AI assistant (ask about your programs, propose changes)

Sparkus lets you connect an AI assistant (such as Claude) to your programs so you can ask, in plain language, how a program is set up — instead of clicking through admin screens to find out. It can also propose a configuration change for you. It cannot make that change happen: proposing is as far as the assistant ever goes, and an administrator has to approve the proposal before anything changes.

What you can ask it

The assistant answers from your live configuration, and can search Sparkus’s own documentation and cite what it drew from. Today it can:

List your programs — the programs you have access to.
See your tenant at a glance — headline counts for your company: how many programs, users, participants, and cohorts you have, plus people-hub figures (total and recently-active users, pools, recent imports, deactivated users).
Explain a program’s configuration — a single “program config” answer that covers:
- General — name, code, program type, dates, timezone.
- Matching — whether a matching item is configured and the support-giver pool summary (pool name + how many profile attributes are configured).
- Branding — colors, logo and hero imagery, and whether the program uses its own theme or inherits the company default.
- Languages — which languages the program offers.
- All program settings — the full list of a program’s settings as key/value pairs, so you can ask “what is X set to?” (secret values — API keys, passwords, tokens — are automatically redacted; you’ll see that a setting exists, never its secret value).
- AI configuration — the program’s guardrails, prompt, and Sidekick settings (with the same secret redaction).
- Reminders and nudges — the program’s reminder scenarios and nudge rules (when they fire, to which roles, and whether they’re enabled).
- Sessions — a breakdown of the program’s sessions by status (how many are scheduled, completed, and so on).
- Calendar and language — the program’s calendar event settings and language configuration (default translation language and language settings).
Explain a program’s matching setup — which matching algorithm and criteria weights are chosen, the matching stats (how many matched, pending, etc.), and which support-giver roles are enabled. Useful for “why isn’t matching working?”
Summarize cohort statistics — membership counts and average goal progress per cohort. Very small cohorts are hidden to protect anonymity.
Diagnose a program — ask “why isn’t matching working?” or “is this program ready?” and get a plain-language health check: whether a matching algorithm is selected, whether support-giver roles are enabled, whether there are support receivers, and whether cohorts exist and have members. Each finding is flagged as a blocker, a warning, or OK.
Give you a direct link — a deep link straight to a program’s settings page (only for programs you can actually open).
Check participant progress (only where enabled) — per person, how many goals they have and how many are completed, their share/handshake status, and average progress. Never the goal text or any other private content.
List who’s in a pool (only where enabled) — per pool member, their name, their role in the pool, and their membership status. Never their email or other contact details.

The last two are off by default. Each needs its own data-protection sign-off before it turns on — see “Reading about people” below; a program config, matching, or cohort answer never needs one, because it never names a person.

This list grows over time as more of Sparkus becomes answerable.

What it can propose (and never apply on its own)

Beyond answering questions, the assistant can propose a change for one of the following:

A program setting — any existing settings key and its new value.
Which matching algorithm a program uses.
Whether a competency set is active (turning it on or off requires you personally to hold the Competencies-Manage permission — the assistant can propose the change, but only an admin with that permission can approve it).
Whether a training package is active for your company.
Adding or removing one person from a coach/participant pool.

That is the complete list — outside these, the assistant has nothing to propose, and it will tell you so rather than guess.

Proposing is all it ever does. The assistant cannot save, confirm, or apply a change itself — there is no button, shortcut, or “trusted” case that skips this. Every proposal, including a pool-membership change touching a real person, is queued as a pending item that an administrator reviews and approves on an approval page in AdminAlpha before it takes effect. Once approved, the change is recorded in a permanent history alongside every other admin-made change, so it’s always clear who approved what and when.

Reading about people

Two of the assistant’s reads touch information about specific people rather than aggregate counts: participant progress, and the pool roster (who’s in a pool). Both are off by default, and turning either one on takes three separate approvals, all of which must be in place before the assistant can read anything:

Sparkus’s own operational switch for the feature.
Sparkus’s data-protection sign-off clearing the capability at all.
Your own company’s sign-off, naming the exact thing it is agreeing to.

Participant progress and the pool roster are consented separately — your company signing off on one does not open the other, because they’re different information about different things. Every time the assistant reads either one, that access — who read it, when, and how many rows — is written to an access log, even though the log itself never records the people’s individual details.

What it will not do

The assistant is deliberately limited so it’s always safe to use:

It never applies a change. It can only propose one, as described above — every proposal waits for a human approval.
It sees only what you see. Every request runs as you: your company, your permissions. It can’t reach another company’s data, and it can’t see anything you couldn’t already see in AdminAlpha.
It never sees private content. Goal text, 360-degree feedback, session notes, and AI-conversation content are off-limits, no matter what else is enabled.
Secrets stay secret. Any setting whose value is a credential is redacted before it ever leaves Sparkus.

How to connect it

An administrator opens AI assistant access in AdminAlpha (/panel/mcp-access), generates a personal token, and pastes it into their AI client (for example, Claude Desktop). The token proves who you are — it carries no password and no special powers of its own; Sparkus re-checks your real permissions on every request. Keep the token private, like a password; generate a new one whenever you need to.