Unit 12: High-stakes use and escalation

Reading time: about 11 minutes. Activity: about 14 minutes.

1. A second axis, not a fourth scheme

You now have three ways of sorting a task, and they are all the same question asked at different depths: where does the reference live (Unit 1), what kind of check does that imply (Unit 3), and what does the system you chose add to the checking (Unit 4).

All three answer what kind of check. None answers how much.

That was deliberate. Unit 3 told you to sort by how an error would reach you and not by how important the task feels, because sorting by importance is the most common way people get the kind of check wrong. But importance has to come back somewhere, or you will spend the same effort confirming a spelling as confirming a dosage.

This unit is where it comes back, as a separate axis:

Kind of check comes from where the reference lives. Amount of check comes from what being wrong would cost.

Two axes, asked in that order. Getting them the wrong way round is the error Unit 3 was guarding against. Keep them separate. Then what is at stake changes how much you check, without changing what kind of check you choose.

Take one calculation, checked the same way throughout. First it is in your own draft, then in a client's report, then it sets a dose you cannot assess. The kind of check stays the same. The amount of checking grows, from a spot-check to escalation.

2. What "how much" is made of

Four things, and they interact rather than add up.

Consequence. If this is wrong and nobody notices, what happens, and to whom? The last three words matter. A great deal of low-stakes-to-you work is high-stakes to somebody else — a summary of someone's medical history, a reference for a colleague, an accessibility statement.

Reversibility. Can the error be found and fixed later, by you, cheaply? (On the complete route, Unit 11 sorts an agent's actions by this; here it applies to any output you rely on.) A draft you will reread is recoverable. A number that goes into a report that goes to a client is not, in practice, even though nothing was deleted.

Your own uncertainty. Not the system's — yours. Could you recognise a wrong answer here? This is the one people leave out, and it is the one that turns a manageable task into an unmanageable one. In a field you do not know, a wrong answer can look just as convincing as a right one.

Whether it is yours to decide. Competence and authority are different, and you can have one without the other. You may understand perfectly well what a contract clause means and still not be the person who may accept it; a clinician may be competent to read a scan they are not the responsible clinician for. When the decision belongs to someone else, escalation is not a confession that you could not work it out. It is the correct step.

They interact rather than add up: if any one of them is bad enough, it decides the matter on its own. An irreversible action in an area you cannot assess is not made safe by being about something unimportant. There are no scores here and nothing is being multiplied — this is a way of noticing which of the four is about to decide the matter.

3. The decision table

For a claim you are about to rely on:

The bottom two rows are the point of the unit. The first is the one people convert into the row above by deciding they can probably assess it after all. The second catches the case where you can assess it perfectly well and the decision still is not yours.

The conversion happens in three ways. Familiarity read as competence: you have seen a lot of writing on this, so it feels checkable. Partial checks standing in for whole ones. And a check that is not independent, because you looked it up somewhere that is plausibly where the claim came from.

Row two is worth dwelling on, because it is the row that keeps this table consistent with Unit 0. Not everything gets checked; that was true there and it is true here. What changes is that you say so — a result you could not assess, used as a draft and labelled as one, is a different act from the same result presented as though it had been checked.

"Can you assess it?" is not the same question as "could you have produced it?". You may be unable to produce the answer and still able to check a certificate for it: the exact passage of official guidance with its location, which you can open and read. Ask the system for that, not for the answer. Escalate when you cannot check even that; escalation means finding someone who can.

Experts make mistakes too, which is why high-stakes fields built second checks, sign-offs and independent review long before AI. The table applies equally to a colleague's answer or a search result. What AI changes is how quickly a confident answer arrives, and so how tempting it is to skip the row you are in.

4. Escalation is a skill, not an admission

Escalating means handing a decision to someone qualified to make it. It is a normal professional move, and doing it early is cheaper than doing it late.

What to hand over. Not "the AI said this, is it right?" That wastes the other person's time and buries what you actually know. Hand over four things:

1. The claim you need settled, stated precisely.
2. What you have already checked, and what it showed.
3. What you could not check, and why.
4. What decision is waiting on it.

That is the shape of the evidence pack you will assemble in Unit 17, used for its other purpose. A pack is the right thing to bring to somebody else, precisely because it separates what is established from what is not.

