Glossary

Plain definitions of the terms and acronyms that come up in conversations about AI, software modernisation and cloud security.

Technology conversations tend to be dense with acronyms. These definitions are deliberately short, and aim to give enough understanding to follow the discussion and ask useful questions.

Modernisation and architecture

How software is structured, and how an ageing product gets updated.

API Application programming interface
A defined way for one piece of software to talk to another. A clean API lets the user interface and the backend change independently.
B2B Business to business
Selling to other businesses, as opposed to consumers.
Big-bang rewrite
Replacing a whole system in one go. Nothing reaches customers until the end, which makes it expensive and risky for a business that depends on the product. See Modernising legacy SaaS without a big-bang rewrite.
CI/CD Continuous integration and continuous delivery
Automation that builds, tests and deploys code each time it changes, so releases are frequent and repeatable. See Moving a 20-year-old SaaS product onto GitHub.
Feature flag
A switch that turns a feature on or off without a new deployment. Useful for releasing changes to a few customers first.
Key person risk
The risk that essential knowledge of a system sits with one or two people and leaves when they do.
Legacy system
An older system the business still depends on. It's typically built on ageing technology and has become slow or risky to change.
Lift and shift
Moving a system to the cloud with as few changes as possible. Quick to do, though the system keeps its existing design and limitations.
Microservices
An application split into small services that are deployed independently. Teams can work separately, at the cost of more operational complexity.
Monolith
An application built and deployed as a single unit. Simple to run, and a sensible choice for many products.
Multi-tenancy
One running instance of an application serves many customers, with each customer's data kept separate.
Proof of concept
A small, quick build that tests whether an idea is feasible before committing to a full implementation. Often shortened to PoC.
Pull request
A proposed code change, presented for review before it's merged into the main codebase.
PWA Progressive web app
A web application that can be installed on a device and work offline, built with standard web technology in place of a separate native app.
Refactoring
Improving the internal structure of code without changing what it does. Usually done in small steps, backed by tests.
SaaS Software as a service
Software hosted by the supplier and accessed over the internet, usually on subscription.
SPA Single-page application
A web application that loads once and then updates the page in the browser, fetching data from an API as needed.
Strangler fig pattern
A way of modernising a system gradually. A new system grows around the old one, taking over functionality in slices, until the old one can be retired. Martin Fowler named it after a fig that grows around a host tree. See Modernising legacy SaaS without a big-bang rewrite.
Technical debt
The accumulated cost of shortcuts and ageing decisions in a codebase. It shows up as changes that take longer and break more often.
Technical due diligence
A review of the technology behind a business, carried out for a buyer or investor before a transaction.
Vendor lock-in
Dependence on one supplier's technology to the point where moving away would be costly or disruptive.

AI and machine learning

The foundations that sit beneath every AI product.

AGI Artificial general intelligence
A hypothetical AI that matches or exceeds people across most intellectual tasks. There's no agreed definition or test for it.
AI Artificial intelligence
An umbrella term for software that performs tasks normally associated with human intelligence. It covers machine learning, large language models and much else.
AUC Area under the curve
A single number summarising how well a model separates one class from another. ROC-AUC and PR-AUC are the two common versions, and PR-AUC tends to be more informative when the outcome of interest is rare.
Bias
Systematic errors in a model's output that disadvantage particular groups, usually inherited from the training data.
Classification
Predicting which category something belongs to, such as whether a transaction is fraudulent.
Computer vision
The field of AI concerned with interpreting images and video.
Deep learning
Machine learning that uses neural networks with many layers. It's behind most recent progress in language, images and speech.
Explainability
The ability to understand why a model produced a particular output. It matters most where decisions affect people.
Feature engineering
Turning raw data into inputs a model can learn from, such as deriving the day of the week from a timestamp.
GPU Graphics processing unit
The type of chip used to train and run most AI models. Supply and cost of GPUs shape the price of AI services.
Gradient boosting
A machine learning technique that combines many small decision trees, each correcting the errors of the ones before. It performs strongly on tabular business data. XGBoost is a well-known implementation.
Inference
Running a trained model to get an output. Most of the ongoing cost of an AI feature is inference.
Machine learning
Software that learns patterns from data, so the rules are inferred from examples and don't have to be written by hand. Often shortened to ML.
MLOps Machine learning operations
The practices for deploying, monitoring and maintaining machine learning models in production.
Model
The result of training. It takes an input, such as a question or a row of data, and produces an output or prediction.
Model drift
A model's performance declines over time as real-world data moves away from the data it was trained on.
Neural network
A model built from layers of simple connected units. The strength of each connection, known as a weight, is learned from data.
NLP Natural language processing
The field of AI concerned with understanding and producing human language.
OCR Optical character recognition
Extracting text from images or scanned documents.
Overfitting
A model memorises its training data and then performs poorly on anything new.
Precision and recall
Two measures of a classifier. Precision is the share of flagged items that were correct. Recall is the share of items that should have been flagged and were. The F1 score combines the two.
Random forest
A machine learning technique that averages the predictions of many decision trees, each trained on a different sample of the data.
Regression
Predicting a number, such as a price or a delivery time.
Reinforcement learning
Learning by trial and error, guided by rewards for good outcomes.
Supervised learning
Learning from examples where the correct answer is already known, such as past quotes marked as won or lost.
Training
The process of fitting a model to data. For large models it's expensive and done by the model provider.
Training data
The examples a model learns from. Its quality and coverage largely decide how well the model performs.
Unsupervised learning
Finding structure in data that has no labels, such as grouping similar customers together.

