Client simulation — pushback, bad news, and unblock

Expected question

"Role-play: the customer's CTO is on the line. The deployment slipped three weeks / they want an unsafe feature / security won't give credentials. Handle it."

Variant forms

"Tell the CTO the go-live slipped by three weeks."
"Customer wants a feature that compromises data governance. Push back without losing the deal."
"Explain to a non-technical VP why RAG cannot guarantee 100% accuracy."
"Customer's IT won't give production credentials. How do you unblock?"
"You disagree with the customer's chosen architecture. Raise it."
"The model is 'getting worse' according to the customer. Investigate live."
"A dependency API times out intermittently in their VPC. Debug aloud."
"Ship Friday or slip — what do you compromise, what don't you?"
"Describe your worst production incident at a customer — what did you tell them, and when?"
"Tell me about a time you said no to a customer and they thanked you later."
"Tell me about a project that failed — whose fault was it?"
"A customer wants the wrong feature. Redirect without losing trust."
"Requirements kept changing — how did you still own the outcome?"

Where this actually gets asked

Anthropic Applied AI and OpenAI FDE client simulations; Palantir-style stakeholder pressure; hiring-manager behavioral probes on conflict and bad news. Grades composure + options + principle, not eloquence.

The question, as it might actually be asked

"I'm the frustrated customer. Convince me you still own the outcome."

The framework

30-second thesis

I'd acknowledge the hit first, put facts on the table in one breath, give two or three options with a clear recommendation, protect HITL on irreversible harm, and own the next checkpoint. I don't surprise people on Friday, and I don't lecture without a path to value.

2-minute method

Acknowledge → facts → options → recommendation → ownership → next checkpoint.

This is a conversation, not a status email. How it sounds:

Scenario	What I'd say	What I'd refuse
Slip / bad news	“You're right to be unhappy — we missed the date. Here's what broke, what's already fixed, and three ways forward. I recommend B. I'll own daily updates until green.”	Surprise Friday; blame-only; “still investigating” with no date
Governance pushback	“I hear you want speed. The risk I'm protecting is a wrong write we can't unwind. Here's a safe alternate that still moves the metric: scoped tool + HITL + shadow, then auto after the reject-rate gate.”	Lecture; silent reject; “policy says no” with no path
Accuracy expectation	“We won't promise 100%. We'll define eval slices, decline when unsupported, and keep a human on the high-harm path.”	Fake certainty
No prod credentials	“Let's unlock staging parity, least-privilege break-glass with an expiry, and dual-run read-only on masked data — without copying prod home.”	Demand god-mode; stall forever
Architecture disagreement	“If the goal is X, here's the failure mode of Y. Want to dual-run A vs B for two weeks?”	Ego fight; “trust me, I'm senior”

CTO slip — spoken under ~90s:

You're right to be unhappy — we missed the date we committed. Root cause in one line: [dependency / scope / access]. We've already [fix]. Three options: (A) reduced scope this Friday with the same safety invariants, (B) full scope in N days, (C) pause until dependency X clears. I recommend B because forcing A still leaves [risk]. I own the daily update until we're green; the metric I'll watch is [tool success / HITL reject / groundedness slice].

Requirements (what “pass” looks like in the room)

Functional

Restate customer goal in their words before arguing.
At least two viable options plus a recommendation.
Explicit next checkpoint with owner and date.

Non-functional

Never trade away authorization / HITL on irreversible harm for a date.
Facts over vibes; label uncertainty.
Leave a reusable operating rule, not just an apology.

Core entities / actors

Frustrated sponsor / CTO: cares about date, risk, optics.
Security / IT: owns credentials and change windows.
FDE: owns communication cadence and technical unblock plan.
Safe alternate: HITL path, reduced scope, shadow, staging proof.

Process flow — live simulation

Rendering architecture diagram…

Deep dive 1: governance pushback without losing the deal

Customer: “Just turn on auto-write. We're losing time.”

