Skip to content

03 · Drill 03 — Three Discovery Objections

Status: Outline. Body fills in Week 2. Voice: principal-level, BFSI-threaded, Apic-calibrated.

What this drill is. The three most common discovery-stage objections — cost, safety, lock-in — each handled in 90 sec at Strong-Hire grade. The objection responses you internalize for live calls.

What this drill is NOT. A full objection-handling library (that's Module 11 Note 05). Not pricing rebuttals.

The three objections

Objection 1 — Cost

"Frontier models are too expensive for our volume. We can't run this at production scale."

Objection 2 — Safety / hallucination

"Last year a vendor's model hallucinated a regulatory fact in a CISO demo. How do we know Claude won't?"

Objection 3 — Lock-in

"We don't want to be locked into one model provider. What's our path if we want to switch?"

Time box

  • Per objection: 90 sec response.
  • Total drill: ~5 min including reading prompts cold.
  • Cold variant: partner picks one of the three at random and adds a follow-up.

Rubric

Strong Hire

For each objection: - Acknowledge the legitimate concern in 1 sentence (not "great question," but "yes, this is the right concern to raise"). - Reframe with a specific architectural answer, not a sales rebuttal. - Anchor with a concrete pattern — model selection matrix, eval framework, multi-cloud deployment surface. - Close with a tradeoff acknowledgment — what you're not claiming. - 75–90 sec wall clock.

Hire

  • Acknowledge + reframe land, anchor is generic ("we have a model selection matrix") not specific.
  • Trade-off acknowledgment missing.

Lean No

  • Sales-rebuttal shape ("Claude is actually cheaper than you think").
  • Mission paraphrase ("Apic is committed to safety").
  • Defensiveness audible.

Strong No

  • Over-claim ("Claude doesn't hallucinate") — Strong No on factual grounds.
  • Dodging ("let's circle back") — Strong No.

The Strong-Hire moves per objection

Objection 1 — Cost

  • Acknowledge: "Frontier-model cost is the architectural concern most often skipped early."
  • Reframe: Model selection matrix — Opus for orchestration / Sonnet for the bulk / Haiku for high-volume classify. Plus prompt caching: long system prompts get cached, hit-rate optimization changes the cost shape.
  • Anchor: "In your support-agent assist case, the right shape is Sonnet 4.6 for the agent loop, Haiku 4.5 for the triage classifier, with shared cache for the policy context. That changes the cost-per-decision math by ~70% vs naive Opus."
  • Trade-off acknowledgment: "This isn't free — you'll need an eval framework that accepts model substitution. We'll design that with you in stage 2."
  • → Module 04: Model Selection Matrix, Prompt Caching Architecture.

Objection 2 — Safety / hallucination

  • Acknowledge: "Yes — this is the concern your CISO will raise too. Claude hallucinates. Every frontier model does."
  • Reframe: The architecture decides whether hallucination matters, not the model. Three controls: retrieval design (citation-grounded answers), refusal gates (when the model isn't confident), human-in-the-loop on customer-impacting actions. Plus eval-gated rollout (shadow → assist → autonomous).
  • Anchor: "For your compliance-doc-review case, we'd design the eval to score faithfulness against the source document, gate at 99%+ before going to assist mode, and never run it autonomous. Hallucination isn't 'made it go away'; it's 'made the failure mode caught and reversible.'"
  • Trade-off acknowledgment: "We're not claiming zero hallucination. We're claiming bounded blast radius."
  • → Module 06, Module 07, Module 09.

Objection 3 — Lock-in

  • Acknowledge: "Right concern. Ten years of cloud history says lock-in matters."
  • Reframe: Three layers of substitutability. (1) Deployment surface — Bedrock, Vertex, first-party. You can move between them with the same Claude. (2) Eval framework — model-agnostic. Same eval runs against any candidate model. (3) Tool schema and orchestration logic — written in your code, not Apic's.
  • Anchor: "In your architecture, we'd design the eval framework as a separate component that scores any candidate model against your acceptance threshold. If a non-Claude model crossed your threshold tomorrow, you could substitute it in under a sprint."
  • Trade-off acknowledgment: "Substitutability isn't free — you pay a small architecture tax for the abstraction. The math works at your scale; might not at smaller scale."
  • → Module 04: Deployment Surfaces. Module 06.

Common Lean No traps

Trap 1 — Sales rebuttal

"Claude is actually cheaper / safer / more open than you think." Reads as marketing. Lean No.

Trap 2 — Over-claim

"Claude doesn't hallucinate at the bar your CISO needs." Strong No on factual grounds.

Trap 3 — Dodging

"Let's set up a follow-up to discuss this." If the architect can't address it cold, the objection wins.

Trap 4 — Mission cosplay

"Apic is deeply committed to..." Strong No.

Trap 5 — Refusing the trade-off

"There are no trade-offs — this is just better." Strong No.

Trap 6 — Premature competitor compare

"Yes, GPT-4 has those issues, Claude is better." Lean No — not what was asked.

How to run this drill

  1. Have a partner (or your phone recorder) pick one objection at random.
  2. Speak the response cold, timer on, 90 sec.
  3. Listen back. Score against the 4-step Strong-Hire structure.
  4. For the criteria you missed, rewrite the response sentence.
  5. Re-run with a different objection.
  6. Application trigger: all 3 objections graded Strong Hire across 2 sessions.

Cross-references

Strong-Hire bar for this drill

  • 4-step structure (acknowledge / reframe / anchor / trade-off) deployed cleanly per objection.
  • Anchors are specific (use case + model + control), not generic.
  • Trade-off explicitly named, never hidden.
  • All 3 objections at Strong Hire across 2 sessions.