Suggestions

Exemplar page — first pass, entirely proposal. No Invoca product ships this pattern today. Nothing below is code fact or measured behavior — it is a proposal offered for review. See Coverage, stated honestly.

What it is

Signal AI has just summarized a call: a 40-second hold, the caller mentioned pricing twice, the call converted. Underneath that summary, a row of chips offers what a manager might plausibly want next — Show similar calls, Flag for QA review, Draft a follow-up email — without requiring the manager to think of those questions or type them out. A Suggestion is that chip: a proposed next action, attached to a specific piece of AI output, that the user can take with one tap or ignore entirely. The same shape recurs for a marketer running a natural-language search over call data: after an answer returns, suggested follow-up questions appear beneath it — narrowing the date range, comparing against a different campaign — phrased as things the marketer might ask next rather than as a menu of unrelated features. Suggestions differ from an Initial CTA in exactly one way that matters: they require an existing context to suggest from. A CTA invites someone who has nothing yet; a Suggestion proposes what to do with something that already exists. A page cannot show Suggestions before it has shown anything.

Choose this when / choose something else when

Situation	Do this instead	Why
The user hasn’t engaged with the feature yet — there’s no output to attach a suggestion to	Initial CTA	Suggestions need existing context; a CTA is what creates that context in the first place
The chip is really adjusting a parameter of the current or next run (tone, scope, depth)	Controls	A Suggestion proposes what to do next; a Control changes how the doing happens
The option set is a fixed list the user is choosing to hold as a durable value, not a one-off action	Select	A Suggestion is a transient proposal attached to one output; a Select’s value persists as form state
The list is a general, always-available set of actions unrelated to any specific AI output	Menu	A Suggestion is contextual to what’s on screen; a Menu is a standing list of things to do, and Menu already exists for that job

Agency tier

Suggests. Tapping a chip proposes an action; by default it does not commit anything on its own — it starts a request the user still reviews (drafting a follow-up email opens a draft, it does not send one). A chip that skips review and acts directly has moved to Drafts or Acts, and that escalation must be stated on the chip itself, not left for the user to discover. Per TITAN-AI overview, escalating a feature’s tier is a decision, never a default — a Suggestion is the easiest place for that line to blur, because tapping a chip already feels like a small, safe action regardless of what it actually triggers.

Anatomy

┌─────────────────────────────────────────────┐
│  This call converted after a 40-second       │
│  hold. The caller mentioned pricing twice.   │
│                                               │
│  ( Show similar calls )  ( Flag for QA )     │
│  ( Draft a follow-up )                       │
└─────────────────────────────────────────────┘

#	Part	Component	Required
1	Source output the suggestions attach to	Card	Yes
2	Suggestion chip	Tag, clickable variant	Yes — one or more
3	Overflow, when more suggestions exist than fit	Menu	No

Outcome states

State	Treatment
Working	Tapped chip shows a loading affordance in place; the row of remaining chips stays interactive
Streaming / partial	Chips may appear progressively as they’re generated rather than all at once — each chip is fully formed the moment it renders, never a partial label
Confident and right	The proposed action matches what the user would have picked manually
Confident and wrong	Most detail below.
Uncertain	Not applicable as a per-chip signal. Suggestions are not individually ranked with a shown confidence — see Disclosure & recourse, question 3.
Refused	Not applicable to the chip itself. If tapping a chip triggers an action that gets refused, that refusal belongs to the invoked Action’s outcome states, not this page’s.
Empty	No suggestions were generated for this context. The row does not render at all — no placeholder chips, no “no suggestions available” message where the row would have been.
Interrupted	User navigates away while suggestions are still being generated; the partial set already shown stays usable, the rest simply never arrives
Degraded	Falls back to a generic, non-contextual suggestion set (e.g., always offering “Export” and “Share” rather than context-specific proposals) — disclosed by the suggestions visibly not referencing the specific record
Rate-limited / over budget	Suggestion generation itself is skipped rather than queued or retried silently; the row simply does not appear that time
Stale	Suggestions were generated against an earlier version of the underlying data (e.g., the call record was re-tagged since); tapping one may act on assumptions no longer true

Confident and wrong. A suggestion chip reads as a proposal, which already carries less weight than a stated answer — but a chip that is contextually wrong looks identical to one that is right, because the chip’s label is plausible on its own (“Draft a follow-up email”) even when it doesn’t fit this record (the call has no email on file). The user notices only after tapping: the drafted action opens against no recipient, or the “similar calls” list returns results with nothing in common with the source call. A wrong suggestion currently has no seam before the tap — this is a named gap, not a solved case; see Gaps.

Disclosure & recourse

Does the user know this is AI? Yes — the chip row is visually distinct from the record’s own manual actions (e.g., a toolbar), and sits attached to AI-generated output the user already knows is AI-generated.
What did it use? The same source the attached output used, plus, for suggestions like “show similar calls,” a comparison across other records — state which, in the chip’s own label where practical (“similar calls” implies a comparison; “draft a follow-up” does not need to).
How sure is it, and does that change behavior? No per-chip confidence is shown. A confidence number on a suggestion would decorate, not inform, unless it changed whether the chip renders at all — which it currently does not. This is intentional, not an omission; see TITAN-AI-03.
How does the user check it? Undecided. Nothing today shows why a specific suggestion was proposed. See Gaps.
How does the user correct it? Dismissing a single chip removes it from view for that session. Whether a dismissal is remembered, or whether the same wrong suggestion reappears next time, is undecided.
How does the user get out? Ignoring the entire row costs nothing — no chip is required to proceed, and no suggestion blocks the surrounding content.

