Decide, Act, and Learn

A resolved decision names the chosen option, why it serves the important outcomes, what tradeoff is accepted, and what happens next. Until action has an owner and date, analysis remains a preference rather than a decision.

Use a decision declaration:

We choose [option] because it gives the best likelihood of [ranked outcomes] while keeping [constraint or downside] within [limit]. [Owner] will take [first action] by [date]. We will monitor [indicators], stop or adapt at [triggers], and review on [date].

The declaration should be short enough to repeat and precise enough to test. If the team cannot say what was chosen or what happens Monday morning, the process has not resolved.

An implementation intention is an if-then plan that links a recognizable situation to a specific action. It reduces dependence on memory, motivation, and improvisation when the moment arrives.

Vague intention	Implementation intention
“Exercise more.”	“If it is 07:00 on Monday, Wednesday, or Friday, I put on running clothes and walk outside for 25 minutes.”
“Escalate serious incidents.”	“If customer data may have been exposed, the incident lead pages security within five minutes and freezes outbound changes.”
“Check the pilot.”	“If the Friday dashboard shows activation below 35%, the product owner pauses acquisition and opens the diagnostic review.”
“Save for tax.”	“When client payment clears, 25% transfers automatically to the tax account.”

Choose cues that are observable and timely. “If the project looks troubled” invites argument; “if projected delivery slips more than ten working days” can trigger.

Assign actions to roles or named people. “The team will respond” hides ownership; “the incident lead pauses deployment” does not.

A trigger is a condition that activates a response, while a contingency is the prepared action for that condition. Together they preserve flexibility without reopening the whole decision whenever conditions change.

Build a trigger table:

Signal	Threshold	Response	Owner
Customer retention	Below 30% for two cohorts	Pause expansion; investigate onboarding and value	Product director
Cumulative spend	Reaches £100,000 before milestone 3	Freeze new commitments; finance review	Programme owner
Safety incident	Any serious uncontained event	Stop immediately; activate response plan	Safety lead
Demand	Above 2,000 weekly orders for four weeks	Add second shift under approved plan	Operations manager
Supplier delay	More than 15 working days	Activate backup supplier	Procurement lead

Include upside triggers. Decision systems often prepare only for failure, then improvise when success creates capacity, quality, or cash problems.

Avoid trigger inflation. Monitor the small set of signals connected to assumptions, hard constraints, or switching thresholds. A dashboard of fifty metrics can hide the three that should change action.

A decision journal is a record made before the outcome is known. It protects the original reasoning from hindsight and creates material for improving estimates and process.

Capture:

Field	Example
Decision	Launch limited service to Manchester customers
Owner and date	Product director, 24 July
Values and constraints	Learn demand; protect privacy; cap loss at £40,000
Options	Full launch, pilot, delay, cancel
Forecast	65% chance retention exceeds 35%; 10% chance loss reaches cap
Critical assumptions	Acquisition channel resembles planned scale
Choice and tradeoff	Pilot; slower revenue in exchange for bounded learning
Emotional state	Excited and under investor timing pressure
Triggers	Stop at privacy breach or cap; scale above stated thresholds
Review	Operational weekly; decision-quality review after six weeks

Record a probability even when it feels uncomfortable. “Likely” cannot later be calibrated; 65% can.

Keep the journal immutable or preserve revisions with dates. New information should update the plan, but it must not erase what was believed before the evidence arrived.

A review cadence is the schedule for checking execution, assumptions, and decision quality. The cadence should follow how quickly useful evidence arrives and how fast harm can compound.

Use three layers:

Review	Purpose	Typical frequency
Operational	Detect triggers, failures, and implementation drift	Real time, daily, or weekly
Decision update	Reassess assumptions when meaningful evidence arrives	Milestone, monthly, or threshold-based
Process review	Examine reasoning and calibration after outcomes mature	Quarterly or after a defined result

Do not review long-term outcomes every day. Noise encourages reactive changes and destroys the experiment. Conversely, a monthly safety review is too slow when harm can escalate in minutes.

Set a minimum commitment period when appropriate: “Do not change the campaign for two weeks unless a hard-stop threshold is crossed.” This protects learning from ordinary variation.

Close stale decisions. When the review date arrives, explicitly continue, adapt, stop, or replace. Repeatedly saying “monitor” without a next threshold leaves the choice unresolved.

Decision quality concerns whether the choice used a clear frame, values, options, relevant information, sound reasoning, and commitment. Outcome quality concerns what actually happened. Uncertainty means the two are related but not identical.

Use a four-cell review:

Process	Outcome	Interpretation
Good	Good	Reinforce process; do not assume every assumption was correct
Good	Bad	Check whether adverse result was within the forecast; improve only where evidence supports
Poor	Good	Do not reward luck; repair the process before stakes rise
Poor	Bad	Separate reasoning failures from unavoidable uncertainty

A 20% risk occurs one time in five over repeated comparable decisions. When it occurs, the decision was not automatically foolish. Ask whether the probability, downside limit, and mitigation were reasonable before the result.

