Financial services
Risk appetite has no limit for a wrong answer
Every risk appetite statement is built from numbers: limits, thresholds, tolerances. Ask what the stated tolerance is for an AI assistant giving a confidently wrong internal answer, and the statement goes quiet.
Cognatum Team · Oct 22, 2026 · 6 min read
A risk appetite statement is a document made of numbers. Credit concentration below ten percent of equity. Non-performing loans below three and a half percent. Unplanned downtime under four hours a quarter. Each line names a risk, a measure, and the point where somebody has to act.
Cognatum governs the entry
source · version · approver · permissions
Now ask a different question. Your assistant answers a policy question out of a procedure you retired in March. What is the stated tolerance for that? Most statements have no line for it. Most firms have no measure either.
The frameworks ask for measures
The Financial Stability Board set the shape of this work in 2013. Its principles cover four things. An effective risk appetite framework. An effective risk appetite statement. Risk limits. And the roles of the people who own them.
The Basel Committee followed in 2015. Its revised principles widen the board's role in overseeing risk systems. They also strengthen the guidance on risk governance. Pay is treated as one more way a board signals which risk taking it accepts.
Measurable is the load-bearing word
The Institute of Risk Management is blunter about why the measures matter. It quotes a head of risk on the point. The appetite statement is the hardest part of this work. And without clearly defined, measurable tolerances, the whole risk cycle is arguably at a halt.
The cascade stops at systems
Appetite cascades. A board sets an appetite, the risk function converts it into tolerances, tolerances become limits, and limits become controls that somebody tests. Guides to writing the statement describe tolerance levels as quantitative thresholds that trigger review, with worked examples like a non-performing loan ratio capped at three and a half percent.
The cascade works because every step can be counted. It reaches portfolios, systems, processes and vendors. It does not reach the content those systems answer from.
A data theme is not a knowledge theme
A published account from a UK public body shows how close the gap sits to the surface. The organization assessed its appetite across thirteen themes, including technology and data information and management, scoring each one from risk minimal to risk eager. It then found that its two hundred and fourteen recorded risks had largely been scored at service level, without reference to the corporate appetite at all.
So there was a theme for data. There was a scale, and there were owners. There was still nothing describing the answers people acted on.
Three reasons knowledge has no number
Nothing sizes the population
A tolerance is a share of something. Three and a half percent of a loan book works because the loan book has a size. Ask how many live knowledge entries your firm holds, how many are past review, and how many have been superseded, and almost nobody can produce the denominator. A percentage with no denominator is not a tolerance. It is a sentiment.
Nothing logs the breach
A limit only functions if a breach is an event somebody records. An unverified answer is generated and delivered in the same moment. Nobody files it, nobody counts it, and the person who received it usually acts on it. There is no breach register for answers, so a threshold has nothing to be measured against.
Nothing owns the measure
Every line in an appetite statement has a named owner. Documents have owners too. Answers do not, because an answer is assembled from several documents at the moment it is asked for, and the assembly belongs to no one.
Where a number does exist, it is not small
One domain has actually measured this. Researchers ran the first preregistered evaluation of retrieval based legal research tools and found that products from LexisNexis and Thomson Reuters each hallucinated between seventeen and thirty three percent of the time, against vendor claims of eliminating hallucinations.
Those are commercial products built on curated case law. An internal library assembled over twenty years, with no approver on most of it, is not a more favorable starting point.
A threshold is required before the measure is settled
The NIST AI Risk Management Framework does not prescribe risk tolerance. It points organizations to applicable requirements or tells them to define reasonable tolerances themselves, and it acknowledges that there is no consensus yet on robust and verifiable measures of trustworthiness.
The EU AI Act is more demanding for high risk systems. Article 9 requires testing against prior defined metrics and probabilistic thresholds appropriate to the intended purpose, and requires that residual risk be judged acceptable. The duty to hold a threshold is arriving ahead of any agreed way to measure one.
What a countable tolerance would need
Four fields make an answer countable. The entry it came from. The version of that entry. The named person who approved it. The date they approved it. An answer carrying those four can be sampled, aggregated and reported. An answer carrying none of them cannot appear in any statement, because it has no properties to measure.
Once entries carry that record, the population becomes visible and so does its condition. Items split into healthy and at risk, where at risk is wider than overdue for review:
- duplicates, where more than one entry answers the same question
- conflicts, where two approved entries disagree with each other
- gaps, where no approved entry exists at all
Those are counts. A board can read a count, set a threshold against it, and ask why the threshold moved.
Where Cognatum sits in this
Cognatum governs the knowledge rather than the model. AI runs seven of the eight steps in the Cognatum Knowledge Loop, and Approve is the one it does not: a named person signs off, and the entry carries that approver, the date and the source from then on. When a source changes, dependent entries are flagged and routed for re-check, and consumers keep the last approved version until a person signs the revision.
Conflicts and gaps, not averages
Where two approved entries genuinely disagree, both are held apart and routed to a person rather than averaged or settled by recency. Gaps are recorded as gaps instead of filled with something that reads well. None of this confers compliance with any framework. It produces the trail and the counts your reviewers ask for.
A test for your own statement
Open your risk appetite statement and count the lines that carry a number. Then count how many of those numbers describe anything your AI assistant answers from. In most firms the second count is zero, and that is the finding rather than the exercise.
Your company's knowledge isn't missing. It's unusable. Cognatum changes that. The eight steps and the single human gate are documented in full on the Cognatum Knowledge Loop.
Sources
- Principles for an Effective Risk Appetite Framework (fsb.org)
- Corporate governance principles for banks (bis.org)
- Risk appetite and tolerance (theirm.org)
- Tools for the job: understanding risk appetite (charterediia.org)
- Guide to Effective Risk Appetite Statements (metricstream.com)
- Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools (arxiv.org)
- Artificial Intelligence Risk Management Framework: AI RMF 1.0 (nist.gov)
- EU AI Act Article 9: Risk management system (europa.eu)
- The Cognatum Knowledge Loop (cognatum.ai)
Common questions
Questions this raises.
Can you set a risk appetite for AI answers?
Not in the usual form, and not until answers carry properties you can count. An appetite line needs a measure, a threshold and an owner. The practical first step is giving every answer an entry, a version, an approver and a date, then setting thresholds against duplicates, conflicts and gaps.
What is the difference between risk appetite and risk tolerance here?
Appetite is the level of risk a board is willing to accept in pursuit of its objectives. Tolerance is the acceptable variation around it, usually written as a quantitative threshold that triggers review. Neither one works on knowledge until the knowledge is countable.
Does a data governance program already cover this?
Only partly. Data governance covers structured data in warehouses, catalogs and master data systems. The answers an assistant assembles come mostly from documents, procedures, policies and chat threads, and those sit outside most data governance scopes.
Does the EU AI Act require a tolerance for AI answers?
For high risk systems, Article 9 requires a risk management system and testing against prior defined metrics and probabilistic thresholds, with residual risk judged acceptable. It does not supply the metric. Neither does the NIST AI Risk Management Framework, which leaves tolerance to the organization. No software confers compliance with either framework; it supplies evidence toward your own obligations.
Where should a knowledge tolerance be reported?
Alongside the operational risk lines it most resembles. The measures worth reporting are the share of live entries past review, the number of unresolved conflicts between approved entries, and the number of recorded gaps where no approved answer exists.