Reference

No model, prompt, tool schema, latency budget, or cost has been defined for how suggestions are generated.

Evaluation

Not evaluated. No eval set exists for suggestion relevance or for the “confident and wrong” case named above.

Content

Element	✅	❌
Chip label	Show similar calls	Similar
Chip label	Draft a follow-up	You could draft a follow-up email
Chip label	Flag for QA review	Flag
Overflow trigger	More suggestions	···

Chip labels name the action, matching Menu’s own content rule: a list of things to do, phrased as verbs, not as a sentence describing the option.

Accessibility

Suggestions arriving progressively are announced through a single aria-live="polite" region summarizing that new suggestions are available, not one announcement per chip — a screen reader user does not need every chip’s arrival narrated individually.
Each chip is independently focusable and activatable by keyboard, in document order, matching a standard Tag’s clickable behavior.
No confidence is expressed by color alone, because none is expressed at all — see Disclosure & recourse, question 3. If that changes, any confidence signal must pair with text per TITAN-COLOR-03.
Tapping a chip that opens a new surface (a draft, a filtered list) moves focus to that surface’s heading; tapping one that acts in place keeps focus on the chip through its loading state.

Constraints

ID	Constraint	Rationale
TITAN-SUGGEST-01	A Suggestion never renders without an existing output to attach to.	Suggestions require context to suggest from; without it, the pattern is actually an Initial CTA.
TITAN-SUGGEST-02	A chip that acts immediately, without a review step, states that on the chip itself — never a bare verb indistinguishable from a chip that opens a draft.	Escalating past Suggests-tier is a decision that must be visible, not discovered after the tap — see TITAN-AI overview, agency tier.
TITAN-SUGGEST-03	An empty suggestion set renders no row at all — never an empty container or a “no suggestions” message in its place.	A visible empty state for an optional, ambient feature adds visual noise with nothing for the user to act on.
TITAN-SUGGEST-04	A degraded, non-contextual suggestion set is only ever generic — it never fabricates a specific-sounding reference to the record it failed to actually analyze.	A generic suggestion that reads as specific is a confident-and-wrong case created by the fallback itself, not by the model.
TITAN-SUGGEST-05	Suggestion chip labels are verbs naming the action, never a sentence describing the option.	Matches Menu’s content rule — a list of things to do reads faster than a list of descriptions.
TITAN-SUGGEST-06	Dismissing the suggestion row never blocks or delays the content it’s attached to.	Suggestions are strictly additive; treating them as a gate the user must clear contradicts the Suggests tier itself.

Divergences

Not applicable — nothing is shipped yet to diverge from.

Gaps

How a user checks why a specific suggestion was proposed is undecided. No mechanism (a tooltip, an expandable reason) exists even as a proposal yet — see question 4 above.
Whether a dismissed suggestion is remembered across sessions, or reappears every time the same context recurs, is undecided.
How many chips may show at once before overflowing into a Menu, and how that threshold is chosen, is undecided.
Whether suggestions should ever be ranked or reordered based on a per-user history of which ones get tapped is undecided, and if so, whether that ranking should be disclosed.

Volatility

This page assumes a model can generate several distinct, independently useful next-step proposals from one piece of output, cheaply enough to show a handful at once. If that generation step turns out to be too slow or expensive to run automatically for every AI output, Suggestions may need to become a pattern the user requests rather than one that appears unprompted — a different agency and disclosure shape. Dated 2026-09-02; revisit on the first real implementation or on any change to per-invocation cost for the underlying model.

AI Experience overview — vocabulary, agency tiers, and the six disclosure questions this page answers
Initial CTA — the entry point Suggestions assumes has already happened
Open input — where a user types a request a Suggestion chip might otherwise shortcut
Controls — for adjusting behavior rather than proposing a next action
Trust builders: Caveat — how the interface admits uncertainty, relevant to the confident-and-wrong gap named above
Tag, Menu, Card — the components this pattern composes

Why it works this way

A Suggestion’s entire value is in costing less than typing the same request out. The moment a chip requires the user to verify its relevance before tapping it, it has cost more than it saved — which is why the confident-and-wrong case above is not a minor edge case but the failure mode that determines whether this pattern is worth building at all. A Suggestion that is wrong often enough to require checking is a Suggestion nobody will trust the next time, including the times it would have been right.

Overview

Wayfinders

Inputs

Tuners

Governors

Trust builders

Identifiers

Actions

What it is

Choose this when / choose something else when

Agency tier

Anatomy

Outcome states

Disclosure & recourse

Reference

Evaluation

Content

Accessibility

Constraints

Divergences

Gaps

Volatility

Why it works this way

​What it is

​Choose this when / choose something else when

​Agency tier

​Anatomy

​Outcome states

​Disclosure & recourse

​Reference

​Evaluation

​Content

​Accessibility

​Constraints

​Divergences

​Gaps

​Volatility

​Related

​Why it works this way

What it is

Choose this when / choose something else when

Agency tier

Anatomy

Outcome states

Disclosure & recourse

Reference

Evaluation

Content

Accessibility

Constraints

Divergences

Gaps

Volatility

Related

Why it works this way