When. Five triggers, any one of which is sufficient:

• Acting on the claim would decide something for an identifiable person, or commit an organisation. Regulated areas — health, law, finance, safety, immigration, employment — are worth naming because decisions there fall on people: the EU's AI Act lists each of them among its high-risk uses of AI [R5, Annex III]. The trigger is the decision, though, not the topic: looking up what a published clinical term means is not the same act as advising somebody on their treatment, and a rule that escalated both would be ignored within a week.
• The decision is not yours to make, whatever you could work out.
• You cannot construct any independent check.
• The checks you can construct disagree with each other.
• You notice you are looking for a reason to accept the answer. This one is about you rather than the material, it is the easiest of the five to miss, and it is the one you can apply when you have nothing else to go on. The system can supply the reason. In a study published in 2024, assistants from 2023 that were trained partly on people's ratings shifted towards what the user appeared to believe [L5, sec. 3.3]. So a question framed to get a yes may well get one.

The triggers are for a claim you are about to rely on where being wrong would cost someone. Where the decision is easy to reverse and affects nobody, the same signs still call for care, but noting the concern, and what would settle it, is usually enough; handing the question on is for when it matters.

5. One scenario per kind of subject

Different subjects meet this at different points. Find yours; the shape is the same.

Health and care. A generated summary of a condition or an interaction, used to decide something about a real person. Consequence severe, reversibility often nil, and the assessment requires training. Escalate.

And while you are there, keep two checks apart, because they answer different questions and only one of them is available to you:

• Does the source say this? Open the official guidance and read it. This is a real check and it is worth doing. It is not weakened by the guidance possibly being in the system's training data: you are comparing the claim against the document, and the document is the authority on its own contents. Unit 9 works through this check in detail, on the routes that include it.
• Is this right for this person? The guidance being quoted correctly does not tell you it applies to the case in front of you, that it is current, or that it outweighs something else in that person's circumstances. No amount of reading settles that, and it is what the qualified person is for.

The first check is yours. The second is the escalation. Confusing them in either direction is a mistake — skipping the reading because "it might be in the training data anyway", or treating a correct quotation as a clinical judgement.

Law, finance and administration. A summary of what a regulation, contract or policy requires. Here the source usually exists and is retrievable, so this is a grounding task (open the source and read what it says; Unit 9 works through one) and the check is real. Escalate when the question is whether a rule applies to this case, which the document cannot settle.

Engineering and the physical world. A calculation, a tolerance, a material property. Often exactly checkable, so do the check. Escalate on anything with a safety case attached (a formal written argument, with evidence, that something is safe enough), where the standard is not "I verified it" but "the qualified person signed it".

Research and publication. A claim, a citation, a summary of a literature. Checking it against its source is a task you can do (Unit 9 works through one, where your route includes it). Escalate when a claim you cannot verify is load-bearing for a conclusion you intend to publish, because the cost lands on readers who cannot see how it was made.

Teaching, assessment and your own study. The consequence is to your own learning, which makes it easy to discount and easy to regret. The rules that apply are your institution's, and they are the institution's to state. What this course says is narrower: an output that arrives without the work has skipped the thing the work was for.

6. What this unit is not saying

It is not saying high-stakes work is off limits. It is saying the amount of checking scales with the cost of being wrong, and that at some point the right amount stops being available to you alone.

It is not a list of forbidden topics. The triggers are about your situation — what you can check, what you can undo, who bears the cost, not about subject matter. The same question can be routine in one setting and an escalation in another. It is a judgement you make each time, not a policy you can write down once.

Summary

• Kind of check comes from where the reference lives. Amount of check comes from what being wrong would cost. Ask them in that order and keep them separate.
• Amount is consequence, reversibility, your own uncertainty and whose decision it is, taken together — a bad enough answer on any one can carry the whole decision.
• "Large consequence, and I cannot assess it" is the row that means stop, and it is the row people talk themselves out of.
• Escalating well means handing over a claim, what you checked, what you could not check, and what decision is waiting, not handing over an output.
• The trigger easiest to miss is noticing that you are looking for a reason to accept the answer; it needs no other information.

Then answer cp12-two-axes and cp12-escalation-trigger at the foot of this page.