Local AI pill

Product intent

Local AI should become available on as many capable devices as possible, but a multi-gigabyte model must never download without informed consent. The product therefore separates instant value from durable capacity:

1. Included AI lets a new person run the first useful operation immediately.
2. Local AI is always visible as a red pill until the person enables it.
3. One explicit action starts the model download while the person continues using CareerVector.
4. Once ready, Smart Routing uses Local AI for task classes where its measured quality is sufficient.
5. Your AI remains an optional user-owned route for stronger or specialized work. It is not a paid CareerVector tier.

“Sporewright” and provider machinery are not user-facing concepts. The product calls the overall behavior Smart Routing and the device route Local AI.

Placement

The pill lives in the workspace subheader immediately to the left of the Scraping pill. The two form a right-aligned device-participation group, with Scraping remaining the rightmost control:

[workspace actions]                         [Local AI · Off] [Scraping · Ready]

Local AI is device-scoped rather than workspace-scoped, so it remains visible in editable and read-only workspace shells. Enabling it once makes the cached model available to all eligible workspaces on that browser profile or desktop installation; workspace privacy and routing policy still apply to every task.

State language

Red is a deliberate, persistent behavioral nudge. It must not imply that the whole product is broken: Included AI may still be serving the current work.

Consent interaction

Opening the red pill shows a compact consent panel, not the full AI settings modal. The primary copy should be concrete:

Run routine AI on this device. Free and private. Download about 1.6 GB while CareerVector is open.

The panel names the recommended model, approximate download size, expected storage use and relevant device support before the person acts. Its primary button is Download & enable Local AI. A secondary Choose another model link opens the existing advanced AI settings.

The same consent can include an enabled-by-default option to keep Local AI ready with background model updates. That option is visible before acceptance and remains reversible. Disabling Local AI does not delete the model; removal is a separate explicit action.

Before consent the app may perform only lightweight capability, quota and manifest checks. It may not fetch model weights. After consent it should request persistent browser storage where supported and begin the download immediately.

Background behavior

In a browser, “background” means the transfer continues while the person uses other parts of CareerVector and while the page remains alive. Browsers do not reliably permit a multi-gigabyte download to continue after the tab or browser is closed, so the UI must not promise that. An interrupted transfer resumes the next time CareerVector opens.

The desktop app may continue while its process remains open, including when its window is hidden if the operating system permits it. Neither target should require the person to stare at a progress screen.

Model replacement is transactional: keep the last validated model routable, download the candidate separately, validate it, then atomically promote it. Failure leaves the working model intact. Downloads are resumable and deduplicated per device.

Included AI continues serving feasible foreground work throughout download and preparation. Readiness changes the set of routes available to Smart Routing; it does not interrupt the person's current task.

Routing and explanation

The pill reports device readiness. Smart Routing still decides whether a local model is appropriate for an individual operation from the tensor address, declared policy and learned evidence. Being green is not a promise that every task runs locally.

Short explanations should connect state to benefit without exposing internal math:

• “Using Local AI for this routine task.”
• “This task needs a stronger model; using Included AI.”
• “Included AI is busy. Local AI will handle compatible steps.”
• “18 tasks completed locally; shared capacity saved.”

User-fixed provider choices, privacy gates and exclusions remain authoritative. Local execution requires both device consent and permission for the task's data class.

Adoption sequence

The nudge should be persistent but not obstructive:

1. Deliver one useful result through Included AI with no setup.
2. Keep the red Local AI pill visible from first use.
3. After an Included AI operation, explain that Local AI keeps the free product responsive and private for suitable work.
4. Download only after the explicit click; let the person keep working.
5. Celebrate readiness briefly, then let the green status become quiet.
6. Offer Your AI only when it provides a concrete quality, availability or specialization benefit.

Success is not the click-through rate alone. Measure consent, completed and resumed downloads, time to readiness, task-class quality, local completion rate, shared-capacity displacement, battery/memory pressure, disable/remove rates and unsupported-device frequency. Smart Routing learns from execution outcomes; the UI must never manufacture consent to improve those numbers.