Every discipline that swallows a new technology hits the same wall first, and the wall is vocabulary. When derivatives arrived in corporate treasury, risk managers had to learn what a delta was. When cloud arrived, audit teams had to work out what a shared responsibility model meant. Both times the new words eventually got domesticated, folded into the existing control catalog with some adjustment and a fair amount of grumbling.
AI is a different kind of vocabulary problem. It is not just that GRC people need new words for new artifacts. Some AI concepts do not fit the semantic structure of governance language at all. They describe behaviors that our existing vocabulary has no way to represent honestly. And when the vocabulary cannot hold the concept, the control framework built on that vocabulary cannot hold it either.
So this piece does two things. First, a working lexicon for anyone who has to speak about AI precisely in a governance setting. Second, the harder argument: even perfect fluency in these terms is not enough without a few new hybrid concepts that bridge what AI actually does and how governance actually works.
The gap has teeth. Someone who does not understand model drift cannot judge whether a validation program is adequate. A board risk committee that has never really grasped emergent behavior cannot assess whether its own oversight structure could detect it. An internal auditor who cannot tell a fine-tuned model from a foundation model cannot design a model governance audit worth the name.
It gets harder because the vocabulary will not sit still. Terms enter active use within months of a research paper. Others get repurposed, redefined, or quietly abandoned. Chasing every development is the AI team's job, not yours. Your job is fluency in the subset that has become load-bearing for governance: the terms that decide which controls are necessary, which risks are being accepted, and who is accountable when something goes wrong.
What follows covers the terms that actually show up in governance, regulatory, and compliance conversations as of 2026.
| Acronym | Full term | Why it matters for GRC |
|---|---|---|
| AGI | Artificial General Intelligence | A hypothetical system with human-level cognitive ability across all domains. Regulatory frameworks increasingly reference AGI thresholds as triggers for enhanced oversight. |
| AIaaS | AI as a Service | Cloud-delivered AI capability. Who owns which risk under the shared responsibility model is the central governance question in enterprise AI procurement. |
| GPAI | General Purpose AI | The EU AI Act classification for broadly applicable foundation models. GPAI models with systemic risk face enhanced obligations, including adversarial testing and incident reporting. |
| GRC-AI | Governance, Risk and Compliance, Artificial Intelligence | The emerging sub-discipline where the two meet: model governance, AI risk management, responsible AI compliance, and AI regulatory compliance. |
| LLM | Large Language Model | Foundation models trained mainly on text at scale. Hallucination, prompt injection, and drift are the LLM-specific risks most model risk frameworks are built around. |
| MCP | Model Context Protocol | A protocol developed by Anthropic for structured communication between models and external tools. Relevant when assessing agentic architectures and tool-use security. |
| MLOps | Machine Learning Operations | The practices and tooling for running models in production. Model governance should be wired into the MLOps pipeline, because that is where monitoring and revalidation actually happen. |
| RAG | Retrieval-Augmented Generation | See the lexicon entry above. The primary architectural mitigation for hallucination in knowledge-intensive applications. |
| RLHF | Reinforcement Learning from Human Feedback | See the lexicon entry above. The training technique that shapes behavioral guardrails. |
| SLA | Service Level Agreement | In AI contracts these still mostly cover platform availability. Extending them to output quality and drift is the live frontier in AI vendor governance. |
| T&S | Trust and Safety | The function and design practice for preventing harmful output and misuse. Increasingly a due diligence dimension in vendor assessment. |
| FLOP | Floating Point Operation | The compute unit used to measure training scale. The EU AI Act sets its GPAI systemic risk threshold at 1025 FLOPs, using compute as a regulatory proxy for capability. |
Everything above describes AI concepts. The harder question is whether the governance vocabulary we would use to manage those concepts, the language of control frameworks, risk registers, audit standards, and board oversight, is up to the job. The answer comes in three parts.
Audit vocabulary rests on an assumption: that a decision can be reconstructed from its inputs, the rules applied, and the output produced. Traceability, walkability, process inspection. Auditability in the conventional sense needs a readable trail.
Black-box models do not leave one. A forward pass through billions of parameters is not a trail; it is a mathematical transformation that cannot be run backward into a human explanation of why this particular output appeared.
Governance has answered with explainability and interpretability, and here is where a quiet error creeps in. Those are not synonyms for auditability. A tool like LIME or SHAP gives you a local approximation of behavior around one prediction. Useful, but it is not an audit trail, and a framework that files explainability output as audit evidence is leaning on a surrogate the underlying method was never built to support.
Risk registers enumerate known risks. They have no slot for a risk that could not have been known in advance, one that arises from a property nobody could predict from the design spec. The whole vocabulary of risk management, likelihood, impact, residual risk, appetite, assumes a universe of enumerable events. Emergent behavior does not respect that assumption.
You can see the regulatory system reaching for this. The EU AI Act's continuous post-market monitoring requirement, the FTC's emphasis on ongoing behavioral testing, calls for system-level AI risk assessment from financial stability bodies: all of it reflects a recognition that a risk register applied to AI without modification produces an analysis that is incomplete by construction. The failure modes that materialize tend to be exactly the ones the register does not contain.
Oversight language assumes a timescale on which human review means something. Review, approval, escalation, four eyes, sign-off. Agentic systems running at scale erase that timescale.
The failure is not that people are too slow to care. It is architectural. The system does not present outputs for review at all. It executes a sequence of actions autonomously, and by the moment a human could weigh in, the action is done.
