The pressure to “do something with AI” is real. Most of it is hype and a distraction. You have probably sat in a meeting where someone insisted the business is falling behind, without anyone being able to specifically say what problem(s) AI would solve. Think of looking down the wrong end of the telescope as a good metaphor. The useful question is not “how do we use AI”; it is “where in this business is there a job where AI adds measurable value?”.
Here’s how to answer that without a big investment of time or a six-figure commitment.
Start with the work, not the technology
AI is a tool, and like any tool it fits some jobs and mangles others. The trap is starting from the technology and hunting for somewhere to put it. That is how you end up with that chatbot nobody uses.
Start instead from the real work your people do all day. Somewhere in that work are tasks that are repetitive, eat hours, and follow a pattern. Those are the places to start looking at carefully. The rest is distraction, and an effective AI partner should tell you so.
The three places AI returns value
Across the projects we see, AI pays off in three fairly consistent patterns. If your idea maps onto one of these, it’s worth pondering seriously.
Assisting repetitive human judgement
Not replacing judgement, assisting it. Think of a person who reviews the same kind of thing over and over: applications, claims, support tickets, contracts, incoming enquiries. The decision needs a human, but most of the work is reading, sorting, and flagging the handful that need real attention. AI is good at doing that first pass and putting the important ones in front of the person, with a reason attached. The human still decides. They just stop wading through the routine to get to the exceptions.
Making dense information usable
Most organisations sit on more data than they can read: reports, monitoring feeds, documents, records that, in theory, hold the answer and, in practice, never get looked at because finding the answer is too much work. This is where AI is strong. It can take a wall of information and turn it into a plain-language answer to the question someone is asking, immediately. The value is not the model. It is that a person who needed an answer now gets one, instead of a spreadsheet they have to work through and interrogate.
We built one of these for ourselves. Public sector opportunities in Scotland are published on a portal with no export and no feed, so the only way to read it was by hand, which meant it went unread. We wrote a script that collects the notices, scores each one for relevance to what we actually do, and sorts them into what needs a response now and what is worth watching. The last run took 1,467 notices down to 164 worth a look, then to 50 published recently enough to matter, and finally to 6 worth acting on. The first pass also turned up a contract that had opened and closed while nobody was watching, which is the honest argument for doing it this way: the information was always public, it was just unreadable at that volume.
Accelerating skilled work under supervision
AI can take a skilled person’s first draft off the table: the boilerplate, the initial version, the routine scaffolding, so the expert spends their time on the parts that need expertise. The keywords are “under supervision”. This works when a competent human reviews and owns the output. It doesn’t work the moment you treat the machine’s first attempt as the finished article.
Notice what these three have in common. There is a repeating task. A human stays accountable for the outcome. And the win is measured in someone’s day getting better, not in having AI on the org chart.
The three places it usually does not work well… yet
Just as consistent are the places where AI tends to disappoint and drain budget.
One-off decisions that carry real weight
If a task happens rarely and the cost of getting it wrong is high, AI is a poor fit. There is no repetition to lean on and no safety margin for a confident wrong answer. Keep those with your best people.
Anywhere the data is missing, messy, or scattered
AI is only as good as what you feed it. If the information lives in five systems, three inboxes and someone’s head, no model fixes that. You would be paying to automate a mess. The best first step is simply to sort the data out, and sometimes that alone delivers most of the value you were chasing.
Work where a plausible wrong answer is worse than no answer
AI is very good at sounding right. In some contexts, a confident answer that happens to be wrong causes more damage than admitting you do not know. If your process cannot catch and absorb that, the risk isn’t worth the savings.
Four questions to score any AI idea
Before you spend anything, run the idea through these. You do not need to be technical to answer them, and if you cannot, that is useful information in itself.
-
Is there real, repetitive human judgement to assist? A genuine pattern that happens often, not a rare one-off dressed up as a process.
-
Is the data actually there, and is it clean? Can you point to where the information lives, and is it in a state a system could use? Be honest here. This is where most projects fail.
-
Is the cost of a wrong answer tolerable? If the system gets it wrong now and then, and it will, can your process catch it and carry on? Or does one bad answer do real harm?
-
Will the people it is for actually use it? The most capable tool in the world is worthless if it does not fit how your team works. If adoption is a maybe, treat that as a no until you have tested it.
An idea that passes all four is worthy of prototyping. One that fails two or more needs to be parked. Move on.
“Our competitors are doing it” is not a reason
It is worth saying plainly, because it drives a lot of wasted spend. A competitor launching an AI feature tells you they decided to, and nothing about whether it works, whether anyone uses it, or whether it delivers meaningful ROI. You are seeing the press release, not the numbers.
The businesses that get real value from AI are not the ones that moved fastest. They are the ones that were honest about where it fit, put a human firmly in charge of the outcome, and were willing to say no to ideas that didn’t pass the test.
Where to start
You do not need a grand AI strategy. You need one well-chosen problem where the answer to all four questions is yes, run as a small, contained test before you commit real money. If it works, you scale it. If it does not, you learned that cheaply, which is the whole point.
The most useful thing a partner can do at this stage is not build you something. It is to help you see clearly where AI belongs in your business and, just as importantly, where it does not. If you’d like to talk through one of your ideas, we are happy to spend 30 minutes on it with no pressure and no sales pitch.