AI Governance: Introspection & Approvals

Business-object introspection (scope introspection:read)

Three read tools let an AI client learn the shape of the business before touching it: which objects exist (including the workspace's own custom fields and custom objects), what each attribute means, how objects relate, and which actions are executable — with their preconditions and current gear. A well-behaved agent calls these first and only proposes actions the gears actually allow.

Write tools under a review gear: pending / refused

When an MCP write tool targets an action whose gear is review, the call does not execute the side effect. It returns a structured result: either the proposal was queued for human approval (pending), or it was refused outright (refused) with a machine-readable reason_code and a suggested_action.

// proposed — waiting for a human
{
  "status": "pending",
  "proposal_id": "prop_9f2c…",
  "action": "issue_coupon",
  "expires_at": "2026-07-10T00:00:00Z",  // 7 days
  "message": "Queued for human approval in
              AI Activity & Approvals."
}

// refused — with a reason and a next step
{
  "status": "refused",
  "action": "issue_coupon",
  "reason_code": "recently_rejected",
  "suggested_action": "A human rejected this
    proposal recently. Do not re-propose;
    consider a different action or ask the
    operator."
}

Every refusal carries a reason_code in this shape plus a suggested_action — treat the codes above as the ones to branch on and the field as extensible. A refusal is a hard boundary read for you, not an error to retry around: never re-word a proposal to slip past a refusal.

Transparency primitives & the glass-box model inventory

The same glass-box discipline — every automated decision recorded and explainable, personal data kept out of model context — extends beyond governed actions: to experiment traffic optimization, to outbound LLM prompts, and to the classic-ML model inventory that powers ranking, prediction and detection across the product.

A/B traffic auto-optimization: the decision log

Recurring campaigns (automations and journeys) can opt in to Auto-optimize traffic over time, which reallocates delivery weights across experiment variants based on observed click performance. Every reallocation is auditable:

• Append-only decision log. Each hourly recompute writes one record with the weights applied and a human-readable rationale. Records are never rewritten — the log is the full history of why traffic looks the way it does.
• The configured split is never mutated.Your experiment's configured split stays untouched as the baseline; optimized weights are computed and logged alongside it.
• Exploration floor: 10%.No variant's weight ever drops below 10%, so every variant keeps receiving traffic and a reversal remains detectable.
• Even split until ≥100 delivered per variant.Weights stay at the configured split until every variant has at least ~100 deliveries — no reallocation on thin data, and no statistical “winner” is ever declared by this mechanism.

LLM PII masking at the provider seam

Before any prompt leaves Flash for an external LLM provider, member personal identifiers are masked — applied at the provider seam, so every AI feature passes through it:

• Deterministic pseudonyms.Email addresses, phone numbers and card-like digit runs are replaced with stable stand-in tokens. The same identifier always maps to the same pseudonym, so the model's reasoning across a prompt stays coherent.
• Default on, no configuration. Masking applies to every outbound prompt from every AI feature — there is nothing to enable, and it cannot be switched off from the product.
• Mask-only, never unmasked. The mapping is one-way at the output boundary: model output is never rewritten to reinstate real identifiers, so raw PII cannot round-trip through an LLM response.

The glass-box model inventory

Every predictive or statistical capability in the product is a classic, deterministic model — no black-box scoring, no sampling, so the same data always produces the same model. Each one carries an explicit honesty gate: below the gate it emits nothing rather than a guess, and every score ships with the numbers needed to recompute it. The current inventory:

None of these models ever acts on its own: they rank, score and flag, and everything that moves money or contacts a member still passes the governed action gears above.

For the full engineering detail on every model — method, inputs, output columns, honesty gates and consumption surfaces — see the Predictive Models Reference.