The Control Stack
The AI controls you need are already written; the decision is whether to run them.
The controls are already specified
Every AI workload that makes consequential decisions needs five controls: a written role for the model, human review before its output acts, tamper-proof logs, continuous testing, and monitoring that pages a person. NIST's AI Risk Management Framework,[1] the OWASP Top 10 for LLM Applications,[2] ISO/IEC 42001,[3] and the EU AI Act[4] already specify all five.
The decision in front of you is whether to fund and run them, sized by what one bad output would cost. Start with roles and logging, the two cheapest, while you inventory your AI workloads. They take days to add during an inventory. Retrofitted after an incident, they cost the incident first.
This paper follows The Price Is Not the Price, whose five questions test whether a model is the right tool at all. A workload that passes them still needs governance. The four costs it carries beyond the vendor's invoice are these controls stated as money: evaluation, review labor, error and rework, and compliance.
AI fails in ways other systems don't
An AI model differs from the systems the enterprise already governs in three ways. It can give a different answer to the same input, so it cannot be tested once and trusted. It reads instructions and data through the same channel, so any document it reads can try to give it orders. And it acts: agents hold tools and credentials and can change systems of record.
The OWASP 2026 list, the first to weight 6,639 documented real-world incidents alongside expert judgment, names the failure modes that follow.[2] Five shape the stack:
Hallucination. The model states wrong answers with full confidence. It has no internal model of the domain, only statistical patterns, so its errors come from the design itself. Cory Doctorow calls this "theory-free" inference.[6] It can be measured and reduced, but not eliminated. The measure is the escape rate: the share of bad outputs that get past every check.
Prompt injection. First on every OWASP list, with no complete fix. Every document, upload, and tool result is untrusted input, so limit what a compromised model can reach.
Sensitive data disclosure. Anything a model can read, it can repeat. Data classification decides what it may see; logs prove what it saw.
Excessive agency. Up from sixth to third in 2026 as agents took on consequential work. With broad tool access, a wrong answer becomes a wrong action.
Unbounded consumption. Cost is a failure mode. Uber spent its planned 2026 AI coding budget in four months, then capped spending per employee.[5] They found out from the bill.
One more risk sits outside the list: vendors update models without notice, and behavior changes with them. Continuous testing catches it.
What each control requires
1. Role definition. For each workload, write down what the model may read, produce, and touch, and who owns the decision its output feeds. For agents, enforce it with least-privilege credentials, an allowlist of tools, and no writes to systems of record except through review. The EU AI Act calls this human oversight.[4]
2. Review gates. A qualified person checks output before it acts. Size review by the cost of one bad output: full review where an error is costly, documented sampling elsewhere. Staff the gate with people senior enough to reject output, and test it. Seed known-bad outputs and count how many reviewers catch, because a reviewer who approves everything is indistinguishable from no reviewer at all. For banks, LLM Workloads in the Model-Risk Gap (available on request) maps this gate to effective challenge.
3. Immutable logging. Keep a tamper-proof record of every decision: capture inputs, outputs, model and prompt versions, retrieved context, tool calls, and the review decision. The same records feed testing, answer an auditor, and meet the EU AI Act's record-keeping obligation.[4] Build it first. It is the cheapest control and the only one you can't backfill.
4. Continuous testing. Keep a versioned set of representative inputs with known-correct answers, including injection attempts. Run it on every change and on a schedule, since vendors don't always announce updates. When escape rates breach a threshold, the whole workload goes back under review. The same suite lets you test a cheaper standby model before you need it.
5. Monitoring with alerting. Track output drift, latency, escape rates, and spend per task against budget. Daily token spend per workload, with anomaly thresholds, catches both the runaway loop and the quiet change that doubled costs. Each alert pages a named person, and each workload has an owner who can pull it and a tested rollback.
Deploy in cost order
Deploy the cheapest controls first, starting during a two-week inventory of every AI workload:
Roles and logging on every workload that passed the five questions. This takes days, and every other control reads from it.
Test suites for the highest-stakes workloads, seeded with injection attempts.
Review gates sized by the cost of one bad output, with the catch-rate check from day one.
Monitoring thresholds and a named incident owner per workload.
Red-team exercises and the remaining workloads follow.

Spend follows the cost of one bad output
Nothing in the stack is a new budget line. A workload where a bad output costs almost nothing earns logging, a light test set, and sampled review. A workload where one error is material earns the full stack. Matching depth to stakes keeps the program affordable where errors are cheap and defensible where they aren't.
