Stop asking AI for a single percentage without inputs

Executives love success probability slides: "72% chance we hit Q3." Without documented assumptions, that number is theater—uncalibratable, unfalsifiable, and forgotten after the meeting.

Responsible AI project success probability works differently. It combines historical signals from your portfolio—throughput, aging, burn, approval latency—with simulate_scenario outputs that compare explicit futures (delay_release, cut_scope, add_people, freeze_project). The product is not prophecy; it is structured uncertainty leaders can debate and log.

WKFGo MCP tools expose simulation and brief data with stated data gaps. Treat probability as a conversation aid, not a substitute for resolve_decision_item.

Problem: fake precision erodes trust

Teams fail forecasting when they:

Export a dashboard green/red and call it "on track"
Ignore known blockers because status fields lag reality
Run one scenario, never log which path was chosen
Present AI output as statistical fact without sample size

A number without a decision_log receipt repeats the same optimism next quarter.

Framework: signals → scenarios → range → decide

Collect signals — flow_cfd, flow_aging, finance burn, review queue depth, dependency chains, capacity heatmap
Run scenarios — simulate_scenario for top executive forks with assumptions listed
Express ranges — "If we cut scope A, date likely moves 1–2 weeks; if not, slip 4+ weeks" beats "68%"
Decide and log — log_decision with forecast fields; calibrate later via decision_log

AI helps assemble signals and narrate trade-offs. Humans choose which scenario matches risk appetite.

Historical signals that matter

Past slip magnitude on similar release types
Team velocity variance sprint-over-sprint
Recurring decision_inbox themes (capacity, scope creep)
Post-mortem wiki pages from prior failures—patterns, not anecdotes

What AI cannot honestly do

Guarantee success from narrative confidence
Predict external vendor behavior without data
Replace missing task discipline with confident prose

Say data gaps aloud in briefs.

Solution: probability workflow in WKFGo

Monthly portfolio review

For each at-risk project:

Pull get_project_report or portfolio_overview for FACTS
Run simulate_scenario for delay vs descope vs add people
Compare five dimensions in output—time, scope, cost, morale risk, dependency exposure (as returned; do not invent metrics)
CEO/CPO frame success as conditional: "Success probability rises if we cut feature X"

Pre-commit gate for customer dates

Before publishing a date externally:

Document top three risks with task links
Attach scenario comparison to steering wiki via update_wiki_page
If probability of hit is low, add_decision_item rather than silent hope

Calibration loop

When the date passes, compare outcome to logged forecast. executive_brief trend memory shows whether last decisions moved target metrics—accountability without blame theater.

Communicating probability to stakeholders

Use language stakeholders can audit:

Conditions — "Hit date if integration tasks finish by May 12"
Indicators — weekly CFD trend, not vanity completion counts
Triggers — "If review queue > N, escalate to inbox"

Avoid single percentages unless tied to a model you will calibrate.

Building a probability review agenda

Schedule a 45-minute probability review before external date commits—not after marketing publishes. Attendees: product owner, engineering lead, finance delegate. Pre-read includes get_project_report, dependency summary, and two simulate_scenario outputs attached to the steering wiki.

Agenda:

List assumptions each scenario requires (vendor delivery, review capacity, scope freeze)
Identify early indicators you will watch weekly—CFD slope, inbox themes, burn variance
Agree trigger that forces re-simulation (e.g., two consecutive sprints missing carry-over target)
log_decision which scenario matches appetite; note what would falsify the forecast

This ritual keeps probability language honest. Stakeholders hear conditions, not magic percentages.

Anti-patterns

One scenario only — always compare at least two forks
Ignoring finance and capacity signals
Letting AI round up because the audience wants good news
Skipping log_decision after leadership picks a path

Probability language should include assumptions and confidence—not single percentages from thin air. simulate_scenario documents forks; leadership picks and logs the bet.

Run a monthly audit: do AI answers cite task IDs from MCP? If not, tighten prompts to require citations before any stakeholder-facing draft leaves the team.

Permission-aware tools mean the same question gets different grounded answers for PM vs contractor—that is a feature, not a bug. Train teams to expect scoped truth.

FAQ

Is this Monte Carlo simulation?
WKFGo simulate_scenario projects structured futures with stated assumptions—not necessarily full Monte Carlo unless your deployment extends it.

Can probability be automated weekly?
Monitors can raise decision_inbox when signals cross lines; humans interpret and resolve.

What if data is incomplete?
Briefs should label gaps; narrow confidence instead of fabricating precision.

How does this relate to Gantt dates?
Gantt shows plan; probability discusses likelihood plan survives contact with reality.

Trade uncertainty with eyes open

Use scenarios and historical signals to discuss success honestly—then log the bet you make.

Try it now

Put these patterns on live project data—not slide decks.