PERSPECTIVES

Straight thinking on AI , transformation and the future of work.

01 PERSPECTIVE - DIAGNOSTIC

Shadow AI is already inside your organisation. Now what?

Most leadership teams still treat AI adoption as a decision they are about to take. In practice, it has already been taken, by their employees.

People paste customer emails into public chatbots, summarise contracts with free tools and draft board papers with whatever model they signed up for privately. Not out of defiance, but because the tools are useful and the official route is slow or missing.

A ban may reduce some use. It can also push the remaining behaviour out of sight. Leaders then lose the visibility they need to assess the risk and understand why people bypass the official route.

The risk lies in both the tools and the absence of clear decisions about their use: who may use what, with which data, and who is accountable for the outcome. Confidential or personal data may be disclosed without authorisation.

Documents and decisions may rely on outputs nobody verified. Unclear ownership makes it harder to detect and address either problem.

There is a second question, and it is the more interesting one: why did people go around the official route in the first place?

Perhaps a two-day process now takes twenty minutes. Perhaps the approved tools create work instead of removing it.

Seen this way, Shadow AI is a compliance and security risk, but also a sensor: it shows where processes, systems or ways of working are falling behind. The answer is neither a ban nor looking the other way, but governance that enables: understand what is in use, assess the risk, offer safe alternatives, and scale what genuinely makes work better.

What leaders can do?

Find out what is really in use. Ask without blame. Combine confidential conversations and a short survey with proportionate technical checks, subject to data-protection and participation requirements. Each method reveals a different part of the picture.
Decide what is allowed. Keep it simple: what is free to use, what needs conditions, and what is off limits.
Offer a sanctioned alternative that is good enough that people prefer it over their private account.
Name an owner in the business, not only in IT. Shadow AI is a question of how work gets done. Executive management needs clear ownership, and the supervisory board needs visibility and oversight.
Involve employee representatives early. In co-determined companies, even monitoring AI use can trigger participation rights, and trust is what makes people tell you the truth.

A question to take back to your team: if you asked everyone tomorrow which AI tools they use for work, would they tell you? And what does your answer say about your organisation?

02 PERSPECTIVE - DIAGNOSTIC

The EU AI Act is not an IT project.

When the EU AI Act lands on a leadership agenda, it is often handed to IT or Legal. IT is asked to list the systems, Legal to interpret the requirements, and the topic is considered covered. Both are essential. But the Act also raises questions about how AI is used in business processes and who takes responsibility for that use.

That is why the key questions cannot be answered in the server room.

Is the tool used to rank applicants or to evaluate employees?
Does it influence decisions about customers?
Who decides that the output is good enough to act on? These are questions about business processes, and only the business can answer them.

Three situations most organisations will recognise:

1. The recruiting tool. A vendor adds candidate ranking to an applicant system. IT sees an upgrade; HR sees a selection decision. AI used to analyse applications or evaluate candidates can fall within the Act’s high-risk framework.

Whether such a feature counts as high-risk, and which duties and dates apply, has to be assessed for the specific use case. Where it does, oversight, monitoring and information for employees need to be prepared.

2. The helpful assistant in customer service. It drafts replies and suggests discounts. Nobody has defined who checks the output, which cases are excluded, or who answers when it gets it wrong.

3. The purchased service with AI inside. Many suppliers add AI to existing products. Without a business owner, the company does not even know where it has become a deployer of AI.

The real implementation challenge is not classification alone. It is accountability. Training needs to reflect the decisions people make with AI, not just the features they can operate.

What leaders can do?

Name a business owner for every AI use case, not only for the platform. The owner is accountable for purpose, quality and oversight.
Classify by use, not by tool. The same model can be harmless in one process and sensitive in another.
Start with a simple inventory, including AI inside purchased software. Ask the departments, not only IT.
Define what human oversight means in practice: who checks what, and with what authority to overrule.
Build AI literacy where AI is used, and involve Legal, HR, Data Protection and employee representatives from the start.

A question to take back to your team: if a regulator asked tomorrow who is accountable for your most sensitive AI use case, would the answer be a person or a department?

03 PERSPECTIVE - ACADEMY

Prompting is becoming easier. Judgement isn’t.

Prompting is useful, but it is only part of effective AI use. A clear request can improve an answer; it cannot establish whether the answer is supported, appropriate or safe to act on. As tools become easier to use, the human contribution remains demanding: choosing the task, checking the evidence and taking responsibility for the result.

Three situations show where the real skill sits:

The variance commentary. A controller asks AI to explain why costs rose. The text reads convincingly, but one figure is wrong and the cause is invented. Someone who knows the business catches it. Someone who only knows prompting forwards it.
The performance feedback. A manager lets AI draft the annual review. The wording is polished and generic. The employee notices immediately that nobody has really looked at their work.
The supplier contract. AI summarises forty pages in a minute. Whether the summary can be trusted for a decision depends on whether anyone checked the clauses that matter.

The skill behind all three is judgement, and it has four parts: choosing the right task for AI, giving it enough context and structure, verifying the result in proportion to the risk, and owning the outcome.

Whoever signs, owns it, whether a human or a model wrote the first draft.

What leaders can do?

Train on your own cases, including real mistakes. Use approved environments and appropriately anonymised or authorised material.
Set verification standards by risk. A brainstorming note needs a glance. A board paper or a customer commitment needs a documented check.
Make checking visible and normal, so that “I verified this” is a sign of quality, not of distrust.
Let leaders go first, including showing where they got it wrong.

A question to take back to your team: who in your organisation is trained to spot a convincing but wrong AI answer in their own field?

