The AI terms that come up in every deployment conversation.

AI systems that plan, take multi-step actions, and call tools or other systems on their own, rather than just responding to a single prompt.

The discipline of directing AI agents toward a goal, writing the spec, decomposing tasks across roles, and verifying output.

Independent verification that an AI system does what it's claimed to do, at an acceptable and measured error rate, before and after deployment.

Systematic skew in a model's outputs that disadvantages particular groups or scenarios, usually inherited from imbalances in its training data rather than deliberate design.

An AI assistant embedded directly inside existing software to help with a specific task, rather than a standalone tool used on the side.

Telling an audience that AI was used to produce a piece of content, a practice regulation and consumer expectation are both pushing toward, with evidence it doesn't always protect the trust it's meant to.

The mental exhaustion that comes from constantly having to verify AI-generated output, a workload that often outweighs the time the AI saved in the first place.

The policies, ownership structures, and controls an organisation puts in place to decide what AI is allowed to do, who is accountable for it, and how it's monitored.

The working understanding of what AI can and can't reliably do, enough to use it well, question its output, and know when not to trust it.

The ISO/IEC 42001 management system, spanning context, leadership, planning, support, operation, performance evaluation, and improvement, that an organisation runs to govern how it develops or uses AI.

The logging, monitoring, and tracing infrastructure that lets you see what an AI system is doing in production: latency, cost, errors, and output quality.

The coordination layer that decides which model, tool, or agent handles each step of a task, and in what order, when a system involves more than one of them.

Technically competent but generic AI-generated content, marketing copy, images, feed posts, that reads as unattended because no visible human judgment shaped it.

A security operations centre where AI agents handle first-pass alert triage, enrichment, and containment, so human analysts spend their time on the cases that need judgment.

The ability to reconstruct, end to end, why an AI system made a specific decision: what data it used, what it did, and who is accountable at each step.

Singapore's government-backed AI testing framework and toolkit, built by IMDA and stewarded by the AI Verify Foundation, for validating AI systems before deployment.

A company whose entire product and revenue depend on an AI capability, as opposed to an existing business that has added AI features to something it was already selling.

A named person's standing authority to stop a deployment the moment it produces a defect, with no requirement to justify the stop first.

The tendency for a human reviewer to defer to an automated system's recommendation, even when they have the information, or the obligation, to question it.

A measure of whether an organisation's guardrails, logging, and approval controls match the level of independent action a given AI system is being asked to take.

A standardised test used to measure and compare AI model or system performance against a defined task or dataset.

A system whose outputs can be observed and measured, but whose internal decision procedure cannot be inspected, reconstructed, or explained by anyone, including the people who built it.

The scope of systems, data, or damage a single compromised component can reach, and the thing every containment control is ultimately trying to shrink.

A predefined level of AI model capability, agreed in advance, beyond which a developer commits to additional safeguards or restricted release before deployment.

A technique where a model works through intermediate reasoning steps before giving a final answer, often improving accuracy on complex problems.

The executive accountable for how AI is deployed, governed, and measured across an organisation, distinct from a CTO's infrastructure remit or a CDO's data remit.

The communication channel an attacker, or a compromised autonomous agent, uses to receive instructions and send data back once it's inside a target system.

The economic principle that an actor should specialise in whatever it does at the lowest relative cost, not whatever it does best in absolute terms, which is why scarce AI compute keeps some work with humans even as models get better at everything.

AI systems that interpret and act on visual input, images or video, such as detecting defects, tracking progress, or reading a document's layout.

A distribution-free statistical method that wraps any model's output in a prediction set with a guaranteed error rate, instead of a single point estimate backed by the model's own, often overconfident, confidence score.

A training method, developed by Anthropic, where a model critiques and revises its own outputs against a written set of principles instead of relying only on human-labelled examples.

The maximum amount of text, measured in tokens, a model can consider at once when generating a response, including your prompt, any retrieved documents, and its own reply.

The open question of whether humans can reliably direct, constrain, or shut down an AI system whose capability exceeds their own.

The tendency for multiple AI models, or multiple human reviewers, to make the same mistake on the same input, which makes adding more checkers far less protective than the headcount suggests.

A virtual metadata layer that connects an organisation's existing databases and lets AI systems query across them without physically moving or duplicating the underlying data.

The unintended exposure of restricted data through an AI system, either by retrieval serving the wrong document to the wrong user, or by a fine-tuned model memorising and reproducing training data it shouldn't reveal.

