A board deck is not a decision.
Decisions that hold up later rest on a process you can show. Measure how noisy your board is, then write the next decision down properly.
A data-based board decision is a decision whose question, options, evidence, probability, dissent and review date are recorded at the time it is taken, so that the process can be judged independently of the outcome. The record is written before anyone knows how things turn out, and it is filed with the minutes.
Only 20% of respondents say their organisation excels at decision making. McKinsey Global Survey, 2019, n=1'259. 1
Adoption is solved. Value is not.
In McKinsey's November 2025 survey, 88% of respondents reported regular AI use in at least one business function. 39% attributed any EBIT impact to AI, most of them under 5%. 2
Most companies already use the tools. The gap is in how their output enters a decision. A model can draft a market memo in ten minutes, and the board still decides on the strongest voice in the room. Value starts when that memo meets a named question, a written probability and a date on which someone checks whether it was right.
Better analysis is not the lever. Better process is, by a factor of six.
Lovallo and Sibony studied 1'048 major business decisions. Process mattered more than analysis for decision effectiveness, by a factor of six. 3 In that study, process covered habits such as discussing uncertainty openly and giving room to views that contradicted the most senior person present. AI has no feelings - perfect! Are you ready for your board's Jarvis?
Across 136 studies, mechanical prediction was almost always as accurate as or more accurate than clinical judgment. Clinicians were substantially more accurate in only 6% to 16% of studies. 4
Agree the criteria and their weights before the proposal arrives. Score the case against them, then discuss.
CFOs' 80% confidence intervals for stock market returns contained the realised return only 36% of the time, across more than 13'300 forecasts. 5
Ask for a low case and a high case on every key number. If the range surprises nobody, widen it.
Of more than 16'000 projects in Flyvbjerg's database, only 0.5% (one in 200) came in on budget, on time and delivered the promised benefits. 6
Before the plan, ask how comparable projects went. Start from that reference class and adjust only for what is really different.
Your board is noisier than it thinks.
In a noise audit at one insurer, two underwriters pricing the same case differed by a median of 55%. Executives had expected about 10%. 7 Noise of that size stays invisible as long as nobody sees a colleague's number for the same file.
Boards have the same blind spot: members talk, converge and never learn how far apart they started. Run this before a discussion. Each member answers five fictional cases alone, one number per case.
Send it by email or chat. Each member answers alone and sends back an answer code. Keep this tab open until the codes are in.
Show the formula
for every pair of members (i, j) on one case: d_ij = |a_i - a_j| / ((a_i + a_j) / 2) noise index (case) = median of all d_ij board noise = mean of the five case indices
Pairs where both answers are 0 are skipped. An answer above ten times the case median is flagged "check unit". Bands (Dacide working assumption, not a published standard): below 15% low, 15 to 35% material, above 35% high.
Executives at the insurer in the Noise study (2021) expected about 10%. 7
Not computed
Noise is a separate problem from bias, and a shared process reduces it.
| Case | n | Min | Median | Max | Noise index |
|---|
Computed on this device.
Answer alone, before any discussion. There are no right answers on this page. Your name is never part of the code.
Send this code to your facilitator by any channel. It holds the session code and your five numbers, nothing else.
For comparison: the insurer's executives expected about 10% between two people judging the same case. 7
Noise needs at least two people. Send the member link to your board.
Computed on this device.
Judge the decision, not the outcome.
Every outcome mixes the quality of the decision with luck. A board that grades only results will reward a lucky gamble and punish a sound bet that missed, and its next decisions get more timid. The useful review asks what was knowable at the time, and whether the probability written down was honest.
Calibration can be trained: less than one hour of training in probabilistic reasoning improved forecast accuracy (Brier scores) by 6 to 11% across four years of forecasting tournaments. 8 In the IARPA tournament, the Good Judgment Project's forecasts beat the official control group by 60% in year one and 78% in year two. 9 Where a decision has many plausible futures, a spread of outcomes says more than a single number (how we generate one).
- Write probabilities, not adjectives."Likely" can mean 55% to one director and 85% to the next (an illustration, not a study result). A number can be checked later.
- State the base rate first.How often does this kind of project succeed elsewhere? Adjust from there, in writing.
- Name what would change your mind.Two or three observable facts that would stop or reverse the decision, agreed before the vote.
- Review at a fixed date.Put it in the calendar on the day of the decision, with the record attached.
A decision without a record is an opinion with a budget.
Fill in thirteen fields before the vote, from the question to the rationale. Export it, file it with the minutes or in your Company Brain, and open it again on the review date.
0 of 6 elements of a defensible record
- At least two options
- Evidence with a source
- A probability
- Dissent addressed
- A kill criterion
- A review date
Write the decision question first. The export buttons work once it is there.
Computed on this device. Nothing is saved unless you tick the box.
Saved only in this browser. Clear it any time.
The business judgment rule protects the process, not the outcome.