Generative AI and agents

Large language models, and the systems built around them.

Agentic AI
Systems built from one or more AI agents that plan and carry out multi-step work with limited supervision. Agents can be highly capable, though they start with no knowledge of the business and carry no accountability, which is why oversight still matters. See Why agentic AI is about leverage.
AI agent
An LLM that works towards a goal over several steps, choosing which tools to use and reacting to the results.
AI-washing
Describing a product as AI-powered when the AI involved is minimal or absent.
Benchmark
A standardised test for comparing models. Scores are a useful first filter, though they tend not to predict performance on a specific task in production.
Coding agent
An AI agent that reads, writes and tests code inside a codebase. See Using AI on an existing codebase without handing over ownership.
Context engineering
Deciding what information a model is given for a task, and in what form. It often matters more than the wording of the prompt.
Context window
The maximum amount of text a model can consider at once, measured in tokens. It includes the instructions, any supplied documents and the model's own output.
Embedding
A list of numbers that represents the meaning of a piece of text, so that similar content sits close together. Embeddings make search by meaning possible.
Eval
A test suite built for one application, measuring whether the AI does that particular job well. Short for evaluation.
Fine-tuning
Further training of an existing model on specific examples to adjust its behaviour.
Foundation model
A large model trained on broad data that can be adapted to many different tasks.
Frontier model
One of the most capable models available at a given time. The label moves on with each new release.
Generative AI
AI that produces new content, such as text, images, code or audio.
Grounding
Tying a model's answer to supplied sources, so claims can be checked against them.
Guardrails
Checks placed around a model that constrain what goes in and what comes out.
Hallucination
Output that is stated confidently but is false or unsupported.
Human-in-the-loop
A person reviews or approves the AI's work at defined points. Accountability stays with that person.
Knowledge cutoff
The date after which a model has no training data. Anything more recent has to be supplied to it.
LLM Large language model
A model trained on very large amounts of text that works by predicting what comes next. LLMs power products such as ChatGPT, Claude and Gemini.
MCP Model Context Protocol
An open standard, introduced by Anthropic in 2024, for connecting AI models to external tools and data sources in a consistent way.
Multi-agent system
Several agents with distinct roles working on one task, coordinated by software or by another agent.
Multimodal
Able to work with more than one kind of content, such as text, images and audio.
Open-weight model
A model whose trained weights are published, so it can be run on an organisation's own infrastructure.
Prompt
The instructions and information given to a model for a task.
Prompt engineering
Writing and refining prompts so that a model produces reliable output.
Prompt injection
An attack where content the model reads, such as a web page or an email, contains instructions that the model then follows.
RAG Retrieval-augmented generation
Relevant documents are retrieved and passed to the model along with the question, so the answer draws on them. A common way to make a model useful with private or recent information.
Reasoning model
A model that works through a problem in steps before answering. It's slower and costs more, and tends to do better on complex tasks.
Structured output
Constraining a model to reply in a defined format, such as JSON, so other software can use the result.
System prompt
Instructions set by the developer that frame how a model behaves throughout a conversation.
Temperature
A setting that controls how varied a model's output is. Lower values give more predictable answers.
Thin wrapper
A product feature that is little more than a prompt over a third-party model. It's quick to build and can be a sound choice, though competitors can copy it just as quickly.
Token
The unit of text a model reads and writes, roughly three quarters of an English word. Usage is typically priced per token.
Tool use
The model asks the surrounding software to run something, such as a search or a database query, and uses the result. Also called function calling.
Transformer
The neural network design behind modern LLMs, introduced by Google researchers in 2017.
Vector database
A database that stores embeddings and finds the ones closest to a query.
Vibe coding
Building software by describing it to an AI and accepting the output with little review. The term was coined by Andrej Karpathy in 2025.

