Tornado charts and schedule sensitivity analysis

Each bar represents one source of uncertainty — usually a task duration, but it can also be a discrete risk event or a cost driver. The length of the bar measures how strongly that source is correlated with, or contributes to, variation in the project finish date. Long bars at the top are the dominant drivers; short bars at the bottom barely matter, no matter how nervous they make people feel.

Crucially, the ranking is about influence on the outcome, not about how big or how risky a task looks in isolation. A task can have a wide duration range and still sit near the bottom of the tornado if it is rarely on the critical path. Conversely, a task with a modest range can dominate the chart if almost every simulated schedule routes through it.

Sensitivity is derived statistically from the simulation, not assumed up front. During a Monte Carlo run, the engine samples a duration for every task from its three-point estimate (optimistic, most likely, pessimistic — often a PERT-Beta distribution), recomputes the critical path, and records the resulting project finish date. After thousands of iterations you have, for every task, a paired record of "the duration this task took" and "the finish date that run produced."

From those pairs the engine computes a sensitivity measure for each task. Two are common. The duration sensitivity index is the correlation between a task's sampled duration and the project finish date — a high correlation means that whenever that task runs long, the project runs long. The criticality index is the percentage of iterations in which the task landed on the critical path. The tornado bars are these measures, sorted descending.

Work top-down. The tasks and risks at the top of the tornado are where reducing uncertainty buys you the most schedule confidence, so they are the first candidates for active management: fast-tracking, adding a feeding buffer, dual-sourcing a vendor, running a de-risking experiment earlier, or simply tightening the estimate with more data. Reducing the range on a top driver visibly narrows your P50-to-P80 spread; doing the same to a bottom driver does almost nothing.

Read it together with the rest of the analysis. A task that is high on the tornado and high on the criticality index is a true constraint — protect it. A task that has a wide range but a low criticality index is volatile but off the critical path; it deserves a watching brief, not a crash plan. This is also how you avoid the classic mistake of pouring effort into the loudest, most-discussed task when the data says it is not what is moving the date.

These outputs answer different questions and are strongest together. The probabilistic completion S-curve (cumulative distribution) tells you how likely each finish date is and where your P50, P80, and P90 land. The criticality index tells you how often each task is on the critical path. The tornado chart tells you which tasks and risks drive the variance you see in that S-curve — it is the diagnostic that explains why the curve is as wide as it is.

In a healthy workflow you commit to a date from the S-curve, justify it with the criticality index, and decide what to act on from the tornado chart. Skipping the tornado is how teams end up with a defensible P80 date and no idea which three things to fix to improve it.

CritPath AI produces tornado and sensitivity analysis directly from its Monte Carlo engine. Every simulation samples task durations from PERT-Beta three-point estimates, injects discrete risk events, optionally applies duration correlation, and records per-task results across the run. From that it computes the duration sensitivity index and the criticality index for every task and renders a ranked tornado chart alongside the P50/P80/P90 completion distribution.

Because the AI copilot is grounded in your real dependency graph, it does more than draw the chart: it can explain why a specific task tops the tornado, what reducing its range would do to your P80, and how a gate decision or a vendor slip would re-shape the ranking after a retroactive reschedule. All of this is in one modern web app at $10 per user per month, with AI usage billed separately by metered usage — not a $10,000-plus desktop seat.