Art. 716a OR lists the board's non-transferable and inalienable duties, including overall management (Oberleitung) and overall supervision (Oberaufsicht). Seven apply to every company limited by shares, an eighth to listed companies. 10 A board can hand the running of the company to management. It keeps the duty to supervise, and supervision depends on information that arrives in time and in a form the board can question.
Oberaufsicht without data is a signature, not supervision.
Art. 725 to 725b OR set the board's duties on threatened insolvency, capital loss and over-indebtedness, in force since 1 January 2023. 11 These duties are triggered by numbers, so the board needs a liquidity view frequent enough to see a threshold coming.
Swiss courts review business decisions with restraint if they were taken in a proper process, on an adequate information basis and free of conflicts of interest. 12 That review looks at how the board reached its decision: what it knew, whom it heard, who had an interest. A record made at the time answers those questions. One reconstructed after a dispute begins carries little weight.
The EU AI Act's AI literacy duty (Art. 4) has applied since 2 February 2025 and, since July 2026, requires measures to support AI literacy rather than a fixed level. It reaches Swiss companies only where they place AI systems on the EU market, deploy them in the EU, or their output is used in the EU. 13 Check those three conditions before you build a training programme around Art. 4.
In a liability case, the question is not what you decided. It is what you knew, and when.
- What was known, and when.Record field 04, evidence with source and date.
Which figures and reports did the board have on the day, and how old were they?
- Who was informed.Record field 02, decision body and date. Distribution list in the minutes.
Did every member receive the papers in time to read them, or only the committee?
- Which options were weighed.Record field 03, options with upside and main risk.
Was there a real alternative on the table, including doing nothing?
- Was there a conflict.Record field 09, conflicts declared.
Did anyone have an interest in the result, and was it declared before the vote?
- What was minuted.Record fields 08 and 13, dissent and rationale.
Do the minutes show the discussion, the dissent and the reasons, or only the resolution?
This page explains; it does not replace legal advice for your board.
Where this breaks.
- A record documents the process; a bad decision stays bad.
- Minutes written after the fact are worth little, because their author already knows the outcome.
- No tool replaces the board's judgment, and none takes over its liability.
Questions.
Is a decision record legally required?
No statute prescribes this format. It helps because Swiss courts reviewing a business decision look at the process, the information basis and conflicts of interest (BGE 139 III 24). A record made at the time shows all three. It does not replace the minutes, and it does not decide how a court will rule.
Does AI in the board room create liability?
The duties in Art. 716a OR stay with the board, whatever prepared the papers. The exposure comes from treating AI output as evidence without checking it. Record which inputs came from AI, who checked them and what grade they earned.
Who writes the record?
The company secretary or the chair's office drafts the question, options and evidence before the meeting. The board adds probability, dissent and kill criteria in the meeting. The chair approves it with the minutes. The person proposing the decision should not be its only author.
How do we start with a noise audit?
Take 20 minutes at the start of a meeting. Open the Facilitator tab, create a session and send the member link. Everyone answers the five cases alone and returns a code. Compute, show the result, then discuss what made the answers differ.
Does the AI Act apply to a Swiss board?
Only through what the company does. The AI literacy duty in Art. 4 reaches Swiss companies where they place AI systems on the EU market, deploy them in the EU, or their output is used in the EU.
Sources.
- McKinsey & Company. Decision making in the age of urgency. McKinsey Global Survey, 2019. PO
- McKinsey & Company. The state of AI in 2025. Survey, November 2025. PO
- Lovallo, D. and Sibony, O. The case for behavioral strategy. McKinsey Quarterly, 2010. PO
- Grove, W. M. and Meehl, P. E., 1996; Grove, W. M. et al. Clinical versus mechanical prediction: a meta-analysis. Psychological Assessment, 2000. WE
- Ben-David, I., Graham, J. R. and Harvey, C. R. Managerial miscalibration. Quarterly Journal of Economics, 2013. WE
- Flyvbjerg, B. and Gardner, D. How Big Things Get Done. 2023. Review in the Wall Street Journal, 6 February 2023. WE
- Kahneman, D., Sibony, O. and Sunstein, C. R. Noise: A Flaw in Human Judgment. 2021. Review and summary (Aldous, UC Berkeley). WE
- Chang, W., Chen, E., Mellers, B. and Tetlock, P. Developing expert political judgment: the impact of training and practice on judgmental accuracy in geopolitical forecasting tournaments. Judgment and Decision Making 11(5), 509-526, 2016. WE
- Tetlock, P. E. and Gardner, D. Superforecasting: The Art and Science of Prediction. 2015. Summary of the Good Judgment Project results (AI Impacts). WE
- Swiss Confederation. Code of Obligations, Art. 716a OR. Version in force since 1 January 2023.WE
- Swiss Confederation. Code of Obligations, Art. 725 to 725b OR. In force since 1 January 2023.WE
- Swiss Federal Supreme Court. BGE 139 III 24, consideration 3.2. 2013.WE
- European Union. Regulation (EU) 2024/1689 (AI Act), Art. 4 AI literacy, with the Digital Omnibus on AI in force since 27 July 2026 (Lewis Silkin, 2026).WE