Identity and access

Who can sign in, and what they're allowed to do afterwards.

ABAC Attribute-based access control
Access decisions are based on attributes of the user, the resource or the request, such as department, a tag on the data or the time of day. Often used alongside RBAC for finer control.
Authentication
Proving identity, usually with a password, a device or a biometric. Often shortened to AuthN.
Authorisation
Deciding what an authenticated identity is allowed to do. Often shortened to AuthZ.
Emergency access account
Also called a break-glass account. A highly privileged account kept for emergencies, such as an identity provider outage or a lockout caused by a misconfigured policy. It's usually exempt from some sign-in policies and closely monitored.
IAM Identity and access management
The policies and tools that control who can sign in to a system and what they're allowed to do once inside.
IdP Identity provider
The system that holds user accounts and confirms who someone is on behalf of other applications. Entra ID, Okta and Google are common examples.
JIT access Just-in-time access
Elevated permissions are granted only when needed and expire automatically afterwards.
JWT JSON Web Token
A compact, signed token carrying claims about a user or system, such as identity and expiry time. Commonly used as an access token for APIs.
Least privilege
Giving each person or system only the access needed for the task. It limits the damage a compromised account can do.
MFA Multi-factor authentication
Signing in needs two or more kinds of proof, such as a password plus a code from a phone app.
OAuth 2.0
A standard for delegated authorisation. It lets an application access an API on a user's behalf using a token, without handling the user's password. See OAuth explained.
OIDC OpenID Connect
A sign-in layer built on OAuth 2.0 that tells an application who the user is. See OAuth explained.
PAM Privileged access management
Controls around powerful accounts, such as administrators. Typically includes approval, time limits and session recording.
RBAC Role-based access control
Access is granted through roles, each a named bundle of permissions. People and systems are given roles, which keeps permissions consistent and easier to review. In Azure, a role is assigned to an identity at a particular scope.
SAML Security Assertion Markup Language
An older XML-based standard for single sign-on, still widely used for enterprise applications. See SAML versus OAuth.
SCIM System for Cross-domain Identity Management
A standard for automatically creating, updating and removing user accounts in other applications when they change in the identity provider.
SSO Single sign-on
One sign-in gives access to several applications, usually through a central identity provider. See SAML versus OAuth.
Zero trust
A security model built on three ideas. Verify every request explicitly, grant least-privilege access, and assume a breach may already have happened.

Azure and Microsoft

Terms specific to Microsoft Azure and its identity service, Entra ID.

