RESEARCH

AI FinOps research.

Field notes on reducing LLM spend without hiding quality regressions. Start with the category guides, then use the technique pages for specific levers.

Start here

AI FinOps - the operating model for measuring, allocating, optimizing, and reconciling LLM spend.
LLM cost monitoring - what to track, how to tag it, and which dashboards actually matter.
OpenAI cost optimization - the highest-leverage ways to lower API spend in production.
LLM cost optimization: a practical guide ; the full playbook: routing, caching, compression, batch APIs, and provider arbitrage.
How to build an LLM CFO function ; the role, cadence, metrics, and first-90-days plan for owning the AI bill.
Provider price benchmarks ; current list-price references and methodology.
Glossary ; plain-English definitions for billing and optimization terms.

Governance

Budget enforcement belongs inside the agent loop - why spend controls have to be enforcement points, not alerts.
Your agent harness decides how much lock-in you buy - the architecture question that decides whether future AI price cuts reach your P&L.
DeepSeek Harness and the real cost of delegation - why delegation depth, not token price, is the number that moves an agent bill.
When should an agent stop? DeepSeek Harness goals - the stopping rule is the budget, and most agent systems have never written one.
Who owns AI value in the finance function - AI value management is the top skillset FinOps is hiring for. Finance owns the numerator, the business owns the denominator, and the join is the operating model.
AI governance framework — the five-layer operating model and 90-day plan.
AI governance for finance leaders — the pillar guide: policy, allocation, accountability.
What is AI governance? — plain-English definition for 2026.

New and timely

Evaluating an open-source agent harness before you commit - a short checklist for judging an agent harness, worked against DeepSeek's.
What is DeepSeek Harness? A plain explainer - what the harness is, how to run it, and the four questions it should make you ask about your own stack.
73% of agentic AI projects went over budget - 73% of 127 reviewed enterprise agentic implementations ran over budget, averaging $2.3M in costs nobody modelled. The overruns are forecasting failures, not waste.
A quarter of planned AI spend is slipping to 2027 - Forrester expects 25% of planned AI spend to slip into 2027. Deferral is reallocation, and undefended line items are what moves.
FOCUS 1.5 will make AI costs comparable - FOCUS 1.5 lands December 2026 with AI token fields and a Price Sheet dataset. Comparable billing data is negotiating leverage and audit evidence.
Gartner's 2026 AI spending forecast, read as a CFO - $2.59T worldwide, $407B enterprise. What Gartner's 2026 AI forecast is good for in a budget conversation, and where it stops being useful.
A standards body for token costs - the Linux Foundation launched a standards body for AI cost on 4 August 2026 with 29 founding members. What it changes for finance, and what not to wait for.
Claude Max 20x vs Qwen 3.8 Token Plan - hands-on comparison of Opus 5 and Qwen 3.8-Max on real multi-file refactoring, retry loops, and task economics.
Claude Max 20x vs Alibaba Qwen 3.8 - real-user telemetry audit comparing Claude Max 20x against Alibaba Token Pro on heavy agentic coding workloads.
Model concentration risk - what the 19-day Fable 5 shutdown taught CFOs about single-provider exposure, failover readiness costs, and why multi-cloud is not multi-provider.
What the State of FinOps 2026 data tells CFOs - 98% of organizations now manage AI spend, 73% blew their budget; what the survey means for forecasting and governance.
Context compaction is a cost lever - trimming and compressing context cuts input tokens 50–70% on long-running chats and agents.
Evals need cost discipline too - why evaluation quality and evaluation economics now have to coexist.
Token prices fell 80% and your bill still went up - per-token prices dropped 30–50% per year, yet reported AI bills keep rising because usage and reasoning tokens grew faster.
Inference vs training: where the AI GPU bill actually goes - why ~55–80% of enterprise AI GPU spend is now inference, not training.
Multi-cloud AI cost allocation - normalizing token cost across providers, tagging strategy, and the reconciliation headache.
Multimodal costs sneak up faster than text costs - why image, audio, and realtime AI often need their own cost model.
Open-weight model total cost of ownership - self-hosting an open-weight model is only cheaper than a frontier API above a utilization threshold.
AI agent spending limits - budgets and controls for finance workflows: AP automation, procurement, reporting agents.
Coding agent cost control - loop termination and guardrails specific to autonomous code agents like Claude Code and Cursor.
Reasoning tokens are the hidden line item on your AI bill - why invisible output spend is becoming a real production issue.
Agent spend guardrails - the budgets, retry limits, and loop controls that keep agent workflows sane.
AI chargeback and showback are becoming real in 2026 - what has to be true before allocating AI spend internally.
AI FinOps in 2026 - why cost optimization is shifting from isolated prompt tricks to an operating model.
OpenAI Flex vs Batch - the 2026 playbook for moving low-priority work off the expensive standard path.
Prompt caching in 2026 - why teams still leave money on the table even after "turning caching on."
Background mode is an economics feature, not just a reliability feature - why async AI work changes cost architecture, not just uptime.
Built-in tools are not free sidecars anymore - why search, retrieval, and code execution deserve their own budgets.
Your GenAI telemetry schema is now a cost decision - why observability structure now shapes FinOps quality.
Conversation state is a cost lever now - why state handling is becoming a real AI cost architecture problem.

Optimization levers

Every agent tool you ship is billed on every call - the tool catalogue is a recurring charge, not a one-off setup cost.
Batch API routing: 50% off for the work that can wait
Provider arbitrage: same model, different price
Semantic caching for LLMs
LLM cost per request - the unit economics metric and how to segment it.
Model routing without quality regressions
Prompt caching: OpenAI vs Anthropic vs Bedrock

Accounting and tooling

An agent audit trail finance can actually read - what an append-only agent log gives finance that a wrapped SDK never will.
What DeepSeek Harness gets right about failing safe - the reliability failures behind agent bills, and the seven rules that prevent them.
Is your AI spend capex or opex? - a capitalization decision tree, and how to document a defensible position.
Reporting AI spend to the board - the five-part quarterly pack, and the questions directors actually ask.
AI spend variance analysis - splitting the miss into volume, rate, mix, and efficiency.
AI vendor negotiation - the six clauses finance must negotiate, and why timing drives the discount.
Proving AI ROI - the attribution problem, three methods ranked by rigour, and what survives board scrutiny.
LLM cost dashboards - the five views that matter and the fields they need.
Scale Tier vs Flex vs Batch vs Standard
Langfuse vs LiteLLM vs OpenLIT - the OpenTelemetry-native angle on cost and usage telemetry.
LLM usage metering for customer billing - metering points, billing units, discount handling, and reconciliation for SaaS platforms.
LiteLLM vs Helicone vs LangFuse
LLM cost monitoring: what to track and how to control it
How to track AI token usage - token types, per-request fields, and aggregation levels.
Tracking LLM costs with OpenTelemetry GenAI conventions - gen_ai.* attributes and deriving cost from spans.
Cache invalidation cost: the hidden line item
Cache-read tokens: the baseline trap

All research (A–Z)

Generated from the site build — every published article, always.

← Back to llmcfo.com