Where AI sits in the work

Two systems can use the same model and carry entirely different obligations, because the obligations follow the position of the model in the workflow rather than the model itself. Before you classify anything, work out where it sits.

Four positions

Upstream of a person. The model produces something a person then works with: a draft, a summary, a ranked list, a suggested next action. The person is between the output and any consequence.

In the decision. The model produces a recommendation and a person confirms or rejects it. This is the oversight position, and it is where Article 14 lives. It is also the position most vulnerable to quiet degradation, because the person can remain present in the process long after they have stopped functioning in it.

Instead of a person. The output takes effect with no human in the path. This is where the automated decision rules bite: Article 21 of the revised Swiss FADP and GDPR Article 22 turn on a decision based exclusively on automated processing with legal or significant effect, and they apply whether or not AI is involved.

Downstream of everything. The model summarises, monitors or reports on work that has already happened. This looks like the safest position and it carries a specific danger, which is the subject of the next lesson.

Why ranking is not as innocent as it looks

The upstream position is usually treated as low-risk, and usually is. The exception is ranking. A model that orders a list determines what gets attention, and in any process where the queue is longer than the day, what is not at the top is effectively not considered.

A system that ranks job applicants, claims for review or customers for outreach is shaping outcomes for people even though a human makes every decision that follows. Nobody experiences it as automated decision-making, because a person decided each case they saw. The cases they never saw were decided by the ranking.

The test to apply is not whether a human decides. It is whether the human ever sees the alternatives. If the model shows the top twenty of two hundred and nobody looks further, the model made one hundred and eighty decisions.

The person who is present but not deciding

The middle position has a well-known failure: the human confirms. Module 15 covered how to detect it and why throughput reveals it. The point here is structural rather than behavioural.

A process designed with a human in the decision can drift into being a process where the output takes effect automatically, without anyone changing the design or the documentation. The volume rises, the override rate falls towards zero, and the system moves from the second position to the third in everything but paperwork. At that point the data protection duties that attach to solely automated decisions become live questions, and the organisation is not asking them because its documents still describe a human decision.

This is worth a standing check: if the override rate is near zero and the volume is high, ask whether the decision is still meaningfully human, and what would follow if it is not.

Mapping a workflow before classifying it

Five questions, in order.

• What comes in, and does it contain personal data?
• What does the model output: text, a score, a ranking, a classification, an action?
• Who or what receives that output next, and is it a person?
• Can that recipient see what the model excluded, or only what it selected?
• What is the effect on a person at the end, and can it be reversed?

Answer those and the classification usually follows without argument. The fourth question is the one most workflow maps omit, and it is the one that catches ranking.

Chains

Increasingly the recipient of a model's output is another model: a summariser feeding a classifier, a retrieval step feeding a generator. Each link is individually defensible and the chain as a whole may have no point at which a person sees the intermediate result.

Two consequences. Errors compound rather than cancel, because the second model has no way to know the first one was wrong and treats its input as fact. And the natural place for oversight, which is where a person could inspect an intermediate output, may not exist anywhere in the chain. Where a chain exists, map each link separately and note where a person could look, not just where one does.

As of 24 September 2026.