AI has become part of almost every conversation about digital transformation. Suppliers have AI roadmaps, conferences are full of AI sessions, and most organisations are exploring where the technology might improve services or productivity.
There are good reasons for the interest, but I’m less interested in how many AI use cases an organisation can identify than in a much simpler question:
Where does AI genuinely create enough value to be worth using?
In local government, I think some of the clearest opportunities are relatively unglamorous. They involve reducing repetitive administration, making information easier to find and giving skilled people more time to do the work that actually requires their expertise.
That doesn’t mean every administrative process needs AI. Often a simpler process change or conventional automation will be cheaper, more reliable and easier to operate. The useful question is where AI adds something those approaches cannot.
Start with the work, not AI
Starting with “Where can we use AI?” almost guarantees a long list of possible use cases. It doesn’t tell you whether any of them are particularly valuable.
I prefer to look at where people are spending significant amounts of time on work such as reading and categorising information, searching for things, summarising documents, producing repetitive first drafts or extracting useful information from unstructured material.
Typical examples include:
- reading and categorising large volumes of email
- searching multiple sources for relevant information
- drafting routine responses
- summarising lengthy case histories
- producing meeting notes and actions
- preparing first drafts of reports
- finding policies, procedures and guidance
These activities are necessary, but they often consume time that could be spent on work requiring professional knowledge, judgement or direct interaction with residents.
That makes them worth investigating. It doesn’t yet make AI the answer.
Know when ordinary automation is enough
This distinction is important because AI and automation are often talked about as though they are interchangeable.
If a process has clear rules and structured data, conventional automation may be the better solution. If an incoming form contains a reference number in a known field and needs to be routed according to a fixed rule, you probably don’t need an AI model to do it.
AI becomes more interesting when the work involves language or other unstructured information that traditional rules struggle to handle. Understanding what an email is about, summarising a long case history, finding the relevant passage across hundreds of documents or producing a useful first draft are different kinds of problems.
The objective shouldn’t be to maximise the amount of AI in a process. It should be to use the simplest technology that solves the problem reliably.
Email and correspondence
Shared mailboxes are a good example of the sort of administrative workload where AI can be useful. A busy service may receive hundreds or thousands of emails each week, all of which need to be read before somebody can decide what they are about and what should happen next.
AI can help identify the subject, urgency and likely destination of an email, extract useful information and prepare an acknowledgement or draft response. Depending on the service and the level of confidence, some of those actions may eventually be automated while others remain subject to human review.
The potential saving isn’t difficult to understand. If staff currently spend hours each week reading correspondence simply to categorise and route it, reducing that handling time releases capacity for other work.
The important part is measuring the real process rather than assuming the saving. How many emails are there? How long do they take to handle? What proportion can the system classify reliably? What happens when it gets one wrong? Those answers determine whether the use case is actually worthwhile.
Finding information
Another promising area is knowledge retrieval.
Council staff often work across policies, procedures, guidance, previous decisions and information held in different systems or document libraries. Finding the right information can take longer than using it.
AI can make that material easier to interrogate by allowing somebody to ask a question in ordinary language and receive a concise response based on the relevant source material. Done properly, the answer should retain links or citations back to those sources so the officer can check what it is based on.
This can be particularly useful in customer services or for people who are less familiar with an organisation or service. The benefit isn’t that AI becomes the authority; the underlying policy or guidance remains the authority. AI reduces the effort involved in finding the relevant part of it.
That distinction matters. A confident answer without a reliable source is not an improvement.
Drafting and summarising
Generative AI is also good at producing a useful starting point from information that already exists.
That might mean drafting a resident email from case information, turning meeting notes into a structured summary, producing a first version of a report or condensing a lengthy case history into the significant events, decisions and outstanding actions.
These are attractive use cases because they can save relatively small amounts of time very frequently. Saving ten minutes on an activity carried out thousands of times can matter more than automating an impressive but infrequent task.
Human review still matters, particularly where communications affect residents or summaries are being used to support important decisions. The value comes from reducing the effort required to get from a blank page or a large body of information to something a person can review and complete.
AI can help specialists spend more time being specialists
For me, this is one of the more useful ways to think about AI productivity.
A housing officer’s value isn’t in repeatedly reformatting information. A planning officer’s expertise isn’t best used searching through folders for a document. A social worker’s time is more valuable working with people than producing routine administration. Transformation teams don’t add much value by spending hours turning meeting notes into board-paper prose.
There will always be administration around professional work, and some of it requires judgement. But where technology can safely reduce the mechanical part of that workload, the organisation gains capacity without pretending that professional expertise has been automated.