Likewise, profit does not prove wisdom. A reckless bet can win once and teach a dangerous lesson.

Calibration means that events assigned a probability occur at roughly that rate over time. Among many 70% forecasts, about 70% should resolve true if the estimates are calibrated.

Track forecasts in probability buckets:

Forecast bucket	Number of forecasts	Number true	Observed frequency	Interpretation
50%	20	11	55%	Close to calibrated
70%	30	18	60%	Overconfident in this sample
90%	20	14	70%	Strong overconfidence

The Brier score is a simple accuracy measure for binary forecasts. For each event, square the difference between probability p and outcome o, where o = 1 if it happened and o = 0 if it did not.

Three forecasts:

Forecast	Probability p	Outcome o	(p − o)²
Supplier delivers on time	0.70	1	0.09
Retention exceeds target	0.60	0	0.36
Audit finds no major issue	0.80	1	0.04
Mean Brier score			(0.09 + 0.36 + 0.04) ÷ 3 = 0.163

Lower is better, with 0 representing perfect forecasts. Compare against your own prior record or a sensible baseline; a score without context does not explain which estimates need improvement.

Calibration improves through frequent, resolvable forecasts, clear outcome definitions, base rates, and honest review—not by rounding every estimate toward 50%.

An update changes a forecast or action because new evidence has arrived. Hindsight bias instead makes the outcome seem as if it should have been obvious all along.

Use an update log:

Date	New evidence	Old estimate	New estimate	Action effect
1 Aug	Pilot activation at 44%	60% chance target holds	72%	Continue pilot
15 Aug	Refund rate rises to 8%	72%	55%	Test pricing explanation
29 Aug	Revised flow cuts refunds to 4%	55%	68%	Prepare staged expansion

Judge the update by what the evidence justified at that time. Do not give full credit for changing only after the answer was already visible.

At review, label findings:

• process lesson: independent estimates exposed a key disagreement;
• model lesson: customer acquisition cost responded more strongly to price than assumed;
• execution lesson: the assigned owner lacked authority to pause spend;
• random variation: an ordinary adverse branch occurred within the forecast;
• new condition: a regulation changed after the decision.

Each category suggests a different improvement. Not every bad result calls for a new rule.

A worked example shows how resolution and learning join. A professional-services firm is choosing whether to introduce a four-day week for one business unit.

Declaration: run a twelve-week pilot with 60 employees because it offers useful evidence on retention and productivity while protecting client service. The managing director owns the choice. Hard constraints are no missed regulatory deadlines and no decline in urgent response below the agreed standard.

Forecasts: 70% chance output per employee remains within 5% of baseline, 60% chance employee retention intention rises at least 10 points, and 15% chance client response time breaches the hard limit.

Implementation intentions: if urgent response falls below 95% in any week, the operations lead adds a Friday rota immediately. If it remains below 95% for two weeks, the pilot pauses. If output and client standards hold for eight weeks and retention intention rises, the owner prepares a broader staged test.

Journal: the team records that recruitment pressure and enthusiasm may be encouraging optimism. It preserves an independent finance estimate and the dissenting view that seasonal workload makes the test unrepresentative.

Outcome: output remains stable, retention intention rises 14 points, but urgent response breaches once and recovers after the rota. The adverse branch occurred and the contingency worked, but one observation cannot validate a 15% forecast. The review checks the estimate’s ex-ante basis—comparable trials, workload data, and assumptions—and reserves calibration judgment for many comparable forecasts. The team expands to one more unit during a busier period rather than declaring universal success.

The result becomes a better next decision, not a trophy for the original advocates.

A decision practice improves through repeated small forecasts and honest reviews, not only occasional major choices. Use thirty days to install the minimum habits.

Week	Practice	Evidence
1	Record five pending choices with owners, dates, and outcomes	Five clear decision statements
2	Add probability forecasts, assumptions, and triggers	At least ten resolvable forecasts
3	Convert chosen actions into if-then plans	Owners and cues visible
4	Review outcomes, process, and calibration	One written process improvement

Final checklist:

• [ ] The resolution names option, rationale, tradeoff, owner, and date.
• [ ] The first action is observable rather than aspirational.
• [ ] Important cues have specific if-then responses.
• [ ] Triggers cover failure, success, spend, and hard constraints.
• [ ] The journal captures forecasts before outcomes are known.
• [ ] Review cadence matches feedback speed and harm.
• [ ] Process quality is judged separately from outcome quality.
• [ ] Probabilities resolve to clearly defined true or false results.
• [ ] Calibration and forecast accuracy are tracked over many cases.
• [ ] Updates preserve the old estimate and identify new evidence.

The mature decision-maker does not seek a record of never being wrong. The aim is clearer commitments, bounded harm, faster learning, and a process that becomes better calibrated every time reality answers back.