— Guide —

Large language models for business,
explained plainly.

A practical, vendor-neutral guide from the Institute of AI, the UK's professional body for AI. What large language models are, what they can and cannot do, where they go wrong, and how UK boards adopt them with confidence.

Sign the Charter Talk to us

— The short answer —

What is a large language model for business?

A large language model, often shortened to LLM, is software trained on vast amounts of text to predict the most likely next words in a sequence, and then further trained on human feedback to follow instructions rather than simply continue text. In business, that prediction becomes useful work: drafting a reply, summarising a long contract, answering a question from your own policies, or writing a first pass of code, in seconds. Familiar names such as ChatGPT, Claude, and Gemini are all products built on large language models.

It helps to be precise about what these models can and cannot do. They are strong at language-heavy, high-volume work where a person still reviews the result: summarising, rephrasing, classifying, extracting, and drafting. They are weak at anything that requires guaranteed accuracy or current facts they were never given, and they are unreliable at arithmetic and quantitative reasoning unless connected to a tool that does the calculation and shows its working. They predict plausible text, not verified truth, so they can state invented facts with complete confidence, absorb bias from their training data, and expose sensitive information if staff paste it into the wrong tool.

None of that is a reason to avoid the technology. It is a reason to adopt it with controls. UK organisations use large language models safely by starting from the problem rather than the tool: choosing a small number of measurable use cases, grounding the model in their own trusted sources, keeping a named person accountable for consequential output, and governing the whole thing against clear standards. That is the difference between a demo that impresses and a system a board can stand behind.

— The building blocks —

What a business system is actually made of.

A large language model deployed for business is not one product. It is a stack of decisions. Understanding the parts is how leaders ask the right questions of a vendor or an internal team.

The model itself

A large language model is trained on vast amounts of text to predict the next word, then further trained on human feedback so it follows instructions rather than simply continuing text. Business rarely trains one. The choice that matters is which model, hosted where, and under what data and residency terms.

The context window

Everything a model can consider at once, from your prompt to the documents you attach. It is finite. Feeding a model too much, or the wrong things, is a common and quiet cause of poor answers.

Retrieval and grounding

Connecting the model to your own policies, contracts, and records so answers cite your organisation rather than the open internet. This is what turns a generic assistant into something defensible.

Fine-tuning and adaptation

Adjusting a base model to your tone, formats, or domain language. Useful for narrow, repeated tasks, but usually the last lever to reach for after prompting and retrieval have been exhausted.

Human oversight

A named owner, a review step for anything consequential, and a clear escalation route. A language model drafts and suggests. A person stays accountable for what reaches a customer or the board.

Evaluation and monitoring

Measuring accuracy, safety, cost, and drift before launch and continuously afterwards. Without a baseline you are shipping on faith and discovering failures through complaints.

— Where it goes wrong —

Six failure modes, and the control for each.

Every one of these is common with large language models, and every one is preventable. The pattern is the same throughout: pair the model's speed with a human control and an evidence trail.

01
The confident wrong answer
A model states an invented fact, figure, or legal citation with total fluency, and a busy colleague forwards it to a client as settled truth.
Ground answers in retrieved source documents and surface the source, so any claim can be checked against something real before it leaves the building.
02
Sensitive data pasted into the wrong tool
Staff paste customer records, contracts, or source code into a public consumer model, and that data leaves your control and your jurisdiction.
Provide an approved enterprise route with clear data terms and residency, and set an acceptable-use policy people can actually understand and follow.
03
Nobody owns the output
A model-drafted decision reaches a customer with no human who reviewed it and no one accountable when it turns out to be wrong.
Assign a named owner per use case and require human review on any consequential, customer-facing, or regulated output before it ships.
04
Quiet bias at scale
A model reproduces patterns from its training data in shortlisting, pricing, or eligibility, applying them at scale and out of sight.
Test outcomes across groups before launch, keep a person in the loop, and record the lawful basis for any automated decision that affects people.
05
The prompt-injection back door
A model connected to your systems follows hidden instructions buried in a web page or an email, and acts against you on someone else’s behalf.
Treat retrieved content as untrusted input, constrain what tools the model can call, and never let an unreviewed model take an irreversible action.
06
Cost and lock-in creep
A single provider changes its model, pricing, or terms, and usage costs and behaviour drift under a service you cannot easily leave.
Keep an inventory, monitor spend and drift, and design for substitution so one supplier is never a single point of failure or a runaway bill.

— How the Institute of AI helps —

Independent standards, not a product pitch.

As the UK's professional body for AI, IoAI helps organisations adopt large language models responsibly through three routes, reselling no third-party software and taking no implementation kickbacks.

— Free to sign —

The UK AI Readiness Charter

A public commitment to getting your organisation ready for AI, set out as five practical pledges. Choose at least three, sign for free, and signal intent to staff, customers, and partners.

Sign the Charter

— Standards —

Organisation accreditation

Independent assessment against the professional standards of the Institute of AI, giving boards external assurance that language models are governed, skilled, and accountable rather than merely present.

Explore accreditation

— Advisory —

Independent AI advisory

Practical, vendor-neutral guidance for leaders setting a language-model position: readiness assessment, governance, and a costed roadmap from a body that resells no third-party software and takes no implementation kickbacks.

Talk to an adviser