TLDR
Every architectural choice - pattern, model tier, guardrail overhead - has to trace back to a named business-value pillar an executive sponsor recognizes: efficiency, transformation, productivity, cost, or performance SLAs. Anthropic: Building Effective Agents Anthropic's enterprise guidance frames system sophistication as NOT the goal - measurable outcomes are. The exam trap is a technically elegant architecture with no stated pillar justification.
What it is
The exam names five value pillars an architecture decision should map to: efficiency (cutting time/cost on existing work), transformation (a genuinely new capability, not just a faster old process), productivity (more work handled by the same team), cost (spend versus the baseline it replaces), and performance SLAs (a named latency/accuracy target the business depends on). This exact five-pillar wording doesn't appear verbatim in a single published Anthropic taxonomy - treat it as the exam's own synthesis of standard enterprise-value categories, grounded in real Anthropic guidance on outcome-anchored, trusted AI adoption rather than a direct quote.
The architect's job is to name which pillar (or pillars) each design decision serves, not to assume that a better-performing system is automatically better-value. Anthropic's guidance ties cost discipline directly to complexity: "Start with the simplest solution possible, and only increase complexity when needed" - unjustified architectural complexity is itself a cost-pillar violation, not merely a technical inefficiency.
How it works
Efficiency and productivity are related but distinct: efficiency asks whether an existing unit of work now costs less time or money, while productivity asks whether the team can now handle more volume with the same headcount. A workflow that replaces a manual triage step is an efficiency win; a batch pipeline that lets one team clear ten times the document volume is a productivity win - a single architecture can claim both if it earns them.
Cost is the pillar most often left implicit. Model tier, workflow-vs-agent choice, and guardrail overhead are all cost decisions in disguise - picking Opus "for quality" without stating which pillar justifies the added spend is the textbook exam trap. The fix is a stated cost ceiling per unit of work, compared explicitly against the labor or system cost it replaces.
Performance SLAs connect the architecture's technical configuration (model tier, retry/guardrail depth, context size) directly to a number the business depends on - e.g. "95% of triage decisions returned within 30 seconds." An architecture with no stated SLA has left this pillar unaddressed, which the exam treats as a gap, not a neutral omission.
Transformation is the pillar most often confused with efficiency. Doing the same manual process faster is only efficiency; surfacing a capability that did not exist before - a new fraud-pattern signal, a previously invisible exception trend - is transformation. An architecture that just automates an existing script one-for-one should not claim transformation.
Trust as a precondition, not a sixth pillar. Anthropic's enterprise guidance on "Building trusted AI in the enterprise" frames trust (safety, reliability, security) as necessary for realizing value at scale: an architecture that is fast and cheap but not trusted by stakeholders won't achieve adoption-driven value regardless of its technical merits. Naming a pillar without addressing trust is an incomplete justification.
Where you'll see it
Loan-document exception triage
Efficiency (less manual triage time), cost (per-document spend under the labor cost it replaces), SLA (95% of decisions in under 30 seconds), and transformation if it surfaces exception patterns no human previously tracked.
Support ticket deflection
Efficiency (faster resolution) and cost (cheaper than added headcount) and SLA (response-time target) - but not transformation if it's just automating an existing decision script one-for-one.
Side-by-side
| Pillar | What it measures | Example architecture decision | Anti-pattern |
|---|---|---|---|
| Efficiency | Time/cost per unit of existing work | A routing workflow instead of an open-ended agent for a fixed-path task | Adding agentic autonomy where the task doesn't need it |
| Productivity | Volume of work the team can now handle | A batch pipeline for high-volume classification | Leaving manual review in the loop for high-volume, low-risk decisions |
| Cost | Spend versus the baseline it replaces | A Haiku-tier model for straightforward extraction, with a stated cost ceiling | Defaulting to the largest model "for quality" with no stated ceiling |
| Performance SLA | A named latency/accuracy target the business depends on | "95% of triage decisions returned within 30 seconds" | No stated SLA - "fast enough" left undefined |
| Transformation | A capability that did not exist before | Surfacing an exception pattern no human previously tracked | Rebuilding an existing manual process one-for-one with no new capability |
Question patterns
A proposed loan-triage architecture has no stated cost ceiling and no stated latency target, but claims high accuracy. What's the gap?
"the architecture is fine as long as accuracy is high" - accuracy alone doesn't satisfy the cost or SLA pillars the sponsor actually cares about.An architect justifies choosing Opus over Haiku by saying "it's simply the better model." What's missing from this justification?
"model choice doesn't need business justification, only technical accuracy" - technical accuracy is not itself a value pillar the exam recognizes.A task has a fully known, fixed sequence of steps, but the team proposes an agent anyway "to be safe." Which pillar does this violate?
"agents are always the safer architectural choice" - safety here is not a stated value pillar, and unjustified complexity is itself a cost-pillar violation per Anthropic's own simplicity guidance.A new system is fast and cheap, but stakeholders don't trust its outputs and refuse to adopt it org-wide. Is this a value-alignment failure?
"the value pillars don't include trust, so this isn't a design failure" - trust gates adoption, so a distrusted system realizes none of its claimed efficiency or cost value in practice.An architecture cuts document-processing time in half AND surfaces a fraud pattern no human analyst had previously tracked. Which pillar(s) does it map to?
"productivity, since faster processing means the team gets more done" - that conflates the two: proving a per-document time cut is efficiency evidence; claiming productivity needs throughput/volume-handled evidence (the same team now clearing more total documents with the freed-up time), which this scenario does not state. If that volume evidence were added, productivity could be claimed too - but time-cut-in-half alone is an efficiency claim.Frequently asked
Is this a fixed 5-pillar list Anthropic publishes verbatim?
Can one design decision serve multiple pillars at once?
Work this with your AI
Work this concept hands-on with Claude Code, Codex, or claude.ai. Copy a prompt, paste it into your assistant, and practise in tandem. Each one keeps you active (explain it back, get drilled, or build) rather than just reading.
- Drill it like the exam (scenario MCQs)Practice in the exam's scenario-MCQ format with trap awareness.
- Explain it back (Feynman)Build durable, transferable understanding of a concept you can half-state.
- Test me, adapting the difficultyActive recall practice on a concept you think you know.
- Check my prerequisites firstBefore studying a concept that keeps not sticking.
- Find the high-leverage 20%When a domain feels too big and you are short on time.