Logiciel Solutions Contact Us
Success Stories Tech News Investors Contact Us
whitepaper
8 min read

The AI Governance Decision-Criticality Framework.

Most AI governance programs sort systems into three or four risk tiers and apply the same controls to everything in a tier, which breaks down when one platform hosts a harmless FAQ bot and an agent that can delete production data.

Decision criticality is a rating of how much harm results if a specific AI output or action is wrong in its deployment context. This whitepaper sets out a seven-axis scoring method, legal floors drawn from Colorado SB 26-189, California's CPPA ADMT rules, ECOA, FCRA, the EU AI Act and OMB M-25-21, and the controls each of four tiers requires.

The evidence is regulatory text, public standards and analyst forecasts. The weights, thresholds and worked examples are Logiciel's proposed method, not a regulatory standard, and nothing here is legal advice.

In depth

Criticality Belongs To The Decision, Not The System.

01

Criticality belongs to a decision in context, not to a model.

Risk tiering asks what kind of system you have, while decision criticality asks what happens when this specific decision goes wrong. One support assistant might answer shipping questions, summarize a dispute and, if allowed, issue a refund, which are three decisions with three different consequences. The idea sits inside established work: the Map function of the NIST AI Risk Management Framework is where context, impact and likelihood are established, ISO/IEC 42001:2023 requires an AI management system to assess impacts, and ISO/IEC 42005:2025 gives guidance on running and documenting those assessments. Canada's Algorithmic Impact Assessment uses 65 risk questions and 41 mitigation questions to place a system in one of four impact levels. Score each decision point, then let the highest score drive the controls.

In shortthen let the highest score drive the controls
02

Legal definitions set a floor that an internal score cannot go below.

Colorado SB 26-189, effective January 1, 2027, covers automated decision-making technology that "materially influences" decisions about education, employment, housing, lending, insurance, healthcare or essential government services. That standard is broader than making the decision, so ranking, scoring, recommending and classifying can all qualify. California's CPPA rules require businesses using ADMT for significant decisions to comply by the same date, and a human reviewer does not automatically remove a system from scope under either state. ECOA requires specific principal reasons for adverse credit action even when complex models make that hard, and FCRA Section 615 adds notice duties when a consumer report is involved. An output that touches eligibility, price or terms in these domains should be treated as consequential until counsel says otherwise.

In shortAn output that touches eligibility, price or terms i…
03

Agent risk lives at the tool layer.

An agent carries the criticality of the most dangerous action it can execute without a gate, so every tool and every permission scope on that tool needs its own score. OWASP's LLM06:2025 "Excessive Agency" names three root causes of LLM agent harm: excessive functionality, excessive permissions and excessive autonomy. OWASP also recommends enforcing authorization in downstream systems rather than delegating it to the model, which means the policy engine, not the prompt, decides which actions need approval. A code agent with write and delete access to production and the same agent restricted to opening pull requests and reading logs are making different decisions. Permission scope is a design choice, and changing it changes what is being scored.

In shortchanging it changes what is being scored
04

Autonomy is rising faster than classification methods were built for.

Gartner predicts at least 15% of day-to-day work decisions will be made autonomously by agentic AI by 2028, up from 0% in 2024, and that over 40% of agentic AI projects will be canceled by the end of 2027, citing costs, unclear value and inadequate risk controls. McKinsey's 2025 survey found 23% of respondents' organizations scaling an agentic AI system somewhere in the enterprise and another 39% experimenting. Federal agencies already apply seven minimum practices to "high-impact AI" under OMB M-25-21, including pre-deployment testing, human oversight and appeals. Canada's directive requires assessments to be updated when system functionality or scope changes, not only on a schedule. A tier set at launch describes a system that may no longer exist.

In shortA tier set at launch describes a system that may no…
The detail

Three Intake Questions That Decide Most Of An AI Decision's Tier.

A criticality score is only as sound as the answers behind it. Three questions at intake settle most of the outcome, and each can be answered on paper before anyone writes a test. Each also produces a named artifact that counsel, an auditor or a governance committee can check later.

Zone · 01

What does the output touch

Ask whether the output affects eligibility, selection, price or terms in credit, housing, insurance, employment, education or healthcare. Colorado's "materially influences" standard reaches ranking and recommending, and California's ADMT rules cover similar significant decisions, so an underwriter making the final call does not settle the question. A match sets a minimum tier regardless of the weighted score. The artifact is a legal floor entry naming each matched law and its citation.

Zone · 02

Who checks before it lands

Ask how much a human reviews before the output takes effect, and be strict about what counts. California's ADMT rules recognize human involvement only when the reviewer can interpret the output, weigh other information and has authority to change the decision. A reviewer who only clicks approve is monitoring, not deciding. The artifact is a written description of the review step, naming who reviews, what they see and what they can overturn.

