For Risk Managers

30 Practical Ideas for Risk Managers to Stay Cognitively Sovereign

Your SAS Risk AI model produces elegant stress test results. Your IBM OpenPages dashboard summarises board-level findings. But you cannot see the assumptions buried in the code or the tail risks the training data never captured. Risk managers who outsource their thinking to AI systems lose the very judgement that spots emerging threats before they become catastrophes.

These are suggestions. Take what fits, leave the rest.

Download printable PDF

Before You Run the Model

Write down your risk hypothesis before opening the toolbeginner
Before you run SAS Risk AI or Palantir, write one paragraph describing what risks you expect to find and why, using only your professional instinct.
Document every assumption the AI tool made on your behalfbeginner
After the model setup step, export or screenshot the parameter choices, correlation matrices, and distribution assumptions that the AI selected without your explicit input.
Identify which historical periods trained your modelintermediate
Know exactly which years of market data, operational events, or loss history your Azure AI or SAS model learned from, then list three major risks that occurred outside those years.
Run a backwards stress test on last year's emerging risksintermediate
Take the three risks your organisation flagged as emerging last year. Run them through your current AI model inputs and see which ones it would have missed or downweighted.
Map which model inputs you truly understand versus inheritedbeginner
Create a two-column table: inputs you personally validated against source data versus inputs copied from vendor templates or previous models you did not build yourself.
Ask your model builder what happens if correlations breakintermediate
Sit with the person who configured your OpenPages or Azure scenario tool and ask them to show you what correlation assumptions fail if markets stress simultaneously.
Stress test the model's stress testsadvanced
Take the extreme scenarios your AI tool generated and check whether they are actually extreme or just statistical outliers from calm periods in the training data.
Require a sign-off on every parameter you did not personally challengebeginner
Before you present results, identify which model inputs you accepted without questioning and ask your quantitative team member to defend each one in writing.
Record your concerns about the model in a separate documentbeginner
Keep a risk model limitations log separate from your board report, noting uncertainties, assumptions you distrust, and gaps where you know historical data misses real threats.
Compare your model output to competitor or peer modelsadvanced
If possible, obtain anonymised stress test results from peer organisations using different AI tools and identify where your model diverges significantly from theirs.

During Results Interpretation

Reject results that feel too tidybeginner
When your ChatGPT summary or Azure dashboard presents risk findings in clean categories with no cross-cutting uncertainties, ask your team to identify what messiness was smoothed away.
Manually walk through three specific risk scenariosintermediate
Pick three scenarios your model ranked as high-impact. Trace them through your actual business by hand, asking business unit heads whether the AI's cascade assumptions match reality.
Identify risks ranked low by the model but flagged by your operational staffbeginner
Collect anecdotal risk concerns from compliance, operations, and business teams over the past quarter. Compare them to your model's risk ranking and document disagreements.
Test the model's blind spots by asking what it cannot measureintermediate
Ask your technical team what types of risk cannot be expressed in your model's input structure because they do not fit numerical parameters.
Benchmark model outputs against past predictionsintermediate
Pull your risk model outputs from twelve months ago and compare them to what actually occurred. Document where the model was wrong and whether those failures cluster in certain risk types.
Require human-written justification for every top-ranked riskbeginner
For your three highest-risk scenarios from the model, refuse to use the AI's auto-generated explanation. Ask your team to write why they believe each risk is material in their own words.
Create a separate emerging risk list outside the modelbeginner
Maintain a rolling list of risks that do not fit your model's structure because they are novel, cross-cutting, or based on weak signals rather than statistical patterns.
Check whether the model punishes uncertainty or hides itadvanced
Review how your SAS Risk AI or Palantir tool handles low-confidence estimates. Confirm that uncertain risks are flagged as uncertain rather than downweighted and buried.
Interview the model's edge casesadvanced
Ask your technical team to show you the scenarios that the model struggled to score because they fell outside its training distribution. These are often your real risks.
Reverse-engineer the model's risk rankingadvanced
Take your top ten risks from the output and manually recalculate impact and probability using your own data. Identify which results changed when you did not rely on AI aggregation.

Board Reporting and Ongoing Governance

Write a separate risk model uncertainty summary for the boardintermediate
Do not let your OpenPages or Azure summary auto-fill the board risk section. Write a one-page statement of what the model cannot see and where you lack confidence.
Tell the board which risks you removed from the model and whyintermediate
In your board pack, list three to five risks you considered but excluded from your AI-driven model because they were unmeasurable, and explain why you still believe they matter.
Assign accountability for each assumption to a named personbeginner
In your model governance, require that every major parameter choice has a named risk manager who will defend it to the audit committee and update it annually.
Schedule an annual model stress test you design yourselfintermediate
Once per year, run a scenario through your SAS or Azure model that you created without vendor input, based on your own risk intuition and board feedback.
Require your board to challenge one model input per meetingbeginner
Build into your board risk agenda a standing question that asks board members to suggest one assumption they distrust, then commit to testing it.
Document every time you overrode the model's outputbeginner
Keep a log of risk decisions where you disagreed with your AI model's ranking or recommendation, with your reasoning. Review this log quarterly to see if overrides cluster in certain areas.
Create a model risk register separate from your operational risk registerintermediate
Maintain a dedicated log of risks that arise from the model itself: assumption failures, correlation breaks, data quality gaps, tool failures that could affect your risk reporting.
Conduct annual interviews with staff who distrust the modelintermediate
Identify risk staff, business unit heads, or compliance officers who have expressed scepticism about your AI model and hold structured interviews to capture their concerns.
Build a manual scenario library separate from AI-generated scenariosintermediate
Maintain a collection of stress scenarios built by your team based on historical crises, geopolitical events, and sector-specific shocks that do not rely on AI pattern matching.
Plan your model replacement before vendor lock-in occursadvanced
Before your SAS, Azure, or Palantir implementation becomes indispensable, document what you would need to do to rebuild your risk framework using different tools or manual methods.

Five things worth remembering

Related reads

The Book — Out Now

Cognitive Sovereignty: How To Think For Yourself When AI Thinks For You

Read the first chapter free.

No spam. Unsubscribe anytime.