Resources

AI terms, in plain English.

Short definitions you can quote in a meeting, each with why it matters to a business. No maths and no hype.

AI basics

AI agent

An AI agent is software that uses a language model to work through a task in several steps, deciding what to do next and using tools such as search, a database or an email system along the way.

A chatbot answers a question; an agent carries out a job. For example, an agent can read an incoming claim, look up the policy, draft a reply and place it in a queue for a person to approve. The more an agent is allowed to do on its own, the more its permissions and review steps matter.

What an AI Sprint builds
Context window

The context window is the amount of text a language model can take into account at one time, including the instructions, the conversation so far and any documents supplied with the question.

Anything outside the window is invisible to the model for that answer. This is why a system cannot simply be handed every company document at once: it has to select the passages that matter for each question. A long conversation can also push early instructions out of view, which is one reason assistants are tested on long exchanges as well as short ones.

Generative AI

Generative AI is the family of AI systems that produce new content, such as text, images, audio or code, in response to an instruction, rather than only sorting or scoring existing data.

For a business, the practical uses are mostly about language: drafting replies, summarising documents, answering questions and extracting details from forms. A support team might use it to draft a first response that a person then checks and sends. It is a different tool from the predictive models used for forecasting or risk scoring, and the two are often combined.

Where AI pays off first
Hallucination

A hallucination is an answer from an AI system that sounds confident and plausible but is wrong or made up, because the model generated likely-sounding text instead of drawing on a reliable source.

It happens most when a model is asked about something it was never shown, such as your prices or policies. Grounding the system in your own documents and testing it on questions it should decline reduce the problem; they do not remove it. That is why steps that carry real risk should keep a person approving them.

Grounding AI in your own data
Large language model (LLM)

A large language model is an AI model trained on very large amounts of text so that it can read and write language: answering questions, summarising, translating, classifying and drafting.

It is the engine inside most current AI assistants. On its own it knows only what it absorbed in training, so it has never seen your contracts, your product catalogue or last week’s price change. Useful business systems connect the model to your own information and put rules around what it may do.

Prompt

A prompt is the instruction and supporting information given to an AI model for a task. In a business system it usually includes standing rules written by the builders as well as the user’s question.

The standing part, often called the system prompt, is where a company sets tone, limits and what the assistant must never say. For example, a front-desk assistant’s prompt tells it to say it is an AI and to pass price questions to the team. Because small wording changes can alter behaviour, prompts should be kept under version control and re-tested when they change.

Token

A token is the unit of text a language model reads and writes, roughly a short word or a piece of a longer one. Model limits and usage fees are both counted in tokens.

Tokens are why running costs depend on how much text goes in and comes out of each request. An assistant that sends a whole manual with every question costs more per answer than one that sends only the relevant page. Estimating tokens per request during a prototype is how you learn what a feature will cost to run before launch.

Adding AI to your product

Building with AI

Conversational AI voice assistant

A conversational AI voice assistant is a system that listens to a person speaking, works out what they need with a language model, and replies in a natural spoken voice, on a website or over a phone line.

It suits callers who would rather talk than type, and it can work in more than one language. A sound setup tells the caller it is an AI, answers only from approved information, and passes anything else to a person. Voice should go live only after someone has rehearsed real conversations with it on a phone and a laptop.

How the AI front desk works
Embedding

An embedding is a list of numbers that represents the meaning of a piece of text, so that a computer can measure how similar two passages are even when they share no words.

Embeddings are what let a system match the question “can I return this?” to a paragraph headed “Refund policy”. A company’s documents are converted to embeddings once and stored; each incoming question is converted the same way and compared against them. The quality of that matching has a large effect on the quality of the final answer.

Fine-tuning

Fine-tuning is further training of an existing AI model on your own examples so that it adopts a particular style, format or specialised task.

It is useful when you need consistent output in a house format, or a narrow skill such as classifying your own ticket types. It is usually not the way to teach a model your facts, because facts change and retraining is slow; retrieval handles that better. Most business features work well without it, so it is worth asking why before paying for it.

Adding AI to your product
GraphRAG

GraphRAG is retrieval-augmented generation that draws its evidence from a knowledge graph, following the links between facts, instead of relying only on passages that look similar to the question.

It helps when the answer depends on how things relate. “Which suppliers are affected by this rule through any subsidiary?” has no single paragraph that answers it; the answer sits in the connections. Because the system can show the path it followed, GraphRAG is also easier to explain to a reviewer in a regulated setting.

RAG and GraphRAG in plain English
Grounding

Grounding means making an AI system answer from specific, trusted sources, such as your own documents or database, and show which source each answer came from.

A grounded assistant can be checked: a person can open the cited page and confirm the answer. It can also be told to say “I don’t have that” when the sources are silent. For a business this is the difference between an assistant that reflects your current policy and one that reflects the general internet.

Grounding AI in your own data
Knowledge graph

A knowledge graph stores information as things and the relationships between them, for example “this policy covers that product” or “this company owns that one”.

Where a document store holds pages, a graph holds connections, which makes questions about ownership, dependency and history much easier to answer. A logistics firm might use one to trace which shipments depend on a delayed part. Graphs are also a strong foundation for recommendations and for AI that must show which facts and links an answer was built from.

RAG and GraphRAG in plain English
Private LLM (VPC deployment)

A private LLM is a language model that runs inside an organization’s own network boundary, such as its virtual private cloud (VPC), so prompts and documents are not sent to a public AI service.

Organizations choose it when privacy rules, contracts or internal policy limit where data may go. It comes with access control, logging and usage policies, so staff can use generative AI on sensitive material. It costs more to run than a hosted service, so the decision should compare privacy needs, volume and response time.

