Industry: Project Management
Product: ModelRisk
Application: Task Prioritization

Criticality Is a Probability: Prioritizing Tasks under Schedule Uncertainty

On a 312-activity highway-bypass project, the deterministic Critical Path Method (CPM) identified a single sequence of 41 tasks as "the" critical path and a single finish date — 480 days. Take the same network and draw each task duration from its real distribution, and that tidy answer dissolves: the simulated completion is a right-skewed distribution that misses the 480-day plan in 62% of realizations.

The P50 is 487 days, the P90 is 517 days, and the probability of meeting the deterministic 480-day plan is just 38%. The deterministic path is not even reliably the critical one — the labeled critical chain (Path A) carries the project in only 61% of runs, while the alternative path through the bridge-pier sequence is critical in the other 39%. Criticality is not a label; it's a probability — and it is the metric that should drive where the project manager spends their first hour of every morning.

A heavy civil contractor used ModelRisk to rebuild its task-prioritization process around criticality index (the probability a task lies on the critical path) and schedule sensitivity (the impact of a task's variance on the project's variance), giving every task in the WBS a defensible score rather than relying on the deterministic critical-path label.

Why a single critical path is the wrong question

Standard CPM treats each task duration as a point estimate. Take the same network, draw each duration from its actual distribution, and the path through the network changes between iterations. The "critical path" is not one path — it is a probability distribution over paths.

Three modeling choices made this rigorous:

Task durations. PERT with optimistic/most-likely/pessimistic estimates from the project controls historian — e.g. bridge-pier cap pour PERT(8, 14, 28) days, with the long right tail driven by re-pours and weather windows. Tight prefab tasks were given truncated Normal distributions; tasks dependent on agency approvals were given Lognormal because regulatory wait time is famously right-skewed.

Discrete risk events. The model allowed for low-probability, high-impact firings — e.g. a 12% probability of a geotechnical re-design adding 10–45 days to the pier-foundation chain, modeled as a Bernoulli switch on a PERT delay.

Inter-task correlation. Tasks sharing a weather window or a shared crew were given a 0.5 rank correlation on their duration draws — verified at 0.50 in the run, not merely asserted. Independence is the wrong default — when the rain falls, it falls on every outdoor task at once.

Which tasks actually drive the finish date

The same 50,000-iteration run that produced the completion distribution also produced a criticality index for every task in the network, not just the labeled-critical 41:

Several deterministic-critical-chain tasks carry a criticality index above 70% — the permit hearing at 82%, approach-fill compaction at 78% — which is the right place for management attention. But the chart also reveals tasks not on the deterministic critical path with criticality indices as high as 74% (the bridge-pier cap pour) and 68% (the geotechnical re-design) — meaning they lie on the critical path in most realizations and were previously unmanaged. That is the finding the deterministic plan suppressed.

Where the schedule variance actually comes from

A tornado over the project completion-duration variance, not the deterministic plan, gave the priority order for management attention:

Two of the top three drivers — geotechnical re-design event and agency-permit wait time — were single-task lines that the deterministic plan treated as "small risks" because their probability was low. In a probabilistic frame they jump to the top because their conditional impact is large. This is the standard EMV-vs-tail blind spot, and the chart is the antidote.

Two prioritization policies, two delivery distributions

The team compared the as-planned task priority order against two alternatives:

1. Criticality-index-driven priority — re-rank attention by simulated criticality, putting the three high-CI off-deterministic-path tasks into the active-management tier.
2. Schedule-sensitivity priority — re-rank by impact on completion variance (the tornado above), allocating buffer-crew capacity to the top drivers.

Both policies improved the P90 over baseline. The criticality-driven policy cut P90 by 15 days (517 → 503); the schedule-sensitivity policy cut it by 20 days (517 → 498) by directing buffer-crew capacity at the highest-variance tasks, and it pulled the P50 in as well (487 → 483). On a project with $80k/day liquidated damages, the 20-day P90 reduction is worth roughly $1.6M of expected late-finish exposure, justifying the policy change many times over.

What changed

• Off-deterministic-path tasks moved into active management based on criticality indices as high as 74% (bridge-pier cap pour) and 68% (geotech re-design) — none of them flagged by the deterministic CPM.
• P90 finish date pulled in by 20 days under the schedule-sensitivity policy, worth ~$1.6M of expected late-finish exposure.
• Daily stand-up agenda restructured around the top criticality-index tasks, not the deterministic critical path — a list that changed three times during execution as upstream tasks resolved their uncertainty.

ModelRisk Functionality Used

• PERT and LogNormal duration distributions for the 312 WBS tasks, fitted against the project controls historian via the distribution-fitting dialog; truncated Normal for the prefab work where variance is genuinely low.
• Bernoulli-switched PERT delays for the discrete risk events (geotech re-design, supplier slip, agency hearing) — modeling the conditional severity correctly rather than averaging it into the base duration.
• Shared-factor correlation (ρ ≈ 0.5, verified empirically at 0.50) between the competing paths so that a single bad-weather or short-crew iteration lengthens both at once — diversification of the schedule's two arms is partial, not free.
• Criticality index output computed for every task in the network across 50,000 iterations — the single metric that surfaced the off-deterministic-path priorities.
• Tornado over P90 completion variance that selected the schedule-sensitivity priority policy, cutting P90 by 20 days while improving the P50.

The question is not "what is on the critical path" — it is "what is the probability each task ends up on the critical path, and how big is each task's contribution to the P90." Monte Carlo simulation in ModelRisk turns task prioritization from a labeling exercise into a quantified daily decision.