Skip to content

Glossary

The terms you will meet, in plain English.

This glossary explains the terms founders and growing businesses meet when they commission software or AI agents, in plain English, with why each one matters to the decision in front of you. Each entry links to where we cover it in depth.

Terms

Acceptance criteria

Acceptance criteria are the written, testable conditions a piece of software must meet before the work counts as done.

They turn "it should handle refunds" into checks anyone can run, such as "a refund over the original amount is refused". They are what a fixed price is measured against, and what a launch warranty covers. Ours: any defect against the agreed acceptance criteria, found in the 30 days after release, is fixed at no cost.

How fixed pricing works

Agent Readiness Review

An Agent Readiness Review is a two-week, fixed-fee review of where AI agents can safely take on work across a business, ending in a ranked shortlist and a costed roadmap.

It looks at how work actually moves, the data each task needs and the risks it carries, so the first agent is chosen for value and safety rather than novelty.

The Agent Readiness Review

Agentic workflow

An agentic workflow is a process in which an AI agent carries out a sequence of steps towards a goal, choosing its next action from what it finds, rather than following one fixed script.

The value is that it can handle variety a rigid automation cannot. The risk is the same thing, which is why each step it may take on its own, and each step that waits for a person, should be agreed in writing and enforced by access controls.

Custom AI agents

AI agent

An AI agent is software that uses a language model to carry out a defined task, reading from approved sources and acting in approved tools within set bounds.

A good first agent does one frequent, rule-bound task a person already does by hand, such as matching invoices to purchase orders. Payments, deletions and changes to customer records should always wait for a person.

What a custom AI agent costs

Autonomy levels

Autonomy levels describe how much an AI agent may do without a person: observe, suggest, act with approval, act and report, or act without sign-off.

An agent should start at the narrowest level that is useful and move up one level at a time, only when its track record and your agreement justify it.

Keeping agents away from production

Discovery phase

A discovery phase is the paid stage before a build in which the problem, users, risks and scope are worked out well enough to plan and price the build.

You need one when you do not yet know what to build. When you already know the outcome you want, a free scoping that ends in a written plan, a fixed price and the risks is usually enough.

What a discovery phase is for

Fixed price

A fixed price is an amount agreed in writing for a defined piece of work before it starts, so the buyer knows the cost and the supplier carries the risk of estimating.

It only protects you if the scope is written down and changes are re-scoped and re-priced in writing before they are built. Compare it with a day rate, where you pay for time whatever it produces.

Our pricing

Guardrails

Guardrails are the limits placed around an AI system, such as what it may read, what it may do, and when it must stop and ask a person.

Guardrails written only into an agent’s instructions can be talked around. The ones that hold are enforced outside the model, by access controls, separate credentials and approval steps.

How we control agents

Hallucination

A hallucination is an answer from a language model that sounds confident but is not supported by its sources or by fact.

It cannot be removed entirely, so it has to be designed around: agents answer from approved sources, cite them, and send any answer their sources do not support to a person instead of sending it on.

How our agents handle unsupported answers

Human in the loop

Human in the loop means a person reviews or approves an AI system’s work at defined points before it takes effect.

The useful question is where the loop sits. Ours puts a person on payments, deletions, changes to customer records and every release to production, at every level of autonomy.

Agents and controls

Internal tool

An internal tool is software built for a business’s own staff, such as an admin panel, an approval workflow or a dashboard, rather than for its customers.

Internal tools often start as a spreadsheet that became critical. When several of them need to share the same records, an operating platform is usually the better answer than another standalone tool.

Operating platforms

IP assignment

An IP assignment is a written, signed transfer of the intellectual property in work, such as code, from the person or firm who created it to the client.

Under UK law, code written by a freelancer or an agency belongs to them unless it is assigned to you in writing. Without an assignment you usually hold only a licence to use it, which can complicate a raise or a sale.

Owning your code with a software partner

MVP (minimum viable product)

An MVP is the smallest version of a product that real users can use for its main purpose, built to learn whether it is worth building further.

Minimum is about scope, not quality: the parts it has should be production-safe. A prototype shows an idea; an MVP has to hold real customers and their data.

What an MVP costs in the UK

No-code

No-code tools let people build software by assembling visual blocks or describing what they want, without writing code by hand.

