When to Use AI: The Right Tool for the Job

I spent part of the weekend reading social media. There were posts from Data Saturdays around the world, reflections on panel discussions, and some thoughtful conversations about AI. Then I went to Facebook. 😂

One question caught my attention: would you hire someone who said they used AI for 90% of their work? The comments made me laugh and cry at the same time. Some suggested 90% was not enough. Others wanted 95%, or preferably 100%. Other were at the other end of the spectrum.

My view on when to use AI is straightforward: use it when it improves the outcome, reduces the total effort, or helps you learn something worthwhile. If the existing method gets you to a reliable result more easily, use that. You still own the decision and the result.

That is the principle behind much of my AI enablement work. Copilot, Claude and ChatGPT give us more tools to choose from. Choosing well remains part of the job.

AI usage tells me very little about performance

I understand the enthusiasm behind those Facebook comments. Someone willing to explore AI might be curious, adaptable and interested in working more effectively. Those are useful qualities.

But a percentage tells me almost nothing about whether their work is any good. Does 90% mean the number of tasks, the time spent, or the proportion of an output generated by a model? Those are different things.

Someone might use AI throughout a process and still produce a poor result. Someone else might use it for one difficult part and save the team a considerable amount of effort.

If I were interviewing them, I would ask for an example. What was the problem? Why did they choose AI? How did they check the output? What improved? I would also ask when they decided to do something themselves. That answer would tell me something about their judgement.

The drill, the hole and the temptation to experiment

Rory Sutherland has a lovely story that adds a twist to the familiar argument about drills and holes. In his account of an advertising agency debate involving Llewelyn Thomas, Dylan Thomas’s son, the question is whether people buy a drill because they need a hole, or buy the drill and then find things to drill.

I recognise both behaviours in AI adoption. Some people arrive with a specific problem. Others get access to a new tool and start looking for places to try it.

There is value in that second approach. You cannot always write a convincing business case for a capability you have never explored. Experimentation can reveal useful applications you would otherwise miss.

The difficulty comes when the experiment becomes an obligation. Every email needs rewriting. Every spreadsheet needs a conversation. Every small decision becomes another prompt.

At that point, it is worth checking whether the tool is earning its place. Give people room to explore, then evaluate what they discover. Buying the drill can start the learning. It does not mean every surface in the house needs a hole.

When to use AI: the Excel example

Excel is a popular topic in the AI enablement sessions I deliver. People can connect the demonstration to something they do at work, which makes the conversation more useful.

Microsoft’s Analyst agent helps with data analysis. Separately, Copilot in Excel can suggest formulas. In a session last week, I demonstrated creating a calculated column in an Excel table using Copilot.

If you know Excel and already know the formula, you may be able to type it more quickly than you can explain the requirement, wait for a response and check the suggestion. In that situation, typing the formula is a sensible choice. By definition not using AI 100% of the time

Now consider someone who knows what they want to calculate but does not know the formula. They might otherwise search for examples, try several functions or ask a colleague. Copilot could help them reach a usable answer much sooner.

The task has stayed the same. The person’s starting point has changed. That changes the value of the tool.

Knowing a formula does not automatically rule AI out, either. If the wider task involves several unfamiliar calculations, explanations or changes, assistance might still be useful. The sensible comparison covers the work you actually need to complete.

A demonstration shows what is possible. Your working context determines whether it is worthwhile.

Getting an answer and learning are different outcomes

There is another opportunity in that Excel example. Someone can ask Copilot to explain the formula, identify the functions involved and show how changing an input affects the result. Used thoughtfully, the interaction can support learning.

But getting a working answer does not automatically mean someone has learnt how it works. Accepting a formula without understanding it can leave the next problem just as difficult as the first.

I would encourage the learner to ask questions, test a few rows and explain the calculation back in their own words. Can they spot an unexpected result? Can they adapt the formula when the requirement changes?

Sometimes spending longer today is worthwhile because it builds capability for tomorrow. That applies to learning with AI and learning through doing the work yourself. Speed matters, but it is only one of the outcomes we should consider.

Count the effort required to finish the job

One of the easiest ways to overstate AI’s value is to stop the clock when the first answer appears.

A draft might arrive quickly, but somebody still has to brief the tool, read the response, check the facts, fix the omissions and turn it into something usable. Those steps belong in the calculation.

Imagine an email that takes two minutes to write yourself. If prompting, reading and editing an AI draft takes five, you have added work. Those are illustrative timings, but the comparison is the one that matters.

The reverse can also happen. A difficult first draft might take substantial effort, while reviewing a useful starting point is much easier. An experienced person may be particularly good at recognising what needs changing.

Measure the work required to reach an acceptable result, including the checking and correction.

