Skip to content
Back/Harness Engineering

AI Cost Governance — Parameters, Not Policies

View in Graph
Updated 2026-10-26
3 min read
707 words

AI Cost Governance — Parameters, Not Policies

What it is

AI Cost Governance is the operational practice of managing model spend through visible, adjustable parameters instead of fixed annual or quarterly budgets. The core insight: model capabilities, prices, and team workflows change faster than static policies can remain valid. The replacement is a feedback loop made of circuit breakers, real-time usage visibility, earned friction, and evidence-based adjustments.

Every's operations lead Arielle Shipper documented this approach after a single GPT-5.6 Sol test raised the company's daily OpenAI credits from 11,520 to 26,685 — about 230% above baseline — and drained the account before breakfast. The response was not a permanent cap; it was a set of operating parameters that let the team keep exploring while keeping spend explainable.

Why it matters

Fixed token budgets fail in two directions. They either block high-value experiments, or they are blown through by a single unbounded task. The real lever is not the dollar limit but the clarity of the task boundary: what the model is allowed to do, when it must stop, and what the team learns from an expensive run.

When costs are visible in real time and tied to outcomes, the same dollar amount can be judged very differently. A $480 run that builds infrastructure, teaches a cheaper workflow, and produces publishable insight is not the same as a $480 run that checks email every fifteen minutes without proportional value.

Key points

  • Set circuit breakers, not ceilings. Use card limits and Slack alerts to pause spend when a threshold is hit, then require a human decision to continue.
  • Make usage visible to the whole team. When everyone can see daily credits, top consumers, and run duration, spend becomes a conversation rather than a month-end surprise.
  • Let friction be earned. Start with light guardrails and add heavier controls only after evidence shows they are needed.
  • Adjust with evidence. If a team can explain what it spent, what it delivered, and what it learned, autonomy stays. If usage is unexplained or unshared, platform-level restrictions follow.
  • Judge runs by outcome and compounding value. An expensive run that leaves reusable skills, tests, or decision rules is closer to an investment than a cost. A cheap run that repeats indefinitely without learning is closer to waste.

Sub-pages

本页已按主题拆分为三个子页(2026-10-25):

  • 参数与预算机制 — 断路器、effort 档位、配额、例行预算、高端模型子预算、硬路由、订阅档位与定价锚点。执行前的硬边界。
  • 政策与治理模型 — 参数而非政策的组织哲学:摩擦被挣得、凭证据调整、探索预算绑定产物、fast/deep/wait、自动化存在门槛、权限隔离、长时任务停止规则。
  • 观测与归因实践 — 任务级账本与"被接受的结果"成本单位:标准账本字段、构建循环核算、缓存 key 归因、contact sheet 批次化、跨 harness 用途分类。

Evidence across sources

Source Key Claim Relevance
Every — Our AI Costs Jumped 230 Percent Every replaced fixed token budgets with circuit breakers, usage dashboards, and post-run reviews after a Sol test caused a 230% spike. Primary framework and operating parameters.
Every — Life After Automation The cost spike and parameters-not-policies response are featured as a front-of-mind organizational challenge for AI-native teams. The framework is a recurring operations pattern, not a one-time fix.
Every — The Next Era of Great Work Compound Engineering requires capturing what worked so the next similar task becomes cheaper. Connects cost governance to compounding knowledge.

Open questions

  • At what company size do light parameters stop scaling and require formal procurement controls?
  • How should cost governance differ for frontier exploration versus production inference?
  • Can automated evals replace human judgment in deciding whether an expensive run was worth it?

Prompts for witness

  • Which of Jean's current agent workflows have hidden runaway costs because no one reviews the daily spend?
  • What would a circuit breaker look like for the wiki ingest, briefing generation, or coding agent pipelines?
  • When has an expensive AI run produced lasting value, and when has it just felt productive?

Pending Jean review

  • harness-engineering/ai-cost-governance — ai_draft;建议:从三个子页各选一条建议开始试点(参数页选一条预算机制、治理页选 fast/deep/wait、观测页建一行标准账本)。

Sources

Synthesized from 67 sources

Evolution

1 event
  1. absorbed

    Derived from source material

    This page is currently synthesized from 67 sources.