App registration
The record that makes an application known to Entra ID, so it can sign users in or call protected APIs.
Azure Bastion
A managed service for connecting to virtual machines over RDP or SSH, without giving those machines public IP addresses.
Azure Policy
A service that enforces rules on Azure resources and reports on compliance. Examples include restricting which regions can be used or requiring encryption.
Built-in roles
Roles that Azure provides ready-made. Owner has full access and can grant access to others, Contributor can manage resources but can't grant access, and Reader can only view. Many narrower roles exist for specific services.
Conditional Access
Entra ID policies that allow, block or add conditions to a sign-in, based on signals such as the user, device, location and assessed risk. A typical policy requires MFA for administrators.
Defender for Cloud
Microsoft's service for assessing the security posture of cloud resources and protecting running workloads. It produces a Secure Score and a list of recommendations.
Entra ID
Microsoft's cloud identity service, known as Azure Active Directory (Azure AD) until 2023. It holds users, groups and application identities, and handles sign-in for Azure and Microsoft 365.
Entra roles
Roles that control the directory itself, such as managing users, groups and applications. They're separate from Azure roles, which control resources like databases and virtual machines. The two systems are easy to confuse.
Global Administrator
The most powerful Entra role, with access to almost every administrative setting in the tenant. Microsoft's guidance is to keep the number of people holding it very small.
Key Vault
An Azure service for storing secrets, encryption keys and certificates. Access is controlled and logged, which keeps credentials out of code and configuration files.
Landing zone
A pre-built Azure environment with identity, networking, policy and monitoring already in place, ready for applications to be deployed into.
Managed identity
An identity that Azure creates and looks after on behalf of a resource, such as a web app or virtual machine. The application can reach other services without any credentials stored in code or configuration.
Management group
A container that sits above subscriptions. Access and policy applied to a management group are inherited by every subscription beneath it.
Microsoft Sentinel
Microsoft's cloud-based SIEM and SOAR service. It collects security logs, detects threats and can automate parts of the response.
NSG Network security group
A set of rules that allow or deny network traffic to and from resources in an Azure virtual network.
PIM Privileged Identity Management
An Entra ID feature for just-in-time access. Administrators are made eligible for a role and activate it for a limited period when needed, optionally with approval.
Private endpoint
A private network address inside a virtual network that connects to an Azure service, such as a database or storage account. Traffic to the service stays off the public internet.
Resource group
A container for Azure resources that share a lifecycle, such as the parts of one application. Access can be granted at this level.
Role assignment
The way access is granted in Azure RBAC. It joins three things together, a security principal (who), a role definition (what) and a scope (where).
SAS Shared access signature
A token, usually added to a URL, that grants limited and time-bound access to data in Azure Storage. Anyone holding the token can use it, so short expiry times matter.
Scope
The level at which a role assignment or policy applies. Azure has four levels, which are management group, subscription, resource group and individual resource. Access granted at one level is inherited by the levels below.
Security principal
Anything that can be given access in Azure. Covers users, groups, service principals and managed identities.
Service principal
The identity an application or automated process uses to sign in and access resources. It needs a secret or certificate, which has to be stored safely and rotated.
Subscription
A container for Azure resources that is also a billing unit. Separate subscriptions are often used to keep production apart from development and test.
Tenant
A dedicated instance of Entra ID that represents one organisation. Users, groups and app registrations live in the tenant, and Azure subscriptions trust it for sign-in.
VNet Virtual network
A private network inside Azure, where resources can talk to each other in isolation from the public internet.

Protecting systems and data

The controls that keep data and services safe.

Attack surface
Every point where an attacker could try to get in, including public endpoints, user accounts, APIs and third-party dependencies.
CMK Customer-managed key
An encryption key that the customer controls, usually held in a service like Key Vault, in place of one managed by the cloud provider. It gives more control, along with responsibility for looking after the key.
CSPM Cloud security posture management
Tools that continuously check cloud configuration against good practice and flag weaknesses. Defender for Cloud is Microsoft's offering.
DDoS Distributed denial of service
An attack that floods a service with traffic from many sources so that genuine users can't reach it.
Defence in depth
Layering several independent protections, so that one failing doesn't expose the whole system.
Encryption at rest
Data is encrypted while stored, on disk or in a database, so a stolen disk or backup can't be read without the key.
Encryption in transit
Data is encrypted while moving between systems, usually with TLS.
HSM Hardware security module
Tamper-resistant hardware that stores cryptographic keys and performs operations with them, so the keys never leave the device.
mTLS Mutual TLS
Both sides of a connection present a certificate, so each can verify the other. Common between internal services.
Secret
A credential that grants access, such as a password, API key or connection string. Secrets belong in a dedicated store like Key Vault, and never in source code.
Secret rotation
Replacing a secret or key on a regular schedule, or immediately after a suspected leak. It limits how long a stolen one stays useful.
Security posture
The overall state of an organisation's security, taking in configuration, controls, and how well problems are detected and handled.
Shared responsibility model
The split of security duties between a cloud provider and its customer. The provider secures the underlying platform. The customer remains responsible for its own data, identities and configuration, and the exact split varies by service.
TLS Transport Layer Security
The protocol that encrypts traffic between two systems. It's the S in HTTPS.
WAF Web application firewall
Sits in front of a web application and filters out common attacks, such as SQL injection and cross-site scripting.

