Lesson 4 · AI Governance Basics

Proportional risk tiering and human oversight

Length: 28 minutes across 7 sections
You will be able to apply: The Highest Dimension
You will produce: Commit Statement
You will work: 3 gated questions

Personalize the practice

Apply this to your environment

These details adapt the application prompts and coach questions. They do not affect your score.

RoleSectorOrganization sizeAI maturity

Reading

The tier is the highest dimension, not the average of them.
Core question

risk tiering

Assign the tier before you see the scoring method

Required practice

Each OmniCorp use case has incomplete but decision-relevant facts. Assign the risk tier you would defend now, before reading the Highest Dimension framework.

A meeting summarizer turns internal project notes into a recap for the same team, and mistakes are reversible in minutes.Jo Halvorsen uses AI to draft ordinary client project updates, reads every one herself, and sends about thirty a month.An ambient transcription tool creates draft clinical notes for clinician review using patient records.A hardship pre-screening workflow leaves unflagged applications in the standard queue with no review of the negative path.

Complete each part before committing.

Proportionality is the only thing that makes governance survivable. Apply the same scrutiny to a meeting summariser and a credit decision and one of two things happens: the credit decision is under-examined, or the meeting summariser is examined so heavily that people stop declaring their AI use at all. Both outcomes are worse than doing nothing.

So tiering is necessary. It is also the single most gameable control in governance, because the person assigning the tier is usually the person who benefits from a lower one.

Next: The pathology: Rounding Down

Section 2 of 7

The pathology: Rounding Down

Watch a tiering conversation and you will see the same movement every time. The system has one alarming characteristic and five unremarkable ones. The discussion spends ninety seconds on the alarming characteristic, establishes that it is not as bad as it first sounded, and then settles on a tier that reflects the five. Nobody lied. The reasoning was even quite careful. The tier is still wrong.

PrincipleRounding Down is the reflex of assigning risk by the general feel of a system rather than by its worst dimension, in a direction that reliably reduces the burden of approval. It is not fraud. It is the ordinary human preference for a tractable answer, operating on a scale where the tractable answer is always the lower one. The tell is the word overall: any sentence beginning overall this is fairly low risk is averaging, and averaging is the failure.

The counter is a scale that does not permit averaging.

Interactive modelRounding Down versus Highest Dimensionmatrix · 4 elements

Averaging (many low, one high)

A moderate tier that feels reasonable and understates the actual risk.

The tier is the maximum across dimensions, never the average; one high dimension puts the whole system at the top.

Averaging (many low, one high): A moderate tier that feels reasonable and understates the actual risk.
Averaging (all moderate): A moderate tier that happens to be correct but arrived by the wrong method.
Highest Dimension (one high): A high tier driven by the dimension that sets it, with oversight designed around that dimension.
Highest Dimension (all low): A low tier that is genuinely low because no single dimension is elevated.

Next: The Highest Dimension

Section 3 of 7Framework

The Highest Dimension

Definition

Score an AI system independently on six dimensions: impact if wrong, autonomy, data sensitivity, reversibility, scale, and regulatory exposure. Score each one on its own, without reference to the others. The system's tier is the highest score on any single dimension. Never the mean, never the mode, never an overall impression. One dimension at the top puts the whole system at the top, and that is the entire point of the instrument.

This is not negotiable, and it is not a heuristic to apply when convenient. A system that is unremarkable in five dimensions and severe in one is a severe system. The five comfortable dimensions are not mitigation. They are the reason nobody looked at the sixth.

When to Use It

Tier before build, before launch, and on any change to autonomy, audience, data or volume. Volume is the trigger most often skipped: a system that moves from four hundred to forty thousand decisions a month has changed dimension even though nothing about its logic has changed.

Re-tier when a system that was advisory starts being followed without question. Autonomy is behavioural, not architectural: a recommendation nobody ever overrides is a decision, whatever the design document says.

How to Apply It

Score each dimension separately and write the score down before discussing the next one.
Take the maximum, and say the number out loud before anyone argues about mitigations.
Record the dimension that set the tier, because that is what oversight must be designed around.
Apply mitigations to reduce a dimension, and re-score it, rather than adjusting the tier directly.

