Line Speed Root Cause Narrative

Try the Agent

Select a preset query to run the agent. Each response synthesizes data across three tables — line speed metrics, sensor event logs, and digital twin forecasts — and produces a ranked, evidence-based narrative ready for plant leadership.

Agent Configuration & Data

Every agent configuration generated by the Knowledge Work Foundry includes the query logic, data schema, and sample output. What you see below is the actual KWF output for this workflow candidate — three tables, two signal-ranking queries, one threshold rule.

• Agent Queries
• Table Schemas
• Sample Output

-- KWF Agent Query Step 1: Detect material line speed deviation
-- Parameters: {line_id}, {shift_id}, {date}
-- Flag if pct_deviation > 0.10; proceed to signal ranking only if flagged.

SELECT
  line_id, shift_id, date,
  avg_speed, target_speed,
  ROUND((target_speed - avg_speed) / target_speed, 4) AS pct_deviation
FROM  line_speed_metrics
WHERE date      = '{date}'
  AND   line_id  = '{line_id}'
  AND   shift_id = '{shift_id}';

-- Step 2a: Rank sensor events by downtime contribution
SELECT
  station,
  SUM(duration_minutes)                                  AS total_downtime,
  ROUND(SUM(duration_minutes) / total_event_time, 4)   AS pct_of_incident
FROM  sensor_events
WHERE line_id  = '{line_id}'
  AND   shift_id = '{shift_id}'
  AND   date     = '{date}'
GROUP BY station
ORDER BY total_downtime DESC;

-- Step 2b: Check digital twin variability threshold (flag if > 0.15)
SELECT forecasted_variability
FROM  digital_twin_forecasts
WHERE line_id  = '{line_id}'
  AND   shift_id = '{shift_id}'
  AND   date     = '{date}';

-- Business logic:
-- Rank stations by total_downtime; compute pct share of incident.
-- If forecasted_variability > 0.15: include as contributing factor.
-- Assign impact weights combining sensor downtime and digital twin risk.
-- Narrative is assembled from ranked signals + recommended action.

-- Table 1: Line speed metrics (one row per line/shift/date)
CREATE TABLE line_speed_metrics (
  line_id    STRING    NOT NULL,
  shift_id   STRING    NOT NULL,
  date       DATE      NOT NULL,
  avg_speed  DOUBLE    NOT NULL,   -- units/min actual
  target_speed DOUBLE  NOT NULL    -- units/min target
) USING DELTA;

-- Table 2: Sensor event log (one row per station event)
CREATE TABLE sensor_events (
  line_id           STRING   NOT NULL,
  shift_id          STRING   NOT NULL,
  date              DATE     NOT NULL,
  station           STRING,              -- packaging | filler | labeler | ...
  event_type        STRING,              -- jam | fault | maintenance
  duration_minutes  INT
) USING DELTA;

-- Table 3: Digital twin material feed forecast (one row per line/shift/date)
CREATE TABLE digital_twin_forecasts (
  line_id                 STRING   NOT NULL,
  shift_id                STRING   NOT NULL,
  date                    DATE     NOT NULL,
  forecasted_variability  DOUBLE   -- fraction; flag if > 0.15
) USING DELTA;

Cognitive Labor Breakdown

The KWF analyzes every workflow candidate across three layers of cognitive labor. The split below is the actual KWF analysis output for the Line Speed Root Cause Narrative workflow for the Manufacturing Operations Manager role. Note that the judgment allocation is higher here than in a simple point-query workflow — because multi-signal synthesis requires more interpretation and contextual framing before it can be handed off.

The execution layer is mechanical: pull from three tables, apply two ranking rules, and write a structured output sentence by sentence. The logic has been in the manager's head for years — it just hasn't been encoded anywhere. So the Operations Manager re-encodes it manually after every incident.

The judgment layer is where the Operations Manager's experience actually matters. Deciding whether a 12% deviation caused by packaging station jams is a fluke or the third occurrence this week — and what that implies for the shift handover — requires someone who knows the line. The agent provides the data; the manager provides the context.

This split is also why the judgment allocation (40%) is higher here than in a typical point-query workflow. The narrative construction task for multi-signal synthesis contains more interpretation work embedded in it, even at the execution layer. The agent takes the mechanical part; the manager retains the rest.

Business Value Translation

Freeing the execution layer of this workflow does not "save 2.5 hours a week." It changes what Operations Managers do after every line speed incident. Instead of spending the hours before the morning briefing building the narrative, they arrive having reviewed it — and the performance review starts at the judgment question: what to do about the cause, not what the cause was. The consistency gain compounds: every shift, every line, every manager is working from the same evidence-based narrative, not from memory.

There is also a strategic upside this model does not capture directly: earlier, more consistent explanations reduce the lag time between a line speed incident and the corrective decision. If an Operations team using this agent identifies recurring packaging station failures two shifts earlier than they would have manually — because the narrative is already assembled before the morning briefing rather than being built during it — the avoided production delay and the reduced leadership escalation time represent additional value that compounds across a plant over a full year.