume control and persist it. A learning system that can't be told 'enough' isn't assisting, it's nagging with a model attached.
Add Contextual Assistance to your product
Copy the prompt below into Claude Code or Cursor in your repo. It encodes the four moves on the left and asks Claude to find your AI decision surfaces and update them. Claude reports what it changed and asks before adding dependencies.
Check if your product already has this pattern
Upload a screenshot. We'll tell you which of the 36 patterns your AI interface uses and where the gaps are.
Audit My DesignMore in Human-AI Collaboration
Human-in-the-Loop
Balance automation with human oversight for critical decisions, ensuring AI augments human judgment.
Augmented Creation
Empower users to create content with AI as a collaborative partner.
Collaborative AI
Enable effective collaboration between multiple users and AI within shared workflows.