What an AI Sprint builds
Retrieval-augmented generation (RAG)

Retrieval-augmented generation is a method where a system first searches your own documents for passages relevant to a question, then asks a language model to answer using only those passages.

It keeps answers tied to your current information without retraining the model each time something changes. An HR assistant built this way answers a leave question from the actual policy and links to the page. It lowers the rate of wrong answers but does not end them, so it should be paired with a test set and a clear “I don’t know” behaviour.

RAG and GraphRAG in plain English

Running AI safely

AI governance

AI governance is the set of decisions and records an organization keeps about its AI systems: what is deployed, who owns each one, what data it touches, how it is tested, and who approves changes.

It can start as a one-page inventory and grow as use grows. Without it, departments adopt tools on their own, data leaves in prompts, and nobody can answer a regulator’s or a board’s question about what is running. Good governance is a working practice inside the organization, not a certificate bought from a vendor.

Program and AI Sustain
AI readiness assessment

An AI readiness assessment is a short, structured review of an organization’s data, systems, people and governance that shows where AI is likely to pay off and what must be fixed first.

A useful one ends in decisions: a scored view of where you stand, a ranked list of opportunities, a costed roadmap and one first project scoped tightly enough to start. It should be written by people who could build the work and should say what it did not look at. Sometimes the honest result is “not yet”, with the reasons listed.

What an assessment should hand you
Evaluation set

An evaluation set is a fixed list of test questions or cases, with the expected behaviour for each, that an AI system is scored against every time something changes.

It turns “it seems to work” into a measurement. A good set includes questions the system should refuse or pass to a person, because that is where many failures hide. Writing it before the build starts gives everyone the same definition of done.

How a Sprint is measured
Guardrails

Guardrails are the rules and checks placed around an AI system that limit what it may discuss, what data it may use and what actions it may take.

Examples include refusing topics outside its job, blocking sensitive details such as card numbers, limiting how often it can be called, and requiring approval before it sends anything. A front-desk assistant that never quotes a final price is following a guardrail. Guardrails lower risk; they need testing like any other part of the system.

Seven questions before an AI answers customers
Human in the loop

Human in the loop means a person reviews, approves or corrects an AI system’s output at defined points before it has an effect on a customer, a payment or a record.

It is what makes early AI systems safe to use. An agent can draft a claim decision in seconds, and a claims handler still makes the call. The design question is where to place the review: on every item, on a sample, or only on cases the system flags as uncertain.

Where AI pays off first
Model drift

Model drift is the gradual change in how well an AI system performs after launch, caused by changes in the data it sees, the questions people ask, or the underlying model itself.

A system that passed its tests in spring can slip by autumn without anyone touching it: products change, a provider updates its model, customers start asking new things. The remedy is to re-run the evaluation set on a schedule and watch quality and cost over time. This ongoing checking is the main difference between maintaining AI and maintaining ordinary software.

Program and AI Sustain
Proof of concept vs production

A proof of concept shows that an idea can work under friendly conditions; a production system is one real users depend on, with security, monitoring, support and an owner.

The gap between the two is where many AI projects stall. The demo answered ten sample questions; production needs access control, logging, an evaluation set, integration with existing systems and sign-off from risk. Planning for that gap from the first week costs less than discovering it after the demo.

From pilot to working system

Software delivery

CI/CD (continuous integration and delivery)

CI/CD is the practice of automatically building, testing and releasing software each time a developer makes a change, instead of saving changes up for large, risky releases.

With it, a fix can reach users the same day, and a faulty change is caught by tests before customers see it. A team without it tends to release rarely and nervously. Setting it up in the first week of a project is one of the cheapest ways to keep delivery predictable later.

How we build products
Discovery

Discovery is a short phase at the start of a software project where the team agrees who the users are, what the first release must do, which technology fits, and what it will cost.

Its main job is to cut scope before money is spent building. A founder may arrive with thirty features and leave with the eight that have to work on day one, plus a dated plan. A good discovery produces a plan you could take to any competent team.

How a build runs
Fixed scope

Fixed scope is a way of contracting work where what will be delivered, the price and the timeline are agreed in writing before the work begins.

It moves delivery risk to the supplier: if the work takes longer than planned, the price does not rise. In return, the buyer agrees that new ideas are traded against existing items instead of simply added. It works best when a discovery phase has already made the scope specific.

Build, buy or embed
Minimum viable product (MVP)

A minimum viable product is the smallest version of a product that real customers can use and that teaches the company whether the idea is worth building further.

It is small on purpose, but what it includes must work properly. A marketplace MVP might support one category and one city, with real payments. The aim is to learn from actual use before committing the larger budget.

From idea to launched product
Staff augmentation

Staff augmentation is adding outside engineers to your own team for a period, working under your direction, in your tools and processes, while a supplier remains their employer.

It suits a company that has a roadmap and a lead who can set priorities but lacks a skill or enough senior hands. For example, a team might bring in a Kubernetes specialist for six months during a platform move. It works poorly when nobody on the client side can direct the work.

Embedded senior engineers
Technical debt

Technical debt is the extra work created when software is built the quick way instead of the sound way; like financial debt, it charges interest in the form of slower and riskier changes later.

Some debt is a sensible trade, such as a shortcut taken to hit a launch date. It becomes a problem when it is never repaid: every new feature takes longer and breaks something else. Signs include missing automated tests, a release process only one person understands, and estimates that keep growing.

Product development

Heard a term that isn’t here?

Ask us what it means and whether it matters for you. We aim to reply within one business day.

Book a callTalk to our team