Zone · 03

What can the agent execute

Ask for every tool and permission scope the agent holds, including tools it can reach by calling another agent. OWASP names excessive functionality, permissions and autonomy as root causes of LLM agent harm and recommends enforcing authorization outside the model. Refunds, deletions, production deploys and privilege changes each need their own score and gate. The artifact is a tool inventory with a tier and an approval gate recorded against each action.

By the numbers

The figures that make it a board-level conversation.

Over 40%
of agentic AI projects predicted to be canceled by the end of 2027, citing costs, unclear value and inadequate risk controls
15%
of day-to-day work decisions predicted to be made autonomously by agentic AI by 2028, up from 0% in 2024
23%
of respondents' organizations scaling an agentic AI system somewhere in the enterprise, with another 39% experimenting
Inside the report

What you'll take away.

01

Step 1 - Describe each decision and list every tool it can reach

Write one sentence on who is affected, what changes for them and what the AI contributes. Then list each tool and permission scope the system or agent will hold, since each is scored separately.

02

Step 2 - Check the legal floor before scoring anything

Compare the use case against Colorado SB 26-189, California's ADMT rules, ECOA, FCRA, EU AI Act Annex III and OMB M-25-21. Record any match with its citation and involve counsel.

03

Step 3 - Score seven axes and apply the override rules

Rate impact on people, business impact, reversibility, scale, autonomy, explainability and data sensitivity from 0 to 3. Overrides for legal triggers, severe harm, irreversible autonomy and safety only ever raise the tier.

04

Step 4 - Attach tier controls and register reclassification triggers

Each tier fixes its testing, review, monitoring, documentation and approval level. Register rescoring triggers such as a new tool, a model change, a wider population, an autonomy shift, drift or a new law taking effect.

Questions

Frequently asked.

What is decision criticality in AI governance?

Decision criticality is a rating of how much harm results if a specific AI output or action is wrong in its deployment context. It weighs impact on people and the business, reversibility, scale, autonomy, explainability needs and data sensitivity, with legal definitions of consequential decisions setting a minimum.

The resulting tier sets the testing, human review, monitoring, documentation and approval level the system must meet.

How do you classify AI use cases by risk tier?

Score each decision point on seven axes from 0 to 3, apply the weights and map the total to four tiers. Then apply override rules for legal triggers, severe impact on people, irreversible autonomous actions and safety, which only ever raise the tier.

The final tier is the highest of the use-case tier, the highest ungated tool tier and any legal floor.

When does an AI system need a human in the loop?

An AI system needs a human in the loop when its output is consequential, irreversible or must be explained to the person affected, as in credit, insurance, employment, housing, education and healthcare.

California's ADMT rules count a reviewer only when they can interpret the output, weigh other information and have authority to change the decision. Low-impact, reversible tasks can run on-the-loop with sampling.

What counts as a consequential decision under US AI laws?

A consequential decision is one about education, employment, housing, financial or lending services, insurance, healthcare or essential government services, under Colorado SB 26-189, effective January 1, 2027. California's ADMT rules cover similar significant decisions, and credit adverse actions also trigger ECOA and FCRA notice duties.

Under Colorado, technology that only ranks or recommends can still qualify if it materially influences the outcome.

How do you assign criticality to AI agent actions and tools?

Treat every tool and permission scope as its own decision and score it separately. Reads of public data sit low, while sensitive reads, external writes, financial commits, deletions, production deploys and privilege changes sit high or critical.

The agent's tier is the highest action it can take without a gate, so adding gates and narrowing permissions lowers it, in line with OWASP's guidance on excessive agency.

What is in the whitepaper and how far does its evidence reach?

The whitepaper gives you a seven-axis scoring worksheet, override rules, a tier-by-tier control matrix, agent action classes, reclassification triggers and a ten-step intake checklist. Its evidence is regulatory text, public standards from NIST, ISO and Canada, OWASP guidance and Gartner and McKinsey research.

The weights, thresholds and worked examples are Logiciel's proposed method rather than a regulatory standard, and legal status reflects November 2026 sources, not legal advice.

How much does it cost to put decision-criticality governance in place?

The whitepaper gives no sourced cost figure, because cost depends on how many decisions and agent tools you score. Its argument is that design choices are often the cheaper lever: moving an action into human-in-the-loop review, or narrowing an agent's permissions, can lower a tier without adding testing.

Gartner cites inadequate risk controls among the reasons it expects over 40% of agentic AI projects to be canceled by 2027.

Get the whitepaper

Have it emailed to you.

Drop your details and we'll send The AI Governance Decision-Criticality Framework straight to your inbox - no spam, unsubscribe anytime.

Download whitepaper
Next step

Score one live AI decision before anyone asks how you classified it.

Bring one AI use case or agent you run today, with the tools and permissions it holds. We will work through its criticality score, the legal floor it meets and the controls its tier calls for, alongside your counsel rather than in place of them. SECTION 7 - FAQ - 5 to 8 questions

Book a criticality review