Dimension	Low	Moderate	High
Impact if wrong	Rework absorbed internally	A customer is materially inconvenienced	Money, health, liberty or eligibility is affected
Autonomy	A person composes the final output	A person reviews before release	The system acts without review
Data sensitivity	Public or internal	Confidential business or personal data	Special category, children, or criminal data
Reversibility	Undone in minutes by the operator	Undone with effort and a conversation	Cannot be undone once acted on
Scale	Tens of instances a month	Thousands	Continuous, at population scale
Regulatory exposure	No specific regime engaged	Sector guidance applies	A statutory obligation attaches to the decision

Worked example 1 of 3

OmniCorp Financial proposed an AI system to pre-screen hardship applications and flag those likely to qualify for a payment holiday. The design group tiered it as moderate. Marisa Delgado made them score the dimensions one at a time.

Marisa Delgado: Impact if wrong.
Product lead: High. A missed flag means someone in hardship does not get contacted.
Marisa Delgado: Autonomy.
Product lead: Moderate. An adviser reviews every flag before contact.
Marisa Delgado: And the ones it does not flag? Who reviews those?
Product lead: Nobody. They stay in the standard queue.
Marisa Delgado: Then on the negative path autonomy is high, and impact is high. This is a high-tier system. Say the number before we discuss what we can do about it.

The oversight that followed was designed around the dimension that set the tier: a mandatory sample review of unflagged applications, not more review of the flagged ones. The team's original proposal had concentrated all its scrutiny on the path that already had a human on it.

Why This Works

Scoring dimensions in isolation defeats averaging structurally rather than by asking people to resist it. Once the six numbers exist on paper, taking the maximum is arithmetic, and arithmetic does not care how reasonable the room finds the system. The judgment is confined to each individual dimension, where it belongs and where it can be argued about specifically.

Recording which dimension set the tier is what makes the oversight proportionate rather than merely heavy. A high tier driven by scale needs monitoring and thresholds. A high tier driven by irreversibility needs a checkpoint before the irreversible step. Designing the same oversight for both wastes effort in one case and misses the risk in the other.

Worked example 2 of 3Optional depth

Dr. Naomi Ellery tiered an ambient transcription tool at OmniCorp Health that produces draft clinical notes for clinician review. Impact moderate, autonomy moderate, reversibility moderate, scale moderate, regulatory exposure moderate. Data sensitivity: high, unavoidably, because the content is clinical. One dimension, and the system is high tier. The oversight designed around that dimension was about storage, access and deletion rather than about note quality, which the moderate dimensions already handled adequately. The tier did not make the project harder. It made it point at the right thing.

Worked example 3 of 3Optional depth

Priya Raghunathan re-tiered an OmniCorp Logistics route-optimisation system that had been low tier for two years and had performed faultlessly. Nothing about its logic had changed. Its volume had risen from two hundred to nineteen thousand route decisions a day, and dispatchers had stopped overriding it because it was almost always right. Two dimensions had moved, scale and autonomy, without a single line of code changing. The system was now high tier and had been for months, governed by a tiering decision made when it was a pilot.

Edge Cases and NuancesOptional depth

A system's tier can differ between its positive and negative paths, as at OmniCorp Financial, and the tier is the higher of the two. Mitigations legitimately reduce a dimension, but only if they are implemented and verified, so a planned mitigation reduces nothing. Composite systems must be tiered as assembled rather than component by component, because three low-tier components chained together produce autonomy nobody scored. And a system can be high tier and still proceed: the tier determines the oversight required, not permission, and treating a high tier as a refusal is what teaches people to round down in the first place.

Interactive modelThe six dimensions of risk tieringladder · 6 elements

Regulatory exposure

Whether a statutory obligation attaches to the decision the system influences.

Each dimension is scored independently; the system's tier is the highest score on any single dimension.

Regulatory exposure: Whether a statutory obligation attaches to the decision the system influences.
Scale: The volume of decisions or outputs, from tens a month to continuous population-level.
Reversibility: Whether an action can be undone in minutes, with effort, or not at all once taken.
Data sensitivity: The classification of data the system touches, from public to special category.
Autonomy: Whether a person composes, reviews, or never sees the output before it acts.
Impact if wrong: The consequence of an incorrect output, from internal rework to harm to health or liberty.

Knowledge check

A system scores moderate on five dimensions and high on data sensitivity alone. A colleague proposes moderate tier because the system is mostly moderate. What is the correct tier and why?

Answer first, then check.

Next: Common Failure Modes

Section 4 of 7

Common Failure Modes

Failure modeTiering against the mitigation. What it looks like in the moment: someone scores impact as high, and immediately adds that there is a review step so it is really moderate. The correction is spoken in the same breath as the score, so no high score is ever written down and the six numbers on the page are already netted off. The cost when this happens: the tier now describes the system you intend to build rather than the one that exists, and when the review step is quietly dropped for throughput in month four, nothing in the record shows the tier depended on it. The correction: score the raw dimension, take the maximum, then apply mitigations and re-score explicitly, so the dependency is visible.

