Glossary

A retail model where the shopping experience adjusts in real time based on individual behaviour, context, and intent. It's broader than personalisation: personalisation targets segments, adaptive retail responds to the individual session.

A structured, repeatable test that runs target queries across ChatGPT, Claude, Gemini, and Perplexity and records whether a domain appears in citations or body text. The only way to see leading AEO signal most teams have access to.

The structural design of an AI agent: how memory is stored, how tools are accessed, how tasks get decomposed, and how decisions get made. An agent built without persistent memory can't learn across sessions, regardless of model capability.

Systematically testing AI agent behaviour against defined success criteria: accuracy, safety, latency, and failure handling. A capable model can still produce a broken agent if the scaffolding around it is poorly designed.

The specific ways an AI agent can fail: hallucination, tool misuse, infinite loops, context overflow, cascading errors in multi-step tasks. Cataloguing them before deployment, not after, is part of responsible AI risk design.

Coordinating multiple AI agents working in parallel or sequence toward a shared goal. An orchestrator decomposes tasks, routes them to specialists, and synthesises results; it doesn't need to be the most capable model, just the best router.

A structured group of specialised AI agents collaborating on complex tasks, each with a defined role, memory scope, and tool access. The distinction between a well-designed agent team and an expensive mess is usually clarity of role boundaries.

A standard for AI agents to communicate and hand off tasks directly without routing through a central orchestrator. Google's Agent Development Kit ships with native A2A support so peer-to-peer teams stay auditable.

Agentic AI refers to systems that pursue goals across multiple steps without a human confirming each action. The practical distinction: errors at step two propagate through every step that follows.

Google's generative-AI answer panel shown above traditional search results. Replaced the Search Generative Experience (SGE) pilot in 2024. Associated with a measured 47–61% click-through drop on informational queries.

Structuring content so it gets cited by AI search systems like ChatGPT, Perplexity, and Google AI Overviews. Unlike SEO, the goal isn't a click; it's placement inside the AI's answer. See AI search visibility.

A verification interface that surfaces only the agent outputs needing human judgment, ranked by what needs you next. The shift from 'where is this?' to 'what needs me?' is the core of VUX design.

A structured approach to AI deployment that constrains scope, audience, or reversibility of trials so that failure modes remain observable and recoverable before broader rollout.

Choosing between building a custom AI capability in-house or purchasing an existing solution. Foundation models have commoditised what used to require specialist teams, so 'build' should be reserved for genuine differentiation.

An executive responsible for an organisation's AI strategy, governance, and capability development, sitting at the intersection of technology leadership and business strategy. Sasha explores what happens when the CAIO role is filled by an AI agent.

A failure pattern in neural networks where learning new information overwrites previously learned knowledge. Complementary Learning Systems theory offers a biological model for why separate memory stores help prevent this.

A technique that instructs a model to reason step by step before producing an answer. The reasoning trace changes the output because each intermediate step constrains what comes next. See Google's prompt engineering whitepaper.

An AI agent that handles coordination, prioritisation, and routing within a larger agent system. In a multi-agent org, it mirrors its human equivalent: it ensures the right agent gets the right task at the right time.

The state in which an AI system's outputs differ meaningfully from a human baseline while remaining internally consistent and traceable to a defensible reasoning chain.

A neuroscience framework proposing two distinct memory systems: the hippocampus for fast learning and the neocortex for slow consolidation. Sasha's article on AI memory architecture applies CLS theory to agent design.

A multi-agent failure mode where small errors at one step propagate through every step that follows. A handful of near-reliable agents chained together stops being near-reliable by the end. Why verification matters more than generation.

The maximum amount of text (measured in tokens) an AI model can process in a single interaction, covering both input and output. Everything outside it is invisible to the model unless retrieved explicitly.

AI systems designed to conduct natural-language exchanges with users, from simple chatbots to agents capable of multi-turn reasoning, tool use, and multi-step task execution within a single conversation.

Google's set of standardised page-experience metrics — Largest Contentful Paint, Interaction to Next Paint, and Cumulative Layout Shift — used as ranking signals and increasingly relevant to AI-indexed content quality.

A governance construct defining who can make, approve, or veto a given class of decisions. In AI contexts, agents can act faster than human review allows, creating accountability gaps if ownership isn't defined before deployment.

A human-in-the-loop decision mode in which an operator consciously evaluates an AI recommendation before acting, as distinct from reflexive approval or fully automated execution.

The distance between what an AI system can do in a controlled demo and what it reliably does in production. Closing it requires structured evaluation, staged rollout, and instrumentation, not just a better model.

A human-centred methodology prioritising user understanding, iterative prototyping, and validated learning over upfront specification. See Sasha's piece on combining TRIZ with design thinking for AI-assisted innovation.

Google's NLU platform for building conversational interfaces, covering intent classification, entity extraction, and dialogue management. Best suited for transactional tasks where predictability matters more than generative capability.

A multi-sided network where a platform creates value by enabling interactions between distinct user groups, with value compounding as more participants join. See network effects in AI.

The Design Council's process model representing two phases of divergent and convergent thinking: discover then define the problem, develop then deliver a solution. It tells you what mode of thinking belongs in each phase, not what to do inside it.

Google's quality framework: Experience, Expertise, Authoritativeness, and Trustworthiness. E-E-A-T signals (named authors, cited sources, domain experience) are what separate content that ranks from content that doesn't.

A memory type that stores records of specific past interactions or task executions, retrievable by time, context, or outcome. Without it, an agent repeats questions and misses patterns across sessions. See AI memory systems.

