How to be a good manager in the age of AI.

In 2024, 75 % of knowledge workers surveyed in the Microsoft and LinkedIn Work Trend Index said they used AI at work. The study covered 31 000 people in 31 countries. Among these users, 78 % brought their own tools.

In many companies, teams already use these tools to prepare for a meeting, summarize a document, rephrase a message or analyze a spreadsheet. They also use them to write code or produce a first draft.

Practices have moved faster than internal rules. I would therefore start with a very simple discussion with the team, focused on the points where AI already plays a part in everyday work.

Making it possible to talk about AI use at work

I would start with a basic point. An employee must be able to say they have used AI without lowering their voice.

A global Slack Workforce Lab survey, conducted among more than 17 000 office workers, reveals considerable unease. Nearly one person in two, 48 %, said they would feel uncomfortable admitting to their manager that they had used AI for at least one routine task. Some feared giving the impression that they were cheating or being seen as lazy.

I find this figure worrying. When an employee hides their use, the manager loses an opportunity to understand the method. They also lose a chance to see what works and correct weak points.

I would like to hear a simple statement more often in teams. I used AI for this part. Here is what I asked it to do. Here is how I checked the response. From there, the work can be reviewed together, with the areas of risk in plain sight.

The manager sets the tone. A belittling remark about a draft produced with ChatGPT will mainly make people want to hide the next draft.

The freedom to speak openly requires understandable rules. In my team, everyone would know which tools are authorized. Everyone would also know which data must stay within the company. When a result could affect a client or an internal decision, the expected level of verification would be defined before work begins.

These rules benefit from being written with the teams. A general policy does a poor job of covering every situation encountered in a profession. Each activity handles different information. Risks vary from one role to another.

In a work discussion, I would ask a direct question. Show me how you used it and how you checked. I would ask in the same tone as I would about a calculation in a spreadsheet or an analysis prepared by a service provider.

I would also show my own uses. Describing a failed attempt helps a lot. A response can look flawless and contain a glaring error. Saying so in front of the team gives everyone permission to show their attempts before a problem reaches a client.

Giving time to people who are already experimenting

In every team, a few people are already testing new tools. They quickly spot an interesting use. They also see the current limitations. I would give them an ambassador role, with time allocated in their workload.

An ambassador can give a short demonstration using a real case, then record what worked so others can try it. Their feedback can also show that a tool wastes time in a particular situation. This prevents a lot of unnecessary enthusiasm.

The Work Trend Index observed that advanced AI users were 53 % more likely to receive encouragement from their leadership to rethink their work with AI. They also had greater access to training tailored to their role.

I see this as an interesting indication for managers. Familiarity grows when people feel they are allowed to experiment during working hours. I would set aside time for this knowledge sharing by making it part of the team’s organization.

Choosing uses according to risk

When I ask myself whether AI belongs in a task, I first look at the consequences of an incorrect response.

A poor rephrasing can be corrected in a few minutes. An error in a numerical analysis or a recommendation sent to a client requires much more rigorous checking. The sources then need to be traced, the calculations redone or a second review requested.

For a hiring decision, a promotion, a diagnosis or legal advice, the person responsible must be able to explain their decision in their own name. They can use AI to prepare the case. They remain responsible for what is decided.

This need for vigilance is illustrated very clearly in an experiment involving 758 BCG consultants. Participants equipped with GPT-4 completed certain tasks more than 25 % faster, with quality rated more than 40 % above the control group. On a task outside the model’s area of competence, AI users were 19 percentage points less likely to give the correct answer.

This study interests me because it forces us to talk about a specific task. Trust in AI is decided case by case, with appropriate verification.

Keeping decisions human

My principle fits into one sentence. The decision remains human.

The person delivering a piece of work must be able to explain what they asked the machine to do. They must also explain why they trust the result. They retain control over the expected standard of evidence and the possible consequences.

Verification depends on the work performed. A factual summary requires going back to the sources. Code requires tests, along with a thorough review. When verification seems impossible, it is better to say so clearly.

Another study covering 5 179 customer support agents measured an average productivity gain of 14 % with a generative assistant. The gain reached 34 % among novice or lower-performing agents, while it was small among the most experienced.

These differences matter a great deal to a manager. An identical productivity target for everyone is a poor fit for what the study observed. A person’s starting level changes the result. So does the nature of their work.

Saying what happens to the time saved

The time saved through AI must be discussed clearly.

The Slack Workforce Lab reported that employees feared seeing every minute saved turn into additional work. This fear can push a team to hide the tools that save them time.

I prefer to allocate some of that gain to work quality first. We can review a case more carefully or revisit an issue that has been set aside for too long. The rest depends on the team’s circumstances and can help absorb a busy period.

An increase in volume can help at certain times. The manager can then clearly explain how this time is being used and listen to the team’s feedback.

I expect a good manager to know how to discuss actual uses with their team and listen to the difficulties they encounter. When someone finds a good method, the manager gives them time to share it.

When an important result has been produced with AI, the manager asks how it was checked. Even if the response looks flawless at first glance, one question remains open. What evidence allows us to trust it.

AI is already at work. A manager can start by asking their team how they actually use it.

Sources

• Slack Workforce Index, fall 2024. Global survey of more than 17 000 office workers.
• Microsoft and LinkedIn, Work Trend Index 2024. Survey of 31 000 people in 31 countries, supplemented by usage data.
• Dell’Acqua et al., Navigating the Jagged Technological Frontier . Preregistered experiment conducted with 758 BCG consultants.
• Brynjolfsson, Li and Raymond, Generative AI at Work . Study of the deployment of a generative assistant among 5 179 customer support agents.