When the Tool Becomes the Master

A manager has twenty minutes to respond to a customer complaint. She pastes the complaint into an AI tool and asks for a reply. In seconds, it produces a calm, polished message offering a refund. It reads well. There is only one problem: the manager has not checked what was promised, whether the customer received the service, or whether she has the authority to offer a refund. The machine has drafted the email and, quietly, made the decision.
That is how a tool becomes the master. Nobody formally puts it in charge. It starts making the choices we stop noticing. If the machine sets the goal, selects the evidence and proposes the action, the person who simply presses Send has become the tool. Over time, giving away every opportunity to exercise judgement leaves us less practised at making it.
I expect AI to become an ordinary part of work. We will use it to write, analyse, plan and organise, often without opening a separate AI app. As access to capable tools spreads, knowing where to click will matter less. Knowing what problem to solve, what evidence to trust and what trade-off to accept will matter more. We risk making ourselves redundant in a particular decision when we routinely hand over every part of it, including the reasoning someone would otherwise need us for.
The upside is real. So is the boundary.
In a study of 5,172 customer-support agents, Erik Brynjolfsson, Danielle Li and Lindsey Raymond found that agents with AI assistance resolved 15 per cent more issues per hour on average. The gains varied substantially, and the researchers also found evidence that workers learned through using the tool. This is a good reason to bring AI into the workplace, especially when it helps people see what a more experienced colleague might do.
But the same tool can help with one task and mislead on the next. In a field experiment with management consultants, Fabrizio Dell’Acqua and colleagues found that consultants using AI did better on a set of tasks the tool handled well. On one complex task deliberately chosen because it sat outside the tool’s capabilities, AI-assisted participants were 19 percentage points less likely to reach the correct answer. The point is not to generalise from one task to every office. It is to recognise how difficult that boundary can be to see from the quality of the prose.
A 2025 survey of 319 knowledge workers adds another piece: respondents who were more confident in AI reported doing less critical thinking, and they described some of their effort shifting towards checking and integrating answers. A self-report survey cannot tell us whether people lose the ability to think over time. It does raise a more immediate question: what checks do we drop today because the answer looks finished?
Keep three parts of the work in your hands
Decide what you need before you open the tool. Take two minutes to write down the decision, the evidence you have and what would change your mind. The manager answering the complaint might write: “I need to respond today without promising an outcome before I check the agreement.” Now she can ask AI to outline two ways to acknowledge the concern. She has defined the job before the tool can redefine it for her.
Ask for a challenge, then check what matters. When you ask for an answer, also ask the tool to identify missing facts and make the strongest case for a different conclusion. Treat that as a way to find leads, then go back to the original evidence. If the draft says the customer is owed a refund, read the agreement and check the service record yourself. Follow your organisation’s rules about what information you can put into an AI tool; often you can describe the problem without sharing anyone’s personal details.
Own the final choice, and practise making it. Before you send a message, approve a proposal or act on a recommendation, see whether you can explain your decision in two plain sentences. If it matters to the business, record the reason and the name of the person who approved it. Keep some practice without assistance, especially for the skills you want new colleagues to develop. A trainee who never tries to diagnose a problem is less likely to know when the suggested diagnosis misses the point.
Try it on one recurring task this week
Pick something you do often: compare two supplier proposals, review a weekly report or respond to a difficult request. Do the first five minutes yourself. Then use AI to improve the analysis or draft the response. At the end, note how much time you saved, what you caught that the tool missed and what it caught that you missed. After three runs, decide what you can safely delegate and what still needs your attention.
For adults who want to practise with examples from work and business, CAMA College’s AI demo and events page lists sessions at our Richmond Hill campus. Whether you learn with us or try the exercise at your own desk, start with a decision you understand well enough to judge the output.
The question worth asking at the end of the week is simple: if I had to defend this decision tomorrow without the chat open, could I?




Comments