Me: “I get the urgency. What I'm not willing to do is let the agent post an EDI change or a payment-adjacent write without a human until we've seen reject rates stabilize. Alternate that still moves the needle: scoped tool, HITL for two weeks, shadow beside your operators, then remove the gate when the eval says so. That's engineering — not bureaucracy.” Point to concrete controls (O: AegisAI HITL / decline-on-low-confidence RAG) as the alternate.

Deep dive 2: “model getting worse” live investigate

Don't retrain first. Talk it out:

“When you say worse — which intent, tenant, doc version, prompt version?”
Check recent deploys: prompt, index, model, retrieval filters, tool schemas.
Compare golden regression vs online sampled fails.
Shadow previous bundle; roll back if a gate is broken.
Retrain is last, not first.

Deep dive 3: no prod credentials unblock

“I'm not asking for god-mode. Staging with contract tests, time-boxed least-privilege break-glass, dual-run read-only against masked prod, never exfiltrate. Go-live sits on a security-owned window — not FDE heroics.”

Quantitative trade-offs

Decision	Trade-off and reversal evidence	Evidence class
Slip with full scope vs ship reduced Friday	Reduced scope preserves trust on date; reverse when reduced path still hits irreversible risk without HITL	H
HITL hold vs full autonomy to save the deal	Autonomy may close a meeting; reverse only after eval + reject-rate gates and sponsor accepts residual risk in writing	O/P
Daily update cadence vs weekly	Daily costs time; reverse when green for 2 weeks and sponsor prefers weekly	H

Migration / operating rule after the call

Write one rule the account team reuses — e.g. “No irreversible tool without HITL until reject rate < X for Y days” (H until you have measured thresholds) or “Bad news < 24h with three options.” Put it in the account runbook.

Org ownership

FDE owns customer communication until the checkpoint is green.
Security owns credential exceptions with expiry.
Sponsor owns scope trade decisions among the options you offered.
Product absorbs recurring pushback themes as platform defaults.

Situation

Client simulations mirror Lucid stakeholder rooms (P): Commerce / Supply Chain leaders care about dates and risk, not model names. Parallel interview pressure: CTO on the line after a slip, or a demand for fully autonomous writes that bypass governance.

Task

Keep trust while protecting the customer from irreversible harm — deliver options and a path to value, not a lecture or a silent stall.

Action

Open with acknowledgment of impact (date missed / risk felt).
State root cause in one sentence and what is already fixed.
Offer 2–3 options with trade-offs; recommend one aloud.
Refuse to remove HITL/authorization on irreversible paths; offer a safe alternate.
Own daily (or agreed) updates with a watched metric until green.
Convert the incident into a reusable account operating rule.

Result

Composure + options + principle. Relationship intact; unsafe path blocked; unblock plan concrete. The “no” lands as engineering — gateway HITL, decline paths (O) — not policy theater.

The follow-up question you should expect

"Ship Friday or slip — what do you compromise?"
Polish, breadth, and autonomy. Never access control, audit, or HITL on irreversible actions. I'd rather ship a reduced-scope Friday with those invariants than a full-scope lie.

What I'd ask them (in the role-play)

“What does success look like Friday if we cut scope — which workflow still has to work?”
“Who can approve the HITL reviewer queue this week?”
“If we choose B, what's the first thing that would make you lose trust again?”

Candidate-owned evidence prompts

Which real pushback story will you use (HITL on EDI/payments-adjacent paths)?
What metric will you put on the daily update?
What is your go-to “safe alternate” sentence?
Have you practiced the CTO slip script out loud under 90 seconds?

Author reference (do not memorize)

Scripts are scaffolds. Do not invent external customer incident logos. Lucid = internal customer (P); open gateway/RAG = method (O).

Staff+/Principal signal rubric

Mid-level: Apologizes or digs in; no options.
Senior: Clear status and a next step.
Staff+: Options with trade-offs; protects governance; owns the checkpoint cadence.
Principal: Turns the incident into a reusable operating rule the account team can reuse.