Failure modeThe frozen tier. What it looks like in the moment: nothing at all. A system tiered at pilot has grown twentyfold, its recommendations are now followed automatically, and no meeting is scheduled because no change was made. The absence of an event is what hides it: tiering is triggered by change requests, and growth is not a change request. The cost when this happens: your highest-volume systems are the ones governed by their earliest, most optimistic assessment, which is precisely inverted. The correction: put a re-tier trigger on volume and on override rate, and treat a falling override rate as an autonomy increase requiring reassessment.

Next: The scale end to end

Section 5 of 7

The scale end to end

Alan Brixmoor tiered a proposed OmniCorp Public system that would rank benefit review cases by likelihood of error, so caseworkers could work the highest-value queue first. The team had proposed low tier: it makes no decisions, it only orders a list.

Scored separately: impact if wrong, high, because a case ranked low is a case worked late and the affected person is on a reduced payment while they wait. Autonomy, high, because nobody reviews the ordering. Data sensitivity, high, health and financial. Reversibility, moderate. Scale, high, every case in the directorate. Regulatory exposure, high, statutory decision timescales attach.

Five high dimensions on a system described in good faith as only ordering a list. Alan's point to the team was not that they had been careless: it was that ordering a list is a decision about who waits, and that this is invisible precisely because no decision is recorded anywhere. The oversight designed around it was queue-age monitoring with a hard escalation on any case aged beyond a threshold, regardless of rank, plus a monthly sample of the bottom of the queue. The system launched. It launched with a tier that told the truth about it.

Decision point

Jo Halvorsen at OmniCorp Studio uses AI to draft client-facing project updates. Impact if wrong is moderate, autonomy is moderate since she reads every one, data sensitivity is moderate, reversibility is moderate, scale is low at maybe thirty updates a month, and regulatory exposure is low. Her studio manager proposes moderate tier and light oversight. Then Jo mentions that two clients are public sector and their updates are appended verbatim to a statutory progress report. What is the tier, and what oversight follows?

Write your answer before you continue

Confidence before seeing the analysis

What evidence or assumption could change your decision?

Commit, calibrate, and name contrary evidence first.

Next: Self-check

Section 6 of 7

Self-check

Knowledge checkTake one AI system near your work and score the six dimensions separately, writing each down before moving on. Take the maximum. If that number is higher than the tier the system currently carries, you have found the gap this lesson exists to find.

Mark the level that describes you today. Nothing is submitted.

Behaviour	Ready	Developing	Not yet
Deriving the tier	Given the OmniCorp Studio case, takes the maximum across dimensions and lands on high for the two public-sector engagements.	Given the same case, recognises the regulatory dimension is high but settles on moderate overall.	Given the same case, reasons from volume and personal review to a moderate or low tier.
Handling mitigations	Scores the raw dimension, then re-scores explicitly once a mitigation is implemented and verified.	Notes mitigations alongside the score without separating raw from mitigated.	Nets the mitigation off in the same breath as the score, so no high score is recorded.
Keeping the tier current	Sets re-tier triggers on volume and override rate, and acts when the override rate falls.	Re-tiers on formal change requests only.	Treats the launch tier as settled until someone raises a concern.

Next: Commit

Section 7 of 7

Commit

Commit Statement

Complete every line in your own words, then sign and date it. Attach your six scores and the dimension that set the tier.

I commit to

I will stop

I will start

I will know I succeeded when

Signed Date

Window	Field application
Days 1 to 7	Score the six dimensions for one system, separately and in writing, and compare the maximum with its current tier.
Days 8 to 21	Find one system whose oversight was designed around a different dimension from the one that set its tier, and redirect it.
Days 22 to 30	Check the override rate on one advisory system. A falling rate is an autonomy increase, and it requires a re-tier.

Six dimensions and a maximum will give you an honest tier for a system you can see. They will not tell you how to hold that tier when a senior sponsor disputes it, how to evidence the assessment to an external reviewer, or how to run a portfolio where two hundred systems are tiered by forty different people to the same standard. Making tiering consistent and defensible across an organisation is the capability the paid programs develop next.

DisclaimerGeneral guidance only. All organisations and people named in this lesson are fictional. Regulated organisations should confirm requirements with a qualified professional before relying on this material.

Required practice must be complete.

Learner feedback

Did this change what you can do?

What will you use at work?