Also consider quality. A slightly slower process might produce something more accessible or complete. Be clear about the benefit you are seeking before declaring a productivity win.

The commercial benefit needs somewhere to go

For a business, saving time is useful when you can do something worthwhile with it. That might mean responding to customers sooner, improving a proposal, reducing a backlog or giving a stretched colleague some breathing room.

It does not automatically mean the saving turns into additional revenue or lower costs. If the same work is completed slightly faster but nothing else changes, describe the benefit accurately. You may have released capacity. Whether you turn that into a financial return is a separate management decision.

That’s an importnat factor to consider  when you are deciding whether to extend a pilot. Consider the licence, training, setup and review effort alongside the improvement. A popular demonstration is a good way to start a conversation; a repeatable benefit is a stronger reason to keep investing.

There are worthwhile benefits beyond speed, too. Helping someone attempt an unfamiliar task, understand a difficult explanation or communicate more clearly can have real value. Agree what you are trying to improve, then look for evidence of that improvement. Otherwise, you risk measuring whatever is easiest to count.

Sometimes another tool already does the job well

AI has expanded the toolbox. We still have formulas, templates, database queries, established application features and conventional automation.

If the same calculation runs every month with clear rules, a tested formula or repeatable process may be the better answer. You should not need to describe the requirement afresh every time.

AI might help you build that process. It could suggest a formula, draft a query or explain an unfamiliar feature. Once you have checked and established the solution, the routine work may need very little further AI involvement.

That would be a useful outcome, even if your AI usage subsequently fell. The improvement would live in the process.

Equally, a brief conversation with a colleague can sometimes resolve an ambiguity faster than a series of increasingly elaborate prompts. Tool choice includes deciding when another person has the context you need.

The human in the loop needs to do something

Having a human in the loop should mean something. What is that person responsible for checking, and can they actually check it? They need to own the loop.

In Excel, a formula can calculate successfully while answering the wrong business question. It might use the wrong date, exclude relevant rows or treat missing values in a way nobody intended. A plausible number is not sufficient evidence.

The person using the output needs to understand the requirement, test relevant cases and know when to ask for help. If they cannot judge the result, pressing an approval button adds very little assurance.

The consequences also matter. Trying alternative wording for an internal note is different from changing a calculation that informs an important business decision. The latter needs an appropriate level of review.

You also need an approved tool and suitable handling of the information involved. A potentially useful prompt does not remove those responsibilities.

Five questions to ask before you use AI

You do not need a committee meeting before every prompt. You do need a habit of thinking about the task. These are the questions I would start with:

1. What outcome do I need? Be specific about the answer, document, decision or capability you are trying to produce.
2. What would I do without AI? Compare it with the method you would realistically use, including any existing template or automation.
3. Where could AI help? Identify the difficult part: getting started, finding an approach, explaining something or handling repetitive drafting.
4. How will I check the result? Decide what evidence, tests or expertise you need before relying on the output.
5. Was the overall result better? Include effort, quality, learning and the consequences of errors. Keep the approach if it earns its place.

The answer can change as your skills improve or the tools develop. A useful decision today deserves another look when the circumstances change.

Give people permission to exercise judgement

For CIOs and IT directors, this has a practical implication: be careful what your adoption measures encourage.

Usage figures can help you understand whether people are exploring a tool. On their own, they cannot tell you whether the work has improved. If people feel judged mainly on frequency, they may find reasons to use AI where it contributes very little.

Pick a few real tasks, establish the current approach, agree what a good result looks like and compare the alternatives. Give the people doing the work a say in that assessment.

Keep a few examples of what worked and what did not. A short record of the task, the approach and the result is more useful than a growing collection of impressive screenshots.

Some experiments will demonstrate a benefit. Others will show that the current method works well. Both findings help you spend your time and money more sensibly.

Keep some room for learning that has no immediate return. People need opportunities to discover what a tool can do and build confidence in assessing it. Give that exploration a purpose and a boundary, then review what was learnt. Expecting every first attempt to save time can discourage useful experimentation just as much as insisting every task must involve AI.

I would challenge both blanket positions: refusing to explore AI and insisting on using it for everything. Neither is a useful rule for deciding how to handle a particular task.

Use it where it helps. Learn where it might help. Be comfortable choosing another approach when that makes more sense. The judgement remains yours.

Find a sensible starting point

If you are deciding where AI could help your organisation, start with the work you need to improve rather than the tool you have licensed. My free AI Readiness Assessment takes a few minutes and gives you a structured view of where you stand across the dimensions that predict whether a rollout pays back.

Take the AI Readiness Assessment

If you already know the gap is structure rather than licences, the AI Enablement Programme is the follow-through: readiness review, role-based use cases, a defined quality bar and the governance to make the gains hold. The frameworks behind it are in my book, AI That Actually Works.