What governance needs here is vocabulary that does not presuppose review: programmatic guardrails, constraints encoded in the environment that operate at the agent's own speed; capability boundaries, hard limits on what an agent may do without authorization; and action audit trails, reconstructions of what the agent did that make accountability possible afterward even when it was impossible in the moment. None of these map cleanly onto anything already in the oversight vocabulary.
Closing the gap is not a matter of stapling AI words onto a GRC dictionary. It takes a small number of genuinely new hybrid concepts that bridge how AI behaves and how governance institutions actually work. Three are doing the most load-bearing work right now.
Fiduciary duty has always attached to human actors in positions of trust: directors, trustees, advisors, people obligated to act in the interests of those they serve. Now AI systems are making decisions in fiduciary contexts, including investment management, benefits eligibility, medical triage, and lending. That raises a live question: does the deploying organization take on a form of algorithmic fiduciary duty, a heightened obligation of care calibrated to what the system can and cannot do? Several EU AI Act provisions imply a version of it for deployers of high-risk systems.
Not an audit of the model's internal logic, which is not available for black-box systems, but a documented record of behavior over time: training data provenance, evaluation results at deployment, monitoring metrics through the production period, incidents and near-misses, model updates and what they did to evaluation performance, and the human decisions taken in response to model output. This is the artifact that makes accountability possible after the fact when traditional process auditability was never on the table. Several frameworks, including the EU AI Act and FDA guidance for AI-enabled medical devices, are converging on requirements that amount to mandating one for high-risk deployments.
Distinct from human-in-the-loop (per-output review) and human-on-the-loop (monitoring with intervention). Human-in-command is the principle that people retain ultimate authority over the consequences of system behavior even when reviewing every output in real time is impossible. In practice it means three things: the environment encodes the boundaries of acceptable action, escalation triggers fire as those boundaries are approached, and humans get actionable information on a timescale where response still means something. This is the control topology agentic governance actually requires.
This piece is part of a series on governing AI inside corporate accountability frameworks. See also The Four Blind Spots of Force-Fitting AI Into Traditional Governance, and Top 5 Governance Skills on what the work requires of a career. For roles at the frontier of this discipline, explore the AI governance leadership landscape.
Same failure mode, different register. Hallucination is the common term; confabulation is the more precise one borrowed from neuropsychology. Either way the governance implication holds: AI output in high-stakes domains needs verification pipelines, not more human review, because reviewers cannot reliably catch well-formed false output without checking every claim themselves.
Retrieval-Augmented Generation grounds output in retrieved documents before the model generates, which cuts hallucination risk substantially without eliminating it. It also adds new risk dimensions: whether the retrieval source is reliable, whether indexed documents are current, whether attribution is accurate. If you are deploying RAG in compliance or legal work, retrieval source governance belongs in your framework.
Agentic systems complete multi-step tasks autonomously, deciding as they go without review at each step. That demands a different control topology entirely: programmatic guardrails, capability boundaries, action audit trails, and kill switches, rather than approval queues.
Prompt injection hides malicious instructions inside data an AI agent processes, such as a document, a web page or an email, causing the agent to follow the attacker rather than the user. It is a genuinely new attack surface. Traditional application security frameworks assume instructions and data travel on separate channels, and for agentic systems that assumption does not hold. Any agent with tool access or external data reach needs this in its threat model.
Because they answer a different question. A tool such as LIME or SHAP gives a local approximation of model behavior around a single prediction. That is useful, but it is a reconstruction after the fact, not a record of what happened. Audit vocabulary assumes a decision can be walked back through its inputs and applied rules. A forward pass through billions of parameters leaves no such trail. Filing explainability output as audit evidence leans on a surrogate the method was never built to support.
Not an audit of the model's internal logic, which is unavailable for black-box systems, but a documented record of behavior over time: training data provenance, evaluation results at deployment, monitoring metrics through the production period, incidents and near-misses, model updates and their effect on evaluation performance, and the human decisions taken in response to model output. It is what makes accountability possible after the fact when process auditability was never on the table.
Human-in-the-loop means a person reviews each output before it takes effect. Human-on-the-loop means monitoring with the ability to intervene. Human-in-command is the principle that people retain ultimate authority over the consequences of system behavior even when reviewing every output in real time is impossible. In practice it requires three things: the environment encodes the boundaries of acceptable action, escalation triggers fire as those boundaries are approached, and humans receive actionable information on a timescale where response still means something.
AI safety is the research discipline concerned with whether systems are fundamentally aligned with human values. AI governance is the organizational discipline of managing AI risk through policy, controls, oversight, and accountability. They inform each other but operate at different levels. GRC professionals practice governance and follow safety research to understand the landscape they are managing.
It is an alignment method, developed by Anthropic, that trains models to check their own output against stated principles. For compliance it lands as a due diligence question: the alignment approach affects how stable behavioral constraints stay across model updates. If you have specific behavioral requirements around content restrictions, refusals, or output format, you should understand your vendor's approach as part of model risk assessment.
GRCcareers.ai publishes ongoing analysis of the GRC-AI career landscape, and the AI governance career guides cover the roles now emerging, from analyst through to Chief AI Officer. The wider role directory maps governance and compliance positions across sectors.
Follow GRC Careers on LinkedIn
New essays, the frameworks that actually move, and the governance roles worth knowing about. Written by practitioners, published as it happens.
Follow on LinkedIn →