Spend Posture Controls: Governing Organizational AI Spend
CoreConfigure Claude tools and environments for teams · Difficulty 2/5
Explanation
settings.json (7.1.3's control surface) doesn't only govern which tools an agent may call -- it's also where an architect enforces organizational SPEND POSTURE: the deliberate set of controls that decide how much a team can spend, on which models, and how much reasoning depth they default to. Left unconfigured, spend posture defaults to "whatever any individual developer's session happens to invoke," which is exactly the same per-developer drift problem 7.1.2 and 7.1.4 already named for memory and integrations -- just applied to cost instead of context or capability.
The Four Levers
| Lever | What it controls | Example |
|---|---|---|
| Model defaults | What model a session uses unless explicitly overridden | A team defaults to a mid-tier model for routine work rather than the most capable (and most expensive) tier by default |
| Model allowlists / restrictions | Which models a team is permitted to invoke AT ALL | Blocking a cost-sensitive team from ever invoking the most expensive tier, regardless of what any individual session requests |
| Effort / reasoning-depth guidance | Default extended-thinking usage | Steering routine tasks away from maximal reasoning depth by default, reserving it for sessions that are explicitly flagged as needing it |
| Spend / rate / per-user caps | Hard ceilings on total spend, requests-per-minute, or per-seat consumption | A monthly spend ceiling per team, or a per-user request-rate limit that prevents one runaway script from consuming a shared budget |
Why This Belongs in settings.json, Not CLAUDE.md
This is the same enforced-vs-advisory distinction 7.1.3 already established for permissions, applied to cost. A sentence in CLAUDE.md -- "please use the cheaper model for routine tasks" -- is advisory context Claude tries to follow; it offers no guarantee, and a developer (or the model, reasoning about an unusual request) can simply not follow it. A model allowlist or a hard spend cap enforced in settings.json holds regardless of what any session decides, the same way a PreToolUse hook holds regardless of what the model "decided" about a destructive action. If a spend rule needs to be a guarantee rather than a preference, it belongs in settings.json for exactly the reason 7.1.3 already gave for permissions -- not because spend governance is a special new category of concern, but because it's the identical deterministic-vs-probabilistic split showing up on a different axis.
Putting It Together
An architect configuring a new team's Claude Code deployment typically sets model defaults appropriate to that team's typical workload, restricts the most expensive tier to teams or use cases that actually justify it, nudges reasoning-depth defaults toward efficiency for routine work while leaving room to escalate deliberately for genuinely hard problems, and sets a spend or rate cap as a backstop against runaway usage -- all four configured once, centrally, in settings.json, exactly like the permission rules and MCP servers already living there.
Common exam traps
- Writing a spend guideline into CLAUDE.md instead of settings.json. "Please default to the cheaper model" as prose in CLAUDE.md is advisory and unenforced -- the same misconfiguration 7.1.3 names for permissions, just applied to cost. The correct fix is a model default or allowlist rule in settings.json.
- Assuming a single global spend cap is sufficient. A single org-wide ceiling doesn't stop one team or one runaway script from consuming the entire budget before anyone else gets a share; per-team or per-user caps are what actually distribute a spend ceiling fairly across an organization.
Key Takeaways
- Spend posture is governed by four levers: model defaults, model allowlists/restrictions, effort/reasoning-depth guidance, and spend/rate/per-user caps
- All four levers are enforced through settings.json -- the same deterministic control surface as tool permissions and hooks, not CLAUDE.md
- A spend guideline written as prose in CLAUDE.md is advisory and unenforced, exactly the same misconfiguration 7.1.3 names for permission rules
- Model allowlists block a team from invoking a cost tier at all, regardless of what any individual session requests
- Per-team or per-user caps, not just a single global spend ceiling, are needed to distribute a budget fairly and prevent one runaway consumer from exhausting it
Glossary Terms
Related Concepts
CLAUDE.md vs settings.json for Team Configuration
CLAUDE.md is memory/instructions; settings.json is deterministic configuration (permissions, hooks, MCP)
Cost, Latency & Token Optimization Techniques
Prompt caching cuts cost/latency on stable repeated prefixes with no accuracy loss -- order stable content first