What leaders need to know about AI, without the hype

Why this matters at your level

You will be asked to approve AI projects, fund them, defend them and sometimes stop them. Vendors, consultants and enthusiastic colleagues will describe what AI can do in confident terms. Critics will describe what it cannot do in equally confident terms. You do not need to understand the mathematics. You do need a clear, accurate picture of what the technology is, so that you can tell a realistic proposal from an expensive promise.

What "AI" usually means today

Artificial intelligence is a broad label. In most proposals you will see, it means one of two things:

• Generative AI. Generative AI tools, usually built on a large language model, produce text, images or audio in response to instructions. Chat assistants are the familiar example. They are good at drafting, summarising, translating and rephrasing.
• Predictive or classification systems. These learn patterns from past data to sort, score or forecast: flagging unusual claims, predicting demand or routing applications. They are narrower and often more dependable for one well-defined task, but only as good as the data behind them.

Ask which kind a proposal involves. The risks, costs and safeguards differ.

What it does well and where it fails

Current AI is genuinely useful for work that is repetitive, language-heavy and easy to check: first drafts, summaries of long documents, sorting large volumes of correspondence and extracting information from forms. It can save real staff time on these tasks.

It is weak in ways that matter for government:

• It can be confidently wrong. Generative tools produce hallucinations: fluent, specific statements that are false. They do not know when they are wrong.
• It reflects its data. Systems trained on skewed data reproduce that skew. This is bias and it can fall hardest on the citizens with least power to complain.
• It does not understand context the way staff do. It does not know your department's history, the politics of a file or the person behind an application.
• Performance changes over time. A system that worked well at launch can degrade as circumstances change. This is model drift and it needs monitoring.
• It is often hard to explain. Many systems cannot give a clear reason for a particular output, which matters when a decision affects a citizen's rights.

What AI really costs

The licence or subscription is usually the smallest part. A realistic budget includes:

• preparing and cleaning data, which is often the largest single effort
• integration with existing systems and records
• staff time for checking outputs, which never falls to zero
• training, guidance and support
• ongoing compute or usage charges, which can rise sharply with volume
• monitoring, evaluation and the cost of leaving a supplier later

Ask for the full cost over several years, not the first-year price. Ask what happens to costs if usage doubles.

Reading claims with care

When you hear a claim about AI, ask four questions:

1. Compared with what? Better than current practice or merely better than nothing?
2. Measured how? On a benchmark or demonstration? Or on work like ours, in our languages, with our data?
3. Who checked? The vendor or someone independent?
4. What happens when it is wrong? Who notices, how quickly and who is harmed?

A demonstration on clean English text tells you little about performance on handwritten Urdu forms or regional language correspondence. Ask to see the system tested on material like yours.

Your role as a leader

You are not expected to be the technical expert. You are expected to ask clear questions, insist on evidence and keep accountability with named people. The rest of this pathway builds those habits. Officers should also read the explainer on the National AI Policy 2025.

Try it

Find a public article or vendor web page that makes a claim about what AI can do for government. Apply the four questions above. Note which questions the material answers and which it leaves open.

Check yourself

1. Name the two broad kinds of AI system you are likely to see in proposals.
2. List three costs beyond the licence fee.
3. Why does a demonstration on clean English text tell you little about your department?

Answers: (1) Generative AI and predictive or classification systems. (2) Any three of: data preparation, integration, staff checking time, training, usage charges, monitoring, exit costs. (3) Performance depends on the material; your department's documents, languages and data may be very different.