They are a fast way to test an idea or run an internal tool. The app usually runs on the vendor’s platform, so you own the configuration rather than the engine, and moving off it can mean rebuilding the logic.

No-code or custom build

Operating platform

An operating platform is a bespoke internal system that brings a business’s clients, pipeline, finance and documents into one set of records, often with agents working across them.

It replaces a patchwork of spreadsheets and disconnected tools, and it is built around your workflow rather than a vendor’s data model, hosted in your own accounts.

Operating platforms

Penetration test

A penetration test is an authorised, simulated attack on a system by testers, to find weaknesses an attacker could exploit, usually reported against a recognised methodology.

Customers, insurers and regulators sometimes ask for one from an accredited firm: in the UK usually CREST, or CHECK for government work. It reports weaknesses; it does not normally fix them, and it is different from Cyber Essentials certification.

Penetration test or security review

Per-seat pricing

Per-seat pricing charges a fee for every user of a piece of software, usually every month.

It is cheap with a few users and grows with headcount, including people who only need to read or approve. It is worth comparing over three years against software you own outright.

Bespoke or off-the-shelf software

Production

Production is the live environment real users and real data run on, as opposed to development or staging environments used for building and testing.

Only a named engineer should be able to release to it. The agents that write your code should have no access to it at all, by construction rather than by instruction.

Keeping agents away from production

Prompt injection

Prompt injection is an attack in which text a user types, or a document an AI system reads, steers the model into ignoring its instructions, revealing data or taking actions it should not.

It cannot be fixed by better instructions to the model alone. The defence is to limit what the AI feature can read and do with access controls, and to make anything consequential wait for a person.

Security review for AI-built apps

Proof of concept

A proof of concept is a small piece of work that shows whether an idea is technically possible and useful before a larger investment is made.

Our version is the proof sprint: one bounded feature, workflow or working prototype, live in your own accounts in 1 to 2 weeks, from £3,000.

The proof sprint

RAG (retrieval-augmented generation)

Retrieval-augmented generation is a technique in which an AI system first finds relevant passages in approved documents, then writes its answer from them.

It lets an agent answer from your own material rather than its general training, and lets it cite its sources. Its answers are only as good as the documents behind it, so out-of-date or duplicated material has to be dealt with first.

Custom AI agents

Row-level security

Row-level security is a database feature, in PostgreSQL and so in Supabase, that decides, for each row, which signed-in users may read or change it, enforced by the database itself.

In Supabase apps it is the main thing standing between one customer and another’s data, because the browser talks to the database directly. Apps built fast with AI tools often have it missing or written too broadly.

Checking Supabase row-level security

Security review

A security review is an engineering-led examination of software for the weaknesses attackers look for first, with findings reproduced, ranked, fixed and retested.

Ours is not a penetration test: it is aimed at finding and fixing, with a test for each fix and a second pass to confirm it. If a certified penetration test is needed, we work alongside the firm you choose.

Security review for AI-built apps

Staging environment

A staging environment is a copy of the live system, with its own data and credentials, where changes are tested before they are released to production.

Agents and engineers can work freely there because nothing in it touches real users. A release process then moves tested changes to production under a named engineer’s control.

Keeping agents away from production

Technical due diligence

Technical due diligence is the review an investor or buyer makes of a company’s software, covering how it is built, who owns it and how secure it is.

Expect questions about code and account ownership, security weaknesses and how they were fixed, and whether another team could take the system over. Written evidence of a review, its fixes and its retests answers many of them directly.

Making an AI-built app safe before a raise or a sale

Vibe coding

Vibe coding, a term coined by Andrej Karpathy in 2025, is building software mainly by describing what you want to an AI tool, such as Lovable, Bolt, Cursor, Replit or v0, and accepting what it produces.

It is a fast way to get a working prototype in front of users. Before real customers, payments or personal data arrive, the result needs a review of access rules, keys, limits and what data leaves for AI providers.

Making an AI-built app safe

Start a conversation

Tell us what you need built.

Send a brief. It does not need to be technical. Within two working days you get a first view, and within a week a written plan, a fixed price and the risks.

In plain language: what you want to exist, who it is for, and anything that worries you. It does not need to be technical.

Read by a senior engineer. Used only to reply to you.

Tell us what you need built.