A traceable record of where a piece of data originated, how it moved and transformed, and where it ended up, used to prove exactly what an AI system was trained or retrieved on.

A decentralised data architecture where each business domain owns and manages its own data as a product, under shared, company-wide governance standards.

An attack that corrupts the data an AI system learns from or retrieves, manipulating its outputs without ever touching the model itself.

The deliberate removal of management or coordination layers from an organisation's hierarchy, reducing how many approval steps a decision has to pass through.

A continuously updated virtual model of a physical asset, process, or system, used to simulate, monitor, and predict its real-world behaviour.

AI models that run on hardware physically located where the work happens, such as a factory line, a vehicle or a hospital ward, rather than in a remote data centre.

A numerical representation of text (or images, or audio) that captures its meaning, allowing a computer to compare how similar two pieces of content are.

The European Union's risk-tiered AI regulation, which bans a narrow set of practices outright and imposes graduated obligations, including trained human oversight, on higher-risk uses.

A category of technology that lets less-experienced workers perform specialist-level work, raising output while eroding the market's willingness to pay a premium for scarce expertise.

The degree to which a person can understand why an AI system produced a particular output, rather than treating it as an unreviewable black box.

The Monetary Authority of Singapore's four principles (Fairness, Ethics, Accountability, and Transparency) for the responsible use of AI and data analytics in financial services.

Further training an existing model on a narrower, task-specific dataset so it performs better on your particular use case, tone, or domain.

The operating discipline for forecasting, monitoring, and controlling cloud and AI compute spend so cost scales predictably with usage instead of surprising finance after the fact.

A large, general-purpose model trained on broad data, meant to be adapted to many downstream tasks rather than built for one narrow job.

A general-purpose AI model at or near the current limit of capability, and therefore the primary subject of international AI safety summits and voluntary safety commitments.

Anchoring a model's response to verifiable source material, so it answers from real, checkable information rather than from memory alone.

The technical and procedural checks placed around an AI system to constrain what it can say or do, catching unsafe, incorrect, or out-of-policy outputs before they reach a user.

When an AI model generates a confident, plausible-sounding answer that is factually wrong or entirely fabricated.

A design pattern where a person reviews, approves, or can override an AI system's output before it takes effect, rather than letting the system act fully autonomously.

The process of running a trained AI model to generate an output from a new input, as opposed to training, which is how the model learned in the first place.

Lisanne Bainbridge's 1983 finding that automating the routine part of a task leaves the hardest, least-practiced part to the human operator, right when they're least prepared for it.

The uneven, unpredictable boundary between tasks an AI model handles at or above expert level and tasks that look similarly hard but fall just outside it.

The regulated process of verifying a client's identity, source of funds, and risk profile before and during a banking or financial relationship.

A centralised cloud data platform, such as Snowflake or Databricks, that consolidates an organisation's raw and structured data into a single repository, combining data-lake flexibility with data-warehouse performance.

A model trained on vast amounts of text to predict and generate language, the technology underneath most modern AI chat and writing tools.

The time between sending a request to an AI system and receiving its response, a major factor in whether a use case feels usable.

A build, measure, learn discipline that tests the smallest version of an idea against real evidence before committing to it at scale.

A security principle that grants a system only the minimum access it needs to complete a specific task, and nothing more, so a compromised or misbehaving process can't reach beyond its assigned job.

When "a human reviewed it" is used to redirect accountability for an AI system's error away from how the system was designed and toward whoever clicked approve.

A technique for scoring AI outputs by having a second language model read the response and grade it against a written rubric, in place of exhaustive human review.

A model run on infrastructure you control, on-premise or in your own cloud environment, rather than called through a third-party provider's API.

The research discipline that reverse-engineers a neural network's internal computations into human-understandable components, rather than only studying what goes in and what comes out.

A structured knowledge base of real-world adversary tactics and techniques against AI and machine learning systems, the AI-security counterpart to MITRE ATT&CK.

A model architecture that splits its parameters into many specialised sub-networks and activates only a few of them per request, giving large capacity at a fraction of the compute cost.

The engineering discipline and pipeline for testing, deploying, monitoring, and updating AI models in production, the equivalent of DevOps for machine learning.

An open standard that lets AI models connect to external tools, data sources, and systems through a common interface, instead of a custom integration for every connection.

A technique that trains a smaller, cheaper model to reproduce a larger model's behaviour on a specific task, cutting inference cost without retraining from scratch.

The gradual decline in an AI system's accuracy or relevance over time, as the real-world data it encounters diverges from the data it was trained or tuned on.

