Last November, we published a blog post setting out our definitions of common AI terms. AI has continued to evolve rapidly since then. This update replaces that post and incorporates newer terminology, including “AI Agent,” “Agentic AI,” and “Harness.”

AI governance is hard enough without the added difficulty of assessing risks absent a common understanding of terms like “AI system.” Does that term refer to the model alone, the model and the tools it uses, the model plus the tools and context available to it, or something else? Regulatory definitions are important, but they are often too broad to answer those practical questions. For example, the EU AI Act and the Colorado AI Act describe AI systems in terms such as:

A machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments.

Definitions like these do not provide teams with a practical shared vocabulary for governance, procurement, risk assessment, and deployment decisions. We therefore use the following working definitions when assisting clients with AI adoption.

Definitions

  1. Model is the core software component trained to process inputs and produce outputs (such as text, images, or code) based on patterns, relationships, or representations learned from data. Examples include GPT-5.5, Claude Opus 4.8, and Gemini 3.5.
  2. Tool is the feature, product, application, API, or user interface through which a User interacts with or deploys a Model, including the connectors that allow the Model to interact with additional data (such as ChatGPT Enterprise, Claude, and Gemini) or other Tools (such as persistent memory or an image generator). It is important to note that Tools can serve as Mitigants or sometimes themselves must be mitigated to protect Users and Affected Individuals.
  3. Harness is the set of constraints, configurations, and controls that govern how a Model and Tool A Harness determines what data the Model can access, what external systems it may call, what actions it is permitted to take, and includes Mitigations at the Tool level, such as content filters, prompt restrictions, connector allowlists, logging, permission limits, or prohibitions on accessing external information. Some Harness elements are built into a Tool and cannot be changed by ordinary Users; others may be configured by administrators or system owners.
  4. Training Data is the data used to train, fine-tune, or otherwise develop a Model before it is used for a particular task. It is used to teach the Model patterns and relationships the Model may apply to future inputs. Examples include public data, licensed datasets, vendor-created datasets, synthetic data, internal historical data, and human feedback data.
  5. Operating Data (also called Reference Data or Context) is data made available to a Tool or Model when performing a task, including user prompts, uploaded documents, connected databases, retrieval results, and other task-specific context such as policy libraries or contract repositories. Operating Data may be persistent or temporary.
  6. Use is the function that we are performing with AI (also referred to as the task or purpose). Examples include resume screening, generating images for the firm’s website, customer service chatbots, final document proofreading, and real-time translation of a Zoom meeting.
  7. Users are the individuals who are using the AI Tool or Model to perform a Use. Examples include the in-house legal department, engineers, HR professionals, finance professionals, customers, and the general public.
  8. Affected Individuals are the people who are impacted by the output of a Tool, Model, or AI System, whether directly or by having access to the Tool’s They may (but do not have to) be the same as the Users. Examples include clients of law firms that use AI for legal work and customers who view AI-generated content on companies’ social media channels.
  9. AI System = AI Application = Model + Tool + Harness + Training Data. We think this formulation is largely consistent with the more complex definitions for AI System and AI Application provided by NIST, the OECD, and the EU AI Act.
  10. Use Case = AI System + Operating Data + Use + Users + Affected Individuals.
  11. Mitigants are measures or controls designed to reduce the risks associated with an AI Use Case that do not significantly reduce the value of the Use Case. We refer to these mitigants collectively as Mitigation. Examples include human review by subject-matter experts (SMEs), use of Retrieval-Augmented Generation (RAG) databases to reduce hallucinations, access controls, logging and monitoring, training, and technical safeguards such as data retention controls, content filters, or prompt restrictions.
  12. AI Solution = Use Case + Mitigants.

From AI Systems to AI Agents

The definitions above provide a framework for understanding any AI deployment: the relevant technology (Model, Tool, Harness, Training Data), how the technology is being used and by whom (Operating Data, Use, Users, Affected Individuals), and how those elements combine into an AI System, Use Case, and (with appropriate Mitigations) an AI Solution.

With the rise of agentic AI, the combination of Models, Tools, Harnesses, and Operating Data is no longer confined to a single prompt that produces a single output. AI can now take actions across multiple steps, and complete sequences of tasks with limited or no human involvement, so we’ve added three more definitions to our list:

  1. AI Agent is an AI Tool, or a capability within an AI Tool, that uses a combination of Model(s), Operating Data, and Harness(es) to pursue a defined goal through multi-step planning, reasoning, and action with limited or no human intervention.
  2. Agentic AI is a mode of AI deployment in which one or more AI Agents pursue a defined goal through multi-step planning, sequencing, and actions, in series or in parallel, with human involvement occurring only at selected checkpoints, exception triggers, or final review, if at all.
  3. Agentic AI Workflow (AAW) is an AI-enabled workflow in which one or more AI Agents complete multiple linked tasks in series or in parallel, toward a defined business or operational goal. In an AAW, the output of one task may become the input for another, and the workflow proceeds without human involvement until completed.

Example: Using AI Tools for Screening Resumes

  1. Model = Claude Opus 4.8.
  2. Tool = a vendor’s Resume Screening Tool.
  3. Harness = system rules that block Claude from viewing names or pronouns, searching the external web, or accessing unapproved internal sources.
  4. Training Data = anonymized historical resumes used for fine-tuning the Model, in addition to the training data Anthropic used when developing Claude Opus 4.8.
  5. Operating Data = new job applications.
  6. Use or Task = screening resumes to identify the 50 most qualified candidates.
  7. Users = the recruiting/HR team.
  8. Affected Individuals = the job applicants.
  9. AI System or AI Application = (1+2+3+4) = Claude Opus 4.8 (Model) + the vendor’s Resume Screening Tool (Tool) + system rules (Harness) + anonymized historical resumes (Training Data).
  10. Use Case = (1+2+3+4+5+6+7+8) = AI System + new job applications (Operating Data) + task to screen resumes (Use/Task) + recruiting team (Users) + job applicants (Affected Individuals).
  11. Mitigants/Mitigation = pilot program testing, ongoing human review by subject-matter experts, and periodic bias assessments.
  12. AI Solution = (1+2+3+4+5+6+7+8+11) = Use Case + Mitigation.

From AI Systems to AI Agents:

  • AI Agent = an AI Tool that identifies missing information in job applications, compares applications against predefined job-related criteria, flags candidates for recruiter review, and generates an explanation of the criteria applied.
  • Agentic AI = Using the Tool to complete multiple screening steps, such as checking minimum qualifications, identifying incomplete applications, and preparing recruiter review packets, with little to no human involvement.
  • Agentic AI Workflow = A workflow in which, without further human instruction, the AI Tool receives applications, ranks candidates for the HR team based on approved role requirements, prepares a ranked review queue, logs the criteria applied, and routes the output to HR personnel for review.

We recognize that these definitions are not perfect, and there are certainly choices we’ve made where reasonable minds may differ. This blog post is only meant to be a starting point for a dialogue among AI practitioners on workable terms. We are continually refining these and other definitions and welcome any suggestions for improvement. Successful AI adoption is extremely challenging, but having a simple, shared vocabulary for what we are all trying to do should not be.

The authors would like to thank Debevoise Summer Associate Kathryn Ann Tso for her contribution to this blog post.

                                                   *      *      *

To subscribe to the Data Blog, please click here.

The cover art used in this blog post was generated by Gemini 3.1 Pro and GPT-5.5.

The Debevoise STAAR (Suite of Tools for Assessing AI Risk) is a monthly subscription service that provides Debevoise clients with an online suite of tools to help them responsibly fast-track their AI adoption. Please contact us at STAARinfo@debevoise.com for more information.