    From Every — Our AI Costs Jumped 230 Percent. I'm Not Setting Token Budgets — Yet., Every — Life After Automation, Every — The Next Era of Great Work, 2026-09-03-02-03, 2026-09-05-08-14To AI Cost Governance — Parameters, Not Policies
    Sources: raw/newsletters/Every/2026-08-17 Our AI Costs Jumped 230 Percent I’m Not Setting Token Budgets Yet.md · raw/newsletters/Every/2026-08-25 Life After Automation.md · raw/newsletters/Every/2026-08-16 The Next Era of Great Work.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-03-02-03.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-05-08-14.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-06-23-01.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-07-08-15.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-07-22-58.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-08-08-14.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-08-23-06.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-10-23-08.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-11-08-05.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-12-08-04.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-12-23-00.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-13-08-15.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-14-08-06.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-15-08-12.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-16-02-18.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-17-22-03.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-17-22-58.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-18-08-04.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-19-08-05.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-23-22-58.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-10-02-17-45.md · /Users/jean/Documents/Obsidian Vault/raw/newsletters/Ben's Bites/2026-09-04 Fable 5.1.md · /Users/jean/Documents/Obsidian Vault/raw/newsletters/Ben's Bites/2026-09-05 Scraping 107M rows of data to build this.md · /Users/jean/Documents/Obsidian Vault/raw/newsletters/Ben's Bites/2026-09-08 The first GPT-6 model.md · /Users/jean/Documents/Obsidian Vault/raw/newsletters/Ben's Bites/2026-09-09 Better design than Fable.md · /Users/jean/Documents/Obsidian Vault/raw/newsletters/Ben's Bites/2026-09-16 Good luck slowing this down.md · /Users/jean/Documents/Obsidian Vault/raw/newsletters/Ben's Bites/2026-10-01 OpenAI DevDay 2026.md · /Users/jean/Documents/Obsidian Vault/raw/newsletters/Every/2026-08-17 Our AI Costs Jumped 230 Percent I’m Not Setting Token Budgets Yet.md · /Users/jean/Documents/Obsidian Vault/raw/newsletters/Every/2026-09-06 A Split Verdict on Fable vs. Astra.md · /Users/jean/Documents/Obsidian Vault/raw/newsletters/Every/2026-09-12 What I Learn When I Run Out of AI.md · /Users/jean/Documents/Obsidian Vault/raw/newsletters/Every/2026-09-16 You're Probably Sleeping On Computer Use.md · /Users/jean/Documents/Obsidian Vault/raw/newsletters/Every/2026-09-18 Why You Should Burn More Tokens.md · raw/briefing/AI Briefing 2026-10-02-17-45.md · raw/briefing/AI Briefing/2026-08-31-23-04.md · raw/briefing/AI Briefing/2026-09-01-08-09.md · raw/briefing/AI Briefing/2026-09-02-00-34.md · raw/briefing/AI Briefing/2026-09-03-02-03.md · raw/briefing/AI Briefing/2026-09-05-08-14.md · raw/briefing/AI Briefing/2026-09-06-23-01.md · raw/briefing/AI Briefing/2026-09-07-08-15.md · raw/briefing/AI Briefing/2026-09-07-22-58.md · raw/briefing/AI Briefing/2026-09-08-08-14.md · raw/briefing/AI Briefing/2026-09-08-23-06.md · raw/briefing/AI Briefing/2026-09-10-23-08.md · raw/briefing/AI Briefing/2026-09-11-08-05.md · raw/briefing/AI Briefing/2026-09-12-08-04.md · raw/briefing/AI Briefing/2026-09-12-23-00.md · raw/briefing/AI Briefing/2026-09-13-08-15.md · raw/briefing/AI Briefing/2026-09-14-08-06.md · raw/briefing/AI Briefing/2026-09-15-08-12.md · raw/briefing/AI Briefing/2026-09-16-02-18.md · raw/briefing/AI Briefing/2026-09-17-22-03.md · raw/briefing/AI Briefing/2026-09-17-22-58.md · raw/briefing/AI Briefing/2026-09-18-08-04.md · raw/briefing/AI Briefing/2026-09-19-08-05.md · raw/briefing/AI Briefing/2026-09-23-22-58.md · raw/newsletters/Ben's Bites/2026-09-04 Fable 5.1.md · raw/newsletters/Ben's Bites/2026-09-05 Scraping 107M rows of data to build this.md · raw/newsletters/Ben's Bites/2026-09-08 The first GPT-6 model.md · raw/newsletters/Ben's Bites/2026-09-09 Better design than Fable.md · raw/newsletters/Ben's Bites/2026-09-16 Good luck slowing this down.md · raw/newsletters/Every/2026-09-06 A Split Verdict on Fable vs. Astra.md · raw/newsletters/Every/2026-09-12 What I Learn When I Run Out of AI.md · raw/newsletters/Every/2026-09-18 Why You Should Burn More Tokens.md

Linked from