Run It With the Agent: Pricing an Insurer’s AI Sovereignty in Six Steps

Companion · Course 34 · Run It With the Agent

A companion to Course 34. It runs the course’s method with an agent on one invented European insurer, from a list of 12 AI workloads to a page of meters a board can read. Every figure is illustrative.

A 7-minute worked example in six steps. The insurer is a composite invented for teaching, and every figure in it is illustrative.

The Short Answer

An agent can classify, trace and price an estate quickly, from the estate’s own records. It cannot decide what a loss is worth, how likely it is or which seller to buy from. The video runs the course’s method end to end on one invented insurer and marks, at every step, where a person has to decide.

The insurer is a midsize European firm. It spends 10 million euros a year on 12 AI workloads, all in one public cloud region, and it treats all 12 alike. You play the architect, and the question at every step is who else can read the workloads, stop them or change them.

1. Classify the Estate

You give the agent the list of 12 workloads with what each costs and where it runs. For each one it asks which verb would hurt, who could exercise it and what the loss would be. The key idea on screen is that until the verb is named, no protection can be priced.

Four workloads have no named verb: marketing copy, the public website assistant, staff handbook search and translation of published documents. They stay on rung one and carry 4 million euros of spend.

Three fear a read by the provider’s support staff and need rung two, at 2 million euros. Three fear a read by anyone who obtains data from the provider and need rung three, the keys, at 2.5 million euros: health claims extraction, fraud scoring and the underwriting assistant.

Two critical workloads, claims triage and the solvency reporting agent, fear a stop and a change. They need rungs four and five, at 1.5 million euros.

The agent refuses a vague answer. Sensitive is not a verb, and the cloud is not a party. The people who would bear each loss estimate it, and each row gets an owner and a date.

2. Trace the Request

Next the agent takes one real claims triage request and follows it through seven hops: the prompt path, the retrieved context, the tool calls, the model, the logs, the traces and the memory. At each hop it records the place, the operator, who holds the key and who owns the operator in law.

The model call sat where the plan said it would. The traces and the evaluation store did not. A workload sits at the rung of its lowest hop, so the placed rung for claims triage is rung one, and the video says it is set by the plumbing, not by the model.

The rung gap is the required rung minus the placed rung. Across the estate, eight workloads sit below the rung they need, and four sit exactly where they should, on rung one. A gap above zero is a risk nobody chose. Which gaps to close first is the person’s call.

3. Read the Offer

A seller now offers a sovereign region. The agent does not ask whether the offer is sovereign. It reads the seller’s own documents and, for each verb, states the highest rung the design really reaches. The seller is a large cloud provider with a local operating entity.

For read, the offer reaches rung two, the operator. It reaches rung three only if the firm holds the keys in an external key manager, because a key the provider stores does not count. For stop, it does not reach rung four, because the parent company can still be compelled by a foreign court. For change, it reaches nothing, because the model roster stays with the model maker.

So the offer fits six workloads and does not fit the two critical ones. Where a document is missing, the agent writes not evidenced and drafts the question for the seller. It gives no legal opinion: that question goes to counsel, and the choice of seller is the person’s.

4. Price the Premium

The agent then prices the estate with the course’s illustrative rates: 10% for the operator, 15% for keys, 25% for law and 40% for law with a model the firm runs. Putting all 12 workloads at rung four would cost 2.5 million euros a year.

Classified, the bill is different. Nothing is spent on the 4 million euros at rung one. The rung two group costs 200,000 euros, the rung three group 375,000 euros and the two critical workloads 600,000 euros. The total is 1,175,000 euros, which is 11.75% of spend and saves 1,325,000 euros a year against the blanket. The two critical workloads sit one rung higher than the blanket put them.

Then the break-even for claims triage. The premium is 400,000 euros a year, and a stop costs 400,000 euros a day, so a 20-day stop costs 8 million euros. The break-even is 5% a year. If the risk committee believes the chance is one in 100, the expected loss is only 80,000 euros a year, and the purchase then rests on the regulator or on whether the firm could survive the loss.

The agent does the arithmetic. The committee decides.

5. Specify and Drill

Each classified workload gets one page, and the agent fills it in. For claims triage the verb is stop, then change, and the parties are a foreign government and the model vendor. The required and placed rung are rungs four and five, so the gap is zero.

Keys sit in an external key manager on hardware the firm holds. The model is a hosted one, reached through a gateway the firm owns. The fallback is an open-weight model whose weights the firm holds, and it passes the frozen suite. The page is reviewed every quarter.

Then the exit drill. The agent writes the exit ledger for one rung three workload and names a second operator. A second region of the same provider does not count. The first drill took nine weeks, and three things broke: identity roles had no equivalent, prompts had to be retuned, and the evaluation set lived in one engineer’s notebook. By the third drill the move took nine days.

A move that works but answers worse is a failure, and a person signs every item accepted as it is. The key idea on screen is that an exit never run is an estimate.

6. Meter and Diagnose

Finally the agent reads the estate as it stood before the work against the course’s ten failure patterns. The traces had left the plan, which is the telemetry trail. Almost every key lived with the provider, which is the resident key. No exit had ever been run, which is the paper exit. And the plan on the table was the blanket rung.

The root was the same each time: nobody had named the verb. After one illustrative year, coverage went from 0 to 100% and the rung gap from 8 to 0. Key custody rose from 10 to 85%. Exposure to foreign compulsion fell from 100 to 60%, and the video labels the remaining 60% as the 6 million euros at rungs one and two.

Both critical workloads hold a model the firm can run. The last drill is 60 days old, and the premium ratio is 11.75%.

Who Did What

The video ends by counting. The agent classified, traced, priced and drafted. The people named the loss, judged the chance, chose the seller and accepted what was left. A practice kit for the course holds every template shown, blank and ready for your own estate.

The method itself is taught in Course 34: Sovereign Engineering.

Try It on Your Own Estate

1. Step 1: for each of your workloads, which verb would hurt, who could exercise it, and what is the loss? Is the answer a verb and a party, or a vague word?
2. Step 2: for your most critical workload, at which hop does the placed rung drop below the required one?
3. Step 3: for one offer, which rung does the seller’s own documentation reach for read, stop and change? What is marked not evidenced?
4. Step 4: what is the break-even probability for your most critical workload, and does the committee judge the yearly chance above or below it?
5. Step 5: when was your last exit drill, and who signed each item accepted as it is?