The spend buys one body of evidence with four audiences: a NIST-aligned procurement questionnaire, an ISO/IEC 42001 audit, EU AI Act documentation, and, for banks, the examiner. Once the records exist, answering a new framework means querying them. Workloads that carry the stack are also the easiest to defend to a CFO, which makes them the last ones cut when budgets tighten.
Where each control sits in the frameworks
Control | OWASP GenAI LLM Top 10 (2026)[2] | NIST AI RMF function[1] | EU AI Act, high-risk[4] |
Role definition | LLM03 Excessive Agency; reduces LLM01 blast radius | Govern, Map | Human oversight (Art. 14) |
Review gates | LLM07 Misinformation; LLM10 Improper Output Handling | Manage | Human oversight (Art. 14) |
Immutable logging | Evidence layer for all ten | Measure | Record-keeping (Art. 12) |
Continuous testing | LLM01 Prompt Injection; LLM05 Data and Model Poisoning; LLM09 Vector and Embedding Weaknesses | Measure | Accuracy and robustness (Art. 15) |
Monitoring with alerting | LLM06 Unbounded Consumption; drift detection | Manage | Post-market monitoring (Art. 72) |
ISO/IEC 42001 sits one level up. It certifies that a management system running these controls exists, which is how you prove the stack to a customer or auditor.[3]
I sell no software and take no vendor margin. The full method is in this paper. Working with me adds speed and pattern recognition from prior deployments.
If your AI workloads reached production before their controls did, I can help you put the stack in place, cheapest controls first.
Notes
National Institute of Standards and Technology, "Artificial Intelligence Risk Management Framework (AI RMF 1.0)," NIST AI 100-1, January 26, 2023, and "Generative Artificial Intelligence Profile," NIST AI 600-1, July 26, 2024. https://doi.org/10.6028/NIST.AI.600-1. The profile names twelve generative-AI risk categories, including confabulation (NIST's term for hallucination), with suggested actions across the Govern, Map, Measure, and Manage functions. NIST states that AI RMF 1.0 is being revised under the White House AI Action Plan; the profile remains the canonical operational reference. https://www.nist.gov/itl/ai-risk-management-framework
OWASP GenAI Security Project, "OWASP Top 10 for LLM Applications," 2026 edition, published August 4, 2026. https://doi.org/10.5281/zenodo.22109014. The 2026 ranking weighted analysis of 6,639 documented real-world incidents at 25 percent alongside expert consensus. Prompt Injection has held first place in every edition, and Excessive Agency rose from sixth to third. The project also publishes a companion Top 10 for Agentic Applications and an Agent Control Standard.
International Organization for Standardization and International Electrotechnical Commission, "ISO/IEC 42001:2023, Information technology: Artificial intelligence: Management system," December 2023. The first certifiable AI management-system standard; crosswalks between it and NIST AI 600-1 are in active practitioner use.
European Union, Regulation (EU) 2024/1689 (Artificial Intelligence Act), as amended by Regulation (EU) 2026/1744 (the AI Omnibus), published in the Official Journal July 24, 2026, in force July 27, 2026. https://eur-lex.europa.eu/eli/reg/2024/1689/oj. High-risk obligations cited here: record-keeping (Article 12), human oversight (Article 14), accuracy and robustness (Article 15), and post-market monitoring (Article 72). The omnibus moved their application to December 2, 2027 for stand-alone Annex III systems and to August 2, 2028 for AI in products covered by EU product-safety law. Amendment dates per Sarah Pearce and Thomas Nietsch (K&L Gates), "EU Digital Omnibus On AI Enters Into Force," Mondaq, August 3, 2026. https://www.mondaq.com/uk/new-technology/1825722/eu-digital-omnibus-on-ai-enters-into-force
Uber CTO Praveen Neppalli Naga, April 2026 disclosure to The Information: the company exhausted its planned 2026 AI coding budget in the first four months of the year. Bloomberg, "Uber Caps Usage of AI Tools Like Claude Code to Cut Costs," June 2, 2026: a cap of $1,500 per employee, per tool, per month. https://news.bloomberglaw.com/artificial-intelligence/uber-caps-usage-of-ai-tools-like-claude-code-to-cut-costs-1. Coverage: TechCrunch, June 2, 2026. https://techcrunch.com/2026/06/02/uber-caps-employee-ai-spending-after-blowing-through-budget-in-four-months/
Cory Doctorow, "Swapping money for expertise," Pluralistic, October 6, 2026. https://pluralistic.net/2026/10/06/nonfungible/. Contrasts theory-free statistical systems with symbolic programs built on an explicit model of the domain; his chess example is a statistical system that sometimes moves a piece onto a square its own piece already occupies.
© 2026 David Nix. Licensed under CC BY-NC-ND 4.0. Questions: dave@dave-nix.com
