AI Brain

The AI decision layer — heads, autonomy policy, HITL, explainability, and the reversibility window.

On this page

What AI Brain is

AI Brain is the decision layer that turns canonical events into ranked, explainable, governed recommendations and (when policy allows) autonomous actions. It is not a chatbot bolted onto a PPM tool. It is a first-class actor in the platform with its own audit trail, its own provenance metadata, and its own governance.

Heads

A "head" is a use-case-specific surface of the AI plane. Each head is a separate model deployment, a separate audit channel, and a separate autonomy policy.

| Head | Output | Default autonomy | |---|---|---| | Status Summarizer | Plain-language exec summary | auto (read-only) | | Risk Predictor | Risk probability + drivers | auto_if conf ≥ 0.9 else recommend | | Budget Forecaster | EAC range with confidence interval | recommend | | Resource Optimizer | Allocation suggestions | recommend (never auto) | | Planner (WBS) | Hierarchical WBS proposal | require_human always | | Prioritizer | Ranked project list | recommend (never auto) |

The defaults are conservative on purpose. Resource Optimizer never auto-applies because it touches careers. Prioritizer never auto-applies because it touches strategy. Planner always requires a human because the cost of a wrong WBS compounds.

Autonomy levels

There are three autonomy levels per head per portfolio:

  • auto — the head executes without human review. Only available

for read-only heads (the Status Summarizer is the canonical example). Available for read-write heads only with explicit per-tenant approval.

  • recommend — the head produces ranked suggestions, with

confidence and counterfactuals, in an inbox. A human approves each one before any state change.

  • require_human — the head produces drafts only. Even after

approval, every action runs through a human review step.

The autonomy level is set per portfolio per head. A risk-tolerant team can run Forecaster on auto_if conf ≥ 0.9; a regulated team can keep everything on require_human and use AI Brain purely for draft generation.

The recommendation contract

Every recommendation carries:

  • The suggestion (what the head thinks should happen)
  • A confidence score, calibrated against historical eval data
  • The drivers — the features that pushed the recommendation in

this direction

  • A counterfactual — what would change if the most influential

driver flipped

  • The rationale, expressed in plain language with citations to the

events that supported the conclusion

  • The proposed actions, expressed as references to pre-declared,

idempotent action templates

  • The model identity (model id, version, signature)
  • The feature freshness — how stale the inputs were at the moment

of inference

  • The residency region the inference ran in

This contract is the core of the explainability story. Every field is auditable. Every field is queryable. None of it is generated post-hoc.

The reversibility window

Auto-executed actions carry a default 7-day reversibility window. During that window the action can be reversed in a single click, which generates a compensating event and rolls the state forward with the reversal recorded as a peer fact in the event log (reversals are never in-place mutations).

HITL inbox

Every recommendation that is not on auto lands in a human-in-the-loop inbox. The inbox shows the standard explainability panel — Headline + Rationale, Confidence + Calibration, Drivers, Counterfactual, Decision Trail — and supports bulk approve / modify / reject with a reason captured. The reason is fed back to the feedback service for labeling and drift monitoring.

Kill switches

Each head has a per-tenant kill switch. If something looks wrong — calibration drift, an unexpected spike in HITL rejection — flipping the kill switch disables the head immediately. Pending recommendations stay in the inbox. New ones do not generate.

Read further