What this section governs
Anything in Invoca’s product where the output is generated by a model rather than rendered from state, or where a feature acts on the user’s behalf rather than only responding to input. A Signal AI call summary, a natural-language search over call data, an assistant that drafts a follow-up or updates a record, a panel that recommends a routing or budget change — all of it lives here, not in Patterns. It does not govern the visual materials an AI surface is built from. A summary still renders in a Card; a streaming answer still uses Foundations motion and color tokens. This section governs what is different about the experience once the thing filling those components is not deterministic.Why this isn’t a Pattern
The rest of this design language resolves top-down through four tiers — Foundations, Components, Patterns, Views — and each tier is defined by a test: is it a token, a package export, a composition with no export, or a page archetype. An AI experience fails all four tests the way a Pattern would pass them, because four assumptions those tiers depend on do not hold here:
So an AI experience page carries every section a Pattern page carries — constraints, content
rules, accessibility, gaps — plus four sections nothing else in this design language has:
agency tier, outcome states, disclosure & recourse, and evaluation. Skipping
straight to Components or Patterns for an AI feature means improvising exactly those four.
Vocabulary
Seven categories, adopted from the public pattern catalogue Shape of AI. The names are taken deliberately: two of the seven — Governors and Trust builders — exist purely for oversight and correction, which is the right weighting for a house system and worth keeping as emphasis rather than reinventing.
Actions is Invoca’s own category, not Shape of AI’s. Shape of AI catalogues the
surrounding chrome — how you prompt, watch, and correct a model — and deliberately stops short
of naming the jobs themselves. Invoca’s agentic surfaces already do specific jobs today —
Signal AI summarizing a call transcript is shipped, named, and real — so this section
names the seven jobs its agentic features are known or expected to do: search, summarize,
draft, update, disambiguate, recommend, and the welcome/empty state every one of them needs on
day one.
Identifiers reads differently from the rest of this section. Avatar, Color,
Iconography, Name, and Personality describe how the AI presents itself, not an interaction —
there is no run to interrupt and no output to be confidently wrong about. Those five pages
use the lighter foundation shape (what it governs, choosing a value, constraints) instead of
the full pattern shape below. Everything else in this section — Wayfinders, Inputs, Tuners,
Governors, Trust builders, Actions — uses the full shape.
The four sections nothing else in this design language has
Agency tier
Every interactive page in this section (everything except Identifiers) states, plainly, how much authority the feature has. Naming the tier is what makes the confirmation-versus-undo question askable at all.
Escalating a feature’s tier is a decision, never a default. A feature that starts
suggesting and ends up acting unsupervised has crossed a line that needs recording, the same
way a token change does — see Constraints.
Outcome states
A component has interaction states — hover, pressed, disabled. An AI experience has outcome states, and every pattern page in this section carries a treatment for each one that applies: working, streaming/partial, confident and right, confident and wrong, uncertain, refused, empty, interrupted, degraded, rate-limited, stale. Confident and wrong is the primary case to design for, not the edge case. A wrong answer delivered with no visible seam is the characteristic failure of this whole section, and every pattern page states how a user notices and corrects one. “The model is usually right” does not count as an answer.Disclosure & recourse
Six questions, on every interactive page, answered or explicitly marked not applicable:- Does the user know this is AI, at the moment it matters?
- What did it use — call data, prior conversation, other records?
- How sure is it, and does that change what the user should do?
- How does the user check it?
- How does the user correct it, and does the correction persist?
- How does the user get out — reach a person, do it manually, turn it off?
Evaluation
Nothing in this section is built, so nothing in this section has been evaluated. Every page says exactly that rather than implying otherwise. Once a pattern ships, its page states the eval set, the metric and threshold, named failure classes from real output, and the date of the last run — a capability claim with no evidence behind it is the failure this section exists to prevent.Coverage, stated honestly
Unlike every other tier in this design language, there is no generated reference layer here
and there will not be one until something ships. A page in this section that looks thinner
than a Pattern page is not missing content — it is missing a product to observe.
Choosing where to look
Wayfinders
Getting a user to try the feature, and to keep going once they’re in it.
Inputs
Letting the user say what they want in their own words.
Tuners
Adjusting scope, tone, or depth before or during a run.
Governors
Showing and checking what the AI is doing.
Trust builders
Admitting what the system doesn’t know.
Identifiers
How the AI looks, sounds, and refers to itself.
Actions
The seven jobs Invoca’s agentic features are known or expected to do.
Constraints
Gaps
- No model, prompt, or vendor has been named for any pattern in this section except Signal AI’s existing transcript summary. Every Reference section below is a placeholder for content that does not exist yet.
- No evaluation framework exists. There is no eval set, no scoring rubric, and no named owner for deciding when an AI feature is good enough to ship.
- The escalation path from Suggests to Acts is undecided. Nothing says who approves a feature moving up a tier, or what evidence that decision requires.
- Whether Identifiers apply per-feature or once for all of Invoca’s agentic surfaces is undecided. Signal AI, a future assistant, and a future copilot panel could each need their own name and avatar, or could share one identity. See Identifiers: Name.