This is also why I am cautious about translating every time saving directly into a financial saving. Releasing several hours of an officer’s week can be valuable because it reduces a backlog, improves response times or allows more cases to be handled. It doesn’t necessarily mean a corresponding proportion of a post can be removed from the budget.
The benefit should be described for what it actually is.
Programme and management work is worth looking at too
AI opportunities aren’t confined to frontline services. There is plenty of administrative work in programmes and management functions as well.
Meeting notes, workshop summaries, document comparison, action extraction, research, first drafts of papers and routine communications can all consume considerable time. Used sensibly, AI can reduce some of that effort.
I use this kind of assistance differently from something that is directly affecting a resident’s case. The consequences of an error are different, so the controls should be different too.
That points to a broader principle: AI governance should be proportionate to what the AI is actually doing.
Put controls around the risk, not around the word AI
Responsible use obviously requires attention to information governance, data security, transparency, record keeping, quality and human oversight. But applying exactly the same controls to every AI use case can make adoption unnecessarily difficult without materially reducing risk.
There is a significant difference between AI helping somebody summarise their own meeting notes and AI contributing to a decision that affects a resident’s entitlement to a service.
The governance should reflect that difference.
For each use case, I want to understand what information the system can access, what it is being asked to produce, what happens if the output is wrong, whether a person reviews it, what needs to be recorded and how its performance will be checked.
The more consequential the use, the stronger those controls should be.
Measure whether it actually saves anything
An AI pilot producing technically impressive results isn’t necessarily a successful transformation initiative.
If the reason for introducing it is productivity, measure productivity. Establish how much effort the existing process takes and compare it with what happens after the change.
Depending on the use case, useful measures might include:
- handling time
- administrative effort
- response times
- backlog
- first-contact resolution
- accuracy and rework
- capacity released
- employee experience
Quality needs to sit alongside time saved. A tool that produces work more quickly but creates additional checking, corrections or downstream problems may simply have moved effort somewhere else.
It is also worth looking at the whole process. Saving two minutes in one part of a transaction is of limited value if the change creates another five minutes of work later.
Use pilots to replace assumptions with evidence
For many AI opportunities, the initial business case will contain quite a lot of uncertainty. That is a good reason to start with a bounded pilot or Proof of Concept rather than committing immediately to a large implementation.
Choose something with enough volume for a saving to matter, outcomes you can measure, manageable consequences if the technology gets something wrong, and users who are willing to participate.
Before starting, establish the baseline. Then test the technology using real work and measure what changes.
A useful pilot may show that the opportunity is larger than expected. It may also show that accuracy isn’t good enough, integration is harder than anticipated, users don’t trust the output or the supposed saving disappears once human checking is included.
Those are all useful findings. The purpose of a pilot is to improve the investment decision, not to prove that the original AI idea was right.
Don’t use AI to preserve a bad process
There is a wider transformation point here.
Adding AI to an inefficient process doesn’t necessarily make it a good process. It may simply make one part of the inefficiency happen faster.
Before introducing AI, it is worth asking why the work exists in its current form. Can a step be removed altogether? Could residents provide better information at the start? Are two teams doing work that should happen once? Could an existing system already solve the problem? Would a straightforward process change remove the need for the task?
Sometimes AI will still be useful after those questions have been answered. Sometimes the best AI use case disappears because the underlying problem has been fixed more simply.
That is a successful transformation outcome too.
Build from evidence rather than an AI roadmap
AI capability will continue to change quickly. Models will improve, costs will change and more of the technology will become embedded in software councils already use.
That makes me wary of detailed multi-year AI roadmaps built around today’s products and capabilities. They can become obsolete remarkably quickly.
A more durable approach is to build organisational capability around identifying suitable problems, assessing risk, running controlled experiments, measuring outcomes and scaling the things that work.
That creates a way of adopting useful AI as the technology changes without requiring the organisation to chase every new feature.
Where AI genuinely saves time
The most convincing AI opportunities I’ve seen aren’t necessarily the most ambitious. They tend to start with a recognisable operational problem: too much correspondence to process manually, too much time spent finding information, repetitive drafting, lengthy documents to understand or administrative work taking specialists away from their main role.
The test is then practical. Can AI perform a useful part of that work reliably enough? Does it genuinely reduce effort once checking and exceptions are included? Is the risk manageable? Is AI actually a better solution than conventional automation or process redesign?
If the answers are yes, there may be a good case for using it.
Local government doesn’t need AI for its own sake. It needs better services, more productive organisations and scarce professional capacity spent where it has the most value. AI is useful when it demonstrably helps achieve those things.