The role, its neighbours and the interview · The interview

The customer scenario — the CIO wants it on-premises

Decide what you say when a role-played CIO demands on-premises servers in week 2, and which placement the pilot uses for its last six weeks.

casemediumExercise · 3 min · 63 of 64

On this page

An interviewer plays a customer with a firm demand and a reason they have not said yet. The round, as the typical loop describes it, is scored on whether you find the requirement under the demand and keep the customer while changing the answer.

The situation

The interviewer plays the CIO of the regional insurer in week 2 of the 8-week claims-triage pilot: "Claims data does not leave this building. Put the model on our servers." The interviewer will confirm any of the following if asked. The claims system already runs in the insurer's own cloud tenant, in-country, under a cloud agreement the CIO signed last year, and the pilot service and its data sit inside that tenant. The prototype calls a hosted model API, so today each prompt (claim form, adjuster note, policy section) leaves the tenant. The same model is offered as a listing inside their cloud tenant at the same r, about a week to enable. The pilot runs 1,200 claims a day at a mean of 8 seconds of model time each, 60% of them between 09:00 and 12:00. Nobody on the CIO's staff has run a model server, and the open-weights model the servers would run has not been measured on the 200-case gold set.

The question

Decide what you say to the CIO in the next five minutes, and which placement you propose for the remaining six weeks.

A strong answer · Rubric · Where answers go wrong

Write your answer first, then open

A strong answer

Ask before answering: "Is the requirement that claims data stays in this building, in your tenant, or in the country?" The claims system already runs in their cloud tenant, so the data left the building last year under an agreement she signed. The requirement she is likely to confirm is that the prompt never leaves the tenant, and the prototype breaks it today; say so, because she is right about the problem.

Then the three placements, in her terms. The hosted API fails the requirement. The listing inside her tenant meets it with a document she already holds, at the same r, and a week to enable fits inside six. Her own servers meet it too, at a cost she should see in numbers: 1,200 claims × 8 seconds ≈ 9,600 seconds, about 2.7 hours of serving a day, on hardware sized for 720 claims between 09:00 and 12:00, so busy about 11% of the day. They also need a named person on her staff to carry the pager, and they run a model nobody has measured on the 200 cases.

Propose the listing from week 3, so the prompt stops leaving the tenant within a week. On-premises becomes a separate phase after week 8, with its own gate of 95% on the same 200 cases and a named on-call before any hardware is bought. Do not agree to the servers in the room, and do not argue latency, which she did not raise.

What reverses this: a regulator's letter or contract clause that names the building. Then the pilot moves to her servers and the date moves, which is the sponsor's decision.

Rubric

asks whether the requirement is the building, the tenant or the country before answering the demand
uses the claims system's existing tenant to restate the requirement in a form the CIO can confirm
concedes that the prototype's hosted call breaks the requirement today
does the utilisation arithmetic and names the on-call and the unmeasured model as costs of on-premises
proposes the in-tenant listing for the pilot and on-premises as a gated later phase, without agreeing in the room
names the evidence that would make the building the real requirement, and who then decides the date

Where answers go wrong

"On-premises buys nothing: the round trip is 200 ms on a call of several seconds." Loses because it is correct and irrelevant. She raised residency, and arguing latency tells her you did not listen; where the model sits calls this the mistake of treating placement as a latency decision.

"Agreed, we will move to your servers." Loses because it promises a quality figure nobody has measured and an on-call nobody has named, and the week-8 date dies in week 5 when the hardware arrives without an owner.

"The vendor's terms say inputs are not retained, so there is no residency issue." Loses because it answers retention, not residency: the prompt still leaves the tenant, which is the one fact she is right about.