What is Artificial Intelligence?

A working definition for business leaders

Artificial intelligence describes computer systems that perform tasks we would normally associate with human cognition: interpreting language, recognising images, spotting anomalies, forecasting outcomes, summarising information and making recommendations. The academic definition has shifted repeatedly since the term was coined at the Dartmouth workshop in 1956, but for commercial purposes the useful test is simple. If the software improves its output by learning from examples rather than by a developer writing an explicit rule for every situation, it is reasonable to call it AI.

That distinction matters because it changes how you buy, deploy and govern the technology. Traditional software is deterministic: given the same input it produces the same output, and a defect is a bug that can be located and fixed. AI systems are probabilistic: they produce the most statistically likely output given their training and their instructions, which means two near-identical requests can return different answers, and a wrong answer is often not a bug at all but a predictable characteristic of the model. Organisations that treat AI as ordinary software are usually the ones caught out by it.

The OECD's widely adopted definition — reflected in the EU AI Act — frames an AI system as a machine-based system that infers from the input it receives how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments. The emphasis on inference and on influence is deliberate: it is what triggers regulatory obligations, and it is also what creates commercial risk and commercial opportunity.

The main types of AI used in business

How modern AI actually works

Almost all of the AI now reaching desks in ordinary businesses is built on large language models. These are neural networks trained on very large volumes of text to predict the next fragment of language given everything that came before. Through that training they absorb grammar, factual associations, reasoning patterns and writing conventions. When you send a prompt, the model does not look up an answer in a database; it generates one token at a time, choosing what is statistically most plausible given your instruction and any context supplied alongside it.

This explains both the strengths and the failure modes. Language models are extraordinary at tasks with many acceptable answers — drafting, summarising, restructuring, explaining, translating and brainstorming. They are unreliable at tasks with exactly one correct answer that is not present in the context they were given, which is why they can produce confident but fabricated citations, figures or case references. The industry term for this is hallucination, and mitigating it is a design problem rather than something the user can solve with better wording alone.

The most effective mitigation in enterprise deployments is grounding: retrieving relevant, permissioned documents from your own systems and supplying them to the model as context so it summarises real source material instead of drawing on general training. Microsoft 365 Copilot grounds answers in your Microsoft Graph content; enterprise deployments of ChatGPT, Claude and Gemini can be grounded in a connected knowledge base. Grounded systems are markedly more accurate, but they inherit your permission model exactly — which is why access hygiene is the single biggest technical prerequisite for safe AI adoption.

Where AI creates measurable business value

In our experience the reliable returns cluster in four places. The first is written output: proposals, reports, correspondence, meeting notes and internal documentation, where a good first draft removes most of the effort. The second is finding and synthesising information across a scattered document estate. The third is structured extraction — pulling fields from invoices, contracts, forms and applications so they can flow into a system of record without re-keying. The fourth is triage: routing, prioritising and summarising inbound work so that skilled people spend their time on judgement rather than sorting.

What these have in common is that they are frequent, language-heavy, tolerant of a reviewed draft, and measurable. That last property matters most. Organisations that establish a baseline — hours spent, cycle time, backlog, error rate — before deployment can prove value afterwards. Those that do not are left arguing about anecdotes at renewal time. Our AI Productivity & Readiness Assessment exists precisely to establish those baselines and to sequence the opportunities that follow from them.

The limits leaders should understand

AI does not understand your business, does not know what it does not know, and has no concept of accountability. It reflects the data it was trained on, including its biases, and it reflects the permissions of whoever asks it a question. It cannot be a decision-maker in any process with legal or safety consequences; it can only support a person who remains accountable for the outcome. Regulators across the UK, EU and elsewhere have converged on this principle of meaningful human oversight, and every credible internal policy restates it.

None of this argues against adoption. It argues for adoption with structure: defined use cases, grounded data, trained users, recorded decisions and periodic review. That is the difference between organisations where AI quietly compounds into real capacity, and organisations where it produces a burst of excitement followed by an expensive shelf of unused licences.

Sources and further reading

• Recommendation of the Council on Artificial Intelligence (definition of an AI system)

  OECD

• Regulation (EU) 2024/1689 — the EU Artificial Intelligence Act

  European Union

• A pro-innovation approach to AI regulation

  UK Government, Department for Science, Innovation and Technology

• AI Risk Management Framework (AI RMF 1.0)

  NIST