04 PERSPECTIVE - DIAGNOSTIC / ACADEMY

AI has a footprint. Not only an environmental one.

Companies with mature sustainability reporting track waste, water, emissions, safety and supplier standards. Yet an AI proposal may still reach the leadership team with a much narrower headline: “this saves us half an FTE.” That comparison leaves important costs and assumptions unexamined.

AI relies on data-centre infrastructure that uses electricity and, depending on cooling and electricity generation, water. The footprint varies by model, workload, location and system design.

For a company buying cloud or software services, attributing a share of that footprint to a specific use case can be difficult. Provider data and clearly labelled estimates can nevertheless improve the investment decision.

The question is whether the value justifies the resources required, compared with feasible alternatives. Apply the investment discipline already used elsewhere in the business.

Three practical examples that most organisations will recognise:

The chatbot for routine questions. Compare a large model with a maintained FAQ, a search function or a smaller model. Which meets the service need with acceptable accuracy, cost and resource use?
The forecast with a simpler alternative. Compare a complex AI approach with a conventional statistical method on the same data. Consider accuracy, explainability, maintenance and resource use before selecting the solution.
The summary nobody reads. Meeting notes, long documents and email threads are summarised automatically by default, because it is easy, not because someone needs the result. Ask who actually uses the output before scaling it.

There is also a capability cost to consider. When AI takes over the first draft, the first analysis and the first structure, junior colleagues may lose practice that builds judgement.

This is a risk to manage, not an inevitable outcome: AI can also support learning when people test its answers, examine the evidence and receive feedback.

What leaders can do?

Ask the use-case question first. Does this problem need AI, or is a simpler tool enough? Document the answer.
Put the footprint on the same page as the benefit. Even an estimate, based on provider data, is better than silence. Ask your cloud and software providers what they can report.
Anchor it in the ESG framework you already have, rather than creating a new one. Compute and cooling belong in the same conversation as waste and water.
Protect the learning path. Decide deliberately which tasks juniors should still do first without AI.

A question to take back to your team: if the AI initiative in your next board paper had to show its energy and skills cost next to its FTE saving, would the business case still hold?

05 PERSPECTIVE - ADVISORY

Work needs redesign, not another cost-cutting programme.

When performance comes under pressure, organisations reach for familiar levers:

reduce headcount, freeze hiring, cut external spend, consolidate functions. The numbers improve, at least for a while. But many organisations emerge from cost programmes with fewer people and almost exactly the same work. The same approvals, the same reports, the same manual workarounds between systems.

That is not transformation. It is compression.

The more useful question is not where to cut, but what work should exist at all. Which decisions genuinely require five approval levels? Which reports are produced because somebody once asked for them? And where are highly qualified people compensating for fragmented systems and unclear accountabilities?

AI raises the stakes. Layered on top of processes that were never designed for it, AI simply automates the waste faster. Used properly, it lets organisations reconsider how work is organised in the first place. That requires looking at processes, decision rights, skills, systems and governance together — and involving the people who know where work gets stuck.

The measure of success is different, too: capacity released and decisions accelerated, not just headcount removed.

Redesign of this kind rarely fails on the idea. It fails on decision rights, co-determination and governance.

When AI is introduced, an early question in many boardrooms is how many positions it could save. That question is legitimate, but it is premature without understanding the work. AI can automate tasks and, in some cases, replace roles. It also changes what remains: verification, exceptions and customer interaction may take on greater importance. A programme that counts heads before it understands tasks risks cutting in the wrong places.

Three examples:

Invoice processing. Automatic matching reduces routine handling. What remains may include exceptions and supplier disputes requiring experienced judgement. A team designed around the old task mix can end up with excess routine capacity and insufficient expertise for the harder cases.
Customer service. When AI drafts the replies, the agent’s role can shift from typing to deciding: owning complex cases, judging when to deviate and keeping the relationship. That calls for different training and different performance measures.
Maintenance planning in a plant. When AI predicts failures and suggests schedules, the planner’s work can move from building the schedule to challenging it with knowledge of the machines. Experienced planners may become more valuable, and those close to retirement take that knowledge with them.

A recurring risk in cost-cutting programmes is that capacity is removed before the future process is understood. Experience leaves with the people, and the remaining team absorbs the rework. The planned saving should therefore be tested against actual workload, quality and execution risk.

A simple calculation shows why the capacity question matters.

Take a mechanical engineering company with eight sales colleagues who spend about 20% of their time on research, quotations and follow-up (illustrative assumptions). If AI-supported workflows reduce that administrative effort by 35%, the released capacity is 8 × 20% × 35%, or roughly 0.56 FTE. That is not a position to cut. It is a decision: should the capacity go into customer-facing work and growth, or into cost reduction? Both can be legitimate, but they lead to very different programmes.

What leaders can do?

Map tasks, not jobs. Which tasks shrink, which grow, which are new?
Decide what the freed time is for (quality, service, growth, development) before the headcount discussion starts.
Redesign one process end to end instead of introducing tools across the board.
Involve employees and their representatives from the design stage, because they know where the real work happens.
Measure outcomes, such as quality, speed and customer satisfaction, not only hours saved.
Don’t start with the headcount. Start with the work.

A question to take back to your team: if your AI business case were not allowed to count headcount, what would it be about?

06 PERSPECTIVE - ADVISORY

Why intelligent organisations still make bad decisions.

Organisations can be full of intelligent, experienced people and still make poor decisions. More data sometimes helps. But it does not resolve unclear ownership, unchallenged