Finding and responding to problems

Testing, monitoring and what happens when something goes wrong.

Audit log
A record of who did what and when. Essential for investigating incidents and often required for compliance.
CVE Common Vulnerabilities and Exposures
A public catalogue that gives each known vulnerability a unique identifier, such as CVE-2021-44228.
CVSS Common Vulnerability Scoring System
A score from 0 to 10 indicating how severe a vulnerability is.
DAST Dynamic application security testing
Probes a running application from the outside for security flaws.
EDR Endpoint detection and response
Software on laptops and servers that detects suspicious behaviour and helps contain it.
Incident response
The planned process for handling a security incident, from detection and containment through to recovery and review.
OWASP Open Worldwide Application Security Project
A non-profit best known for the OWASP Top 10, a list of the most common web application security risks.
Penetration test
An authorised, simulated attack carried out by security specialists to find weaknesses before a real attacker does. Often shortened to pen test.
SAST Static application security testing
Analyses source code for security flaws without running it.
SBOM Software bill of materials
A list of every component and dependency in a piece of software. Used to check exposure when a new vulnerability is announced.
SIEM Security information and event management
Collects logs from across an organisation's systems and analyses them to detect threats.
SOAR Security orchestration, automation and response
Automates the routine steps of responding to a security alert.
SOC Security operations centre
The team that monitors for security incidents and responds to them. Unrelated to SOC 2.
Threat modelling
A structured way of working out what could go wrong with a system, who might attack it and how, before deciding which protections matter most.
Vulnerability scanning
Automated checking of systems and software for known weaknesses. Broader and cheaper than a penetration test, with less depth.
XDR Extended detection and response
Builds on EDR by combining signals from endpoints, email, identity and cloud services.

Standards and compliance

The bodies, standards and regulations that customers and auditors refer to.

AI governance
The policies, roles and controls an organisation uses to oversee how AI is chosen, built and used.
CIS Benchmarks
Configuration guidelines from the Center for Internet Security for hardening common systems and cloud platforms. Many security tools can check against them.
Cyber Essentials
A UK government-backed certification covering five basic technical controls. These are firewalls, secure configuration, user access control, malware protection and security updates. Cyber Essentials Plus adds independent technical testing.
DPIA Data protection impact assessment
A documented assessment of privacy risk, required under GDPR before processing that is likely to pose a high risk to individuals.
EU AI Act
The EU's regulation on AI, in force since August 2024 with obligations phased in over the following years. It sorts AI systems into risk tiers, with the strictest duties on high-risk uses. It can apply to UK companies that sell into the EU. See AI compliance for SMEs.
GDPR General Data Protection Regulation
The EU law governing personal data. The UK has its own version, the UK GDPR, which sits alongside the Data Protection Act 2018.
IEC International Electrotechnical Commission
Publishes international standards for electrical and electronic technology, often jointly with ISO.
ISO International Organization for Standardization
Publishes international standards across many industries.
ISO/IEC 27001
The international standard for an information security management system, or ISMS. Certification involves an audit by an accredited body and is often requested by enterprise customers.
ISO/IEC 42001
An international standard, published in 2023, for an AI management system. It covers how an organisation governs the development and use of AI. See AI compliance for SMEs.
NCSC National Cyber Security Centre
The UK's technical authority on cyber security and part of GCHQ. It publishes practical guidance for organisations of all sizes.
NIST National Institute of Standards and Technology
A US government body whose publications are widely used as reference points. They include the Cybersecurity Framework and the AI Risk Management Framework.
NIST AI RMF AI Risk Management Framework
A voluntary framework from NIST, published in 2023, for managing AI risk. It's organised around four functions, which are govern, map, measure and manage.
PCI DSS Payment Card Industry Data Security Standard
Security requirements for any organisation that stores, processes or transmits payment card data.
Responsible AI
A general term for developing and using AI in ways that are fair, transparent, safe and accountable. Several vendors publish their own principles under this name.
SOC 2
An independent audit report on a service provider's controls, based on criteria from the American Institute of Certified Public Accountants (AICPA). A Type 1 report covers one point in time and a Type 2 report covers a period. Often requested by US customers.