An attack that reconstructs sensitive information from a model's training data by repeatedly querying the model and analysing what its outputs reveal.

AI systems that can understand and generate across more than one type of input at once, such as text, images, audio, and video, rather than being limited to text alone.

An AI model whose trained parameters are published for anyone to download and run, rather than kept behind a provider's hosted API.

The hardware and software that directly monitors or controls physical equipment and industrial processes, distinct from the IT systems that manage data and applications.

A commercial model where a provider is paid for a measured result, an issue resolved, a lead qualified, rather than for hours worked or a licence held.

A ranked catalogue of the security risks unique to autonomous AI agents, from goal hijacking and tool misuse to insecure inter-agent communication and cascading blast-radius failures.

A ranked catalogue of the most critical security risks specific to large language model applications, from prompt injection through to uncontrolled resource consumption.

One of the numerical weights a model adjusts during training, with the total count used as a rough proxy for a model's capacity and for the hardware needed to run it.

Singapore's core data protection law, governing how organisations collect, use, disclose, and secure personal data, with real penalties for breaches.

A pattern where a model responds to a user's fact-check or pushback by escalating persuasion instead of disclosing its limitation, restating its original answer with more confidence and supporting evidence.

The multi-decade divergence between growth in worker output per hour and growth in worker pay, tracked for the US since 1979 by the Economic Policy Institute.

The practice of structuring instructions, context, and examples given to an AI model to reliably get the output you want.

An attack where malicious text, in a user message or a retrieved document, is crafted to override an AI system's instructions and make it do something it shouldn't.

A small, time-boxed test built to show an AI approach can work at all, before anyone commits to building it for real production use.

A training technique that fine-tunes a model using human rankings of its outputs, rewarding the responses people prefer and penalising the ones they don't.

A technique that retrieves relevant documents from your own data before the model answers, so responses are grounded in real, current information instead of the model's training data alone.

A prioritisation method that scores each candidate project on Reach, Impact, Confidence, and Effort, so teams rank initiatives by a number instead of by opinion.

A permissions model that grants data or system access based on a user's role rather than their individual identity, so what an AI system can retrieve or act on depends on who is asking.

The ability to revert an AI system to a previous, known-good version when a model update, data change, or tool failure degrades its output.

An isolated execution environment used to run untrusted or high-risk code, including AI agent actions, without giving it access to production systems or the open internet.

A cost and latency optimisation that reuses a stored AI response for a new query when its meaning matches a previous one, not just its exact wording.

AI tools employees adopt and use for work without formal approval, visibility, or oversight from IT, security, or governance functions.

Michael Spence's 1973 account of why a costly, hard-to-fake credential works as proof of quality when the quality itself can't be observed directly.

A workflow where a written specification, not a prompt, is the source of truth an AI coding agent builds against, with the generated code treated as a build artifact.

The documented record, required by ISO/IEC 42001 (and ISO/IEC 27001), of every reference control an organisation includes or excludes from its management system, with a justification for each decision.

A hypothetical AI system whose general cognitive capability substantially exceeds the best human minds across essentially every domain, not just one narrow task.

Artificially generated data used to train or test AI models when real data is scarce, sensitive, or too imbalanced to use directly.

A risk that emerges from the interaction of many individually reasonable decisions across an entire system, invisible to any single actor within it and therefore unmanaged by any of them.

The length of a task, measured by how long a skilled human takes to complete it, that an AI agent can complete autonomously at a defined reliability threshold.

The economic model that analyses AI exposure at the level of individual work tasks rather than whole job titles, since most jobs bundle both automatable and non-automatable tasks.

The unit of text a model processes, roughly a word or part of a word, used to measure context length, usage, and cost.

The capability that lets an AI model call an external system, such as a database, an API, or a calendar, and act on the result, instead of only producing text.

A commercial model where the bill scales with consumption, tokens processed, minutes generated, resolutions completed, rather than with the number of people who have a licence.

A database built to store embeddings and quickly find the ones most similar to a given query, the storage layer underneath most retrieval-augmented systems.

The situation where switching away from an AI provider becomes so costly or disruptive that you're exposed to their pricing, policy, and downtime decisions with no real alternative.

Building software by describing what you want in natural language and letting an AI model generate the code, rather than writing it by hand line by line.

AI systems that understand spoken input and respond in natural speech, increasingly in real time and across both text and voice within the same conversation.

A software flaw that's exploited before the vendor knows it exists, so no patch is available and standard defenses have nothing to detect.