Methodology

The Executive Attention cards surface the handful of items that a senior leader on the Texas HHSC Medicaid engagement should know about this week. Each card is a synthesis of multiple underlying signals — a procurement forecast, a CMS release, an anomaly in enrollment data — packaged with a short “why it matters” explanation and a concrete action prompt.

Ranking is produced by Claude (Anthropic) with the dashboard’s current state as input. The model is prompted to weigh three factors:

• Strategic relevance to the Accenture TX HHSC contract position (incumbency, scope adjacency, amendment exposure).
• Time-sensitivity: how quickly the reader has to act before the window closes.
• Impact magnitude: scale of the dollar, policy, or reputational stakes.

Each card carries a confidence indicator (high / medium / low) reflecting source reliability and data recency. Items are regenerated on a schedule; users can mark cards as “reviewed” to dim them without removing them from the feed.

The ranking prompt has access to the signals, alerts, and risk-opportunity feeds below, plus a curated set of procurement and policy sources.

Signals from the Edge tracks external developments that may shape Accenture’s position before they show up in enrollment or spending data. The feed pulls from six categories: procurement, policy, regulatory, OIG, CMS, and legislative.

In the live pipeline, items are gathered from public sources including Texas SmartBuy, CMS.gov, the Federal Register, congressional committee pages, HHS OIG reports, and state procurement portals (Ohio, Louisiana, Georgia, and others). A scheduled crawler normalizes each item into a common shape — title, summary, category, relevance band, affected states, and a source URL — and dedupes against prior runs.

Relevance (high / medium / low) is assigned by a lightweight classifier tuned to the Texas HHSC engagement. A human reviewer can promote or demote items; overrides persist across crawls.

The category filter pills filter client-side; as the feed grows, the same filters will move to server-driven queries.

Intelligence Alerts flags significant changes in Medicaid data that deviate from expected patterns. Alerts span four types: enrollment change, spending spike, policy change, and quality alert.

The detection pipeline compares each state’s monthly CMS data against a rolling baseline and against a peer-state reference group (matched on expansion status and managed care penetration). Deviations beyond a configurable threshold are lifted into the feed with a severity level (critical / high / medium / low).

Alerts are then passed through Claude for a short human-readable summary and a suggested interpretation. The model does not change the severity or add alerts that the detector did not flag — its role is explanatory.

Each alert carries a link to the originating source (CMS dataset, KFF tracker, 1115 waiver page, or similar) so the reader can verify the underlying data.

The Risk & Opportunity Matrix plots tracked items on a two-axis grid: opportunity (x-axis) against risk (y-axis), with bubble size encoding impact. The four quadrants are Backburner (low risk, low opportunity), Pursue (low risk, high opportunity), Defend (high risk, low opportunity), and Executive Priority (high risk, high opportunity).

Each item is scored on three 0–100 dimensions by Accenture intelligence analysis, drawing on the signals feed, alert history, and contract knowledge:

• Risk — downside exposure if the item plays out unfavorably.
• Opportunity — upside if the item is pursued or shaped.
• Impact — overall magnitude (dollars, scope, reputation).

The ranked list on the right of the chart orders items by a weighted priority score:

priority = impact × 0.50 + opportunity × 0.35 + risk × 0.15

Weighting intentionally favors impact and opportunity, so that large positive moves rank above medium-sized defensive plays. The weights are configurable and will be tuned against real outcomes once the live pipeline has a few months of history.

New items are proposed by the signals pipeline and promoted into the matrix after a short human review.

A few things to keep in mind when reading the dashboard:

• AI-assisted, not AI-authoritative. Where AI is used (Executive Attention ranking, alert summarization, the Ask Claude panel), the model is always working from underlying data that a human can verify. Treat AI output as a prioritization and explanation layer, not as ground truth.
• Source recency varies. CMS monthly enrollment data typically lags 60–90 days. Procurement forecasts update on state-specific cadences. Alerts are only as fresh as the most recent data refresh.
• Confidence indicators are directional.High/medium/low confidence bands reflect source quality and data recency, not statistical confidence intervals.
• No PHI or PII. All data on the dashboard is aggregated at the state level or higher. No individual-level records are ingested, stored, or displayed.