The EU's AI regulation, classifying systems by risk level (unacceptable, high, limited, minimal) with corresponding obligations for developers and deployers. Organisations deploying AI into EU markets need to know which tier applies.

Providing a model with a small number of worked examples in the prompt, demonstrating the desired output format before asking it to generalise. Examples often outperform lengthy instructions. See Google's prompt engineering whitepaper.

An AI model at or near the current capability frontier: GPT-4o, Claude, Gemini, and their successors. What's frontier today is mid-tier in eighteen months, so product decisions built around specific capabilities need to account for that drift.

AI systems that produce new content (text, images, code, audio, video) from learned patterns rather than retrieving or classifying existing content. See generative vs predictive AI.

Optimising content for visibility inside generated answers from AI search engines like ChatGPT, Perplexity, and Google AI Overviews. Coined by Aggarwal et al. (2023). Broader umbrella than AEO, which targets direct-answer extraction specifically.

The policies, roles, processes, and controls an organisation puts in place to oversee how AI systems are developed, deployed, and monitored. Sasha's article on AI risk design treats governance as a design problem, not a compliance exercise.

Constraints applied to AI system inputs, outputs, or intermediate steps to prevent unsafe, off-topic, or non-compliant behaviour. A common mistake is treating them as a one-time configuration rather than an ongoing calibration problem.

When an AI model generates plausible-sounding but factually incorrect information with the same confidence as accurate content. Mitigations include RAG (grounding outputs in verified sources) and human-in-the-loop review.

An architecture pattern where a central orchestrator coordinates specialist sub-agents, routing tasks outward and consolidating results inward. The most common multi-agent pattern because it keeps communication manageable.

Workflows where humans and AI systems work together, each handling the parts of a task they do better. The design challenge is knowing where to put the boundary, which shifts as model capability improves.

Designing AI systems that keeps human needs, context, and oversight at the centre of technical decisions rather than optimising for model capability alone. The goal is AI that fits how people actually work.

A design pattern where a human reviews or approves AI outputs at defined checkpoints before they take effect. Where on the HITL spectrum to sit is a risk management decision that should be made before deployment.

Gartner's model tracing a technology's trajectory through five phases — Innovation Trigger, Peak of Inflated Expectations, Trough of Disillusionment, Slope of Enlightenment, and Plateau of Productivity — used to contextualise AI adoption maturity.

In natural language understanding, intent is the goal a user is trying to accomplish, extracted from the surface form of their text. 'Can I cancel my order?' and 'Cancel order #4421' express the same intent despite very different phrasing.

The totality of touchpoints — UI elements, voice interfaces, API endpoints, and agent handoffs — through which a user or system interacts with an AI product; a key unit of analysis in human-centred AI design.

A neural network trained on large volumes of text to predict and generate language, capable of translation, summarisation, reasoning, and code generation. Most of what practitioners call 'AI' in 2025-2026 is built on LLMs.

The narrowest tier of the AI-search optimisation stack: whether your content appears in the training data that shaped a model before any live query ran. Unlike GEO or AEO, LLMO outcomes are measured across months, not days.

A reliability engineering principle from post-war rocketry: system reliability is the product of its component reliabilities, not the average. Applied to AI agent chains, it explains why small errors compound multiplicatively.

An open protocol by Anthropic standardising how AI models connect to external tools, data sources, and services. Tools declare their capabilities and any MCP-compatible model can discover and use them.

The process by which an AI agent moves information from short-term working memory (the context window) into more durable storage for later retrieval. Sasha's CLS-inspired memory architecture applies this principle directly to agent design.

Directing different tasks to different AI models based on capability, cost, or latency. Sending complex reasoning to a frontier model and classification to a smaller one cuts costs significantly. Token economy thinking starts here.

An architecture where multiple autonomous agents interact, each pursuing subtasks toward a shared objective. Well-designed systems define protocols, role boundaries, and conflict resolution before agents are deployed.

A dynamic where a product becomes more valuable as more people use it, creating compounding growth advantages for early leaders. Sasha's piece on network effects and generative AI examines how these dynamics interact with AI-driven platforms.

The failure pattern where an AI pilot succeeds under controlled conditions but stalls before reaching production. Rarely technical; usually reflects missing governance, undefined ownership, or unscoped integration complexity.

AI systems that forecast outcomes, classify inputs, or detect patterns from historical data rather than generating new content. The distinction from generative AI matters: predictive systems are more auditable and easier to validate.

Designing inputs to AI models to reliably produce specific, high-quality outputs: instruction framing, context structuring, example selection, and constraints. See Google's prompt engineering whitepaper.

An architecture that augments a model's generation with real-time retrieval from an external knowledge base, grounding outputs in verifiable sources rather than training data alone. Retrieval quality determines generation quality.

A framework that interleaves reasoning with action steps: the model thinks, takes an action (like a tool call), observes the result, and continues reasoning. Foundational to most agentic AI because it makes decision-making auditable mid-task.

A structured set of policies, processes, and controls for identifying, assessing, and mitigating risks in AI systems; often references NIST AI RMF or ISO 31000 as baseline standards.

The layer of an AI agent's memory that stores general facts and conceptual understanding, updated gradually as episodic memories are processed and generalised. It holds what the agent knows, not what it experienced.

The cost and capacity framework governing AI usage, where tokens are the unit of both billing and context. Every input, output, and tool response consumes tokens. Sasha's article on AI budgets covers why most organisations get this wrong.

A methodology for capturing customers' needs, expectations, and complaints in their own words through interviews, surveys, and support data. See Sasha's piece on AI-powered VoC analysis.

The design activities for interfacing with and managing AI agent teams, where the human role has shifted from tracking work in progress to verifying whether completed output is correct. See VUX in practice.