Setting direction: where AI could help your department

Why direction matters

Without direction from the top, AI arrives in a department in one of two ways. Either staff start using whatever free tools they find with no safeguards. Or a vendor arrives with a product and the department looks for a use for it. Both are poor ways to spend public money and public trust. Your job is to set a clear, modest direction: which problems are worth exploring, in what order and under what conditions.

Start from problems, not tools

The most common mistake is to begin with "we should use AI" and then search for somewhere to apply it. Begin instead with the problems your department already knows it has:

• Where do files and applications wait longest?
• Which tasks consume large amounts of staff time for little judgement?
• Where do citizens most often complain about delay, inconsistency or lack of information?
• Where do errors keep recurring?

For each problem, ask what is really causing it. Often the cause is a missing process, unclear rules, poor records or too few staff. AI will not fix any of those and may hide them. Only when the cause is genuinely the volume or repetitiveness of work is AI worth considering.

Sort ideas by value and risk

Once you have a list of candidate problems, sort them on two simple questions.

How much would it help? Think about time saved, faster service for citizens, fewer errors and better information for decisions.

What happens if it goes wrong? Think about who is affected, how badly and how quickly a mistake would be noticed.

This gives you four groups:

1. High value, low risk. Start here. Internal drafting support, summarising long reports for staff, sorting incoming correspondence for a human to act on.
2. Low value, low risk. Allow staff to explore these within your guidance, but do not invest heavily.
3. High value, high risk. Anything that affects a citizen's entitlement, liberty, money or record. These may be worth doing eventually, but only after you have experience, strong safeguards and evidence. Module 3 covers how to classify them.
4. Low value, high risk. Do not do these.

Low-risk first, for good reasons

Starting with internal, low-risk uses is not timidity. It lets your department:

• learn how the technology behaves on your real work
• build the skills to check output and spot failure
• find out whether your data and systems are ready (module 5)
• build an honest record of what worked and what did not

That experience is what makes later, higher-stakes decisions safer. A department that has never run a small pilot is not ready to automate decisions about citizens.

Turning direction into practice

A short written direction from you is more useful than a long strategy. It might set out:

• the two or three problems the department will explore first and why
• the uses that are not permitted, such as making decisions about individual citizens without human review
• who owns each exploration and who they report to
• what evidence you expect before anything is extended
• how staff may and may not use AI tools in the meantime

Small, time-bound pilots are the usual next step. The how-to guide on running a small, safe pilot describes the mechanics. Module 8 covers how to judge the results.

What to avoid

• Announcing before testing. Public commitments create pressure to declare success.
• Pilots with no question. "Try AI in the department" is not a pilot. "Can this tool cut the time to draft replies to routine letters without reducing accuracy?" is.
• Scaling on enthusiasm. One good demonstration is not evidence.
• Ignoring the people affected. Staff who will use the tool and citizens who will feel its effects should shape the choice of problem.

Try it

List five recurring problems in a public service you know well, using only public information such as published annual reports or service standards. For each, write one sentence on the likely root cause. Then sort them into the four value and risk groups and pick the one you would explore first.

Check yourself

1. Why should you start from problems instead of tools?
2. What are the two questions used to sort candidate uses?
3. Give two reasons for starting with low-risk uses.

Answers: (1) Starting from tools leads to looking for uses to justify a product; starting from problems shows whether AI is the right fix at all. (2) How much would it help and what happens if it goes wrong. (3) Any two of: learning how the technology behaves, building checking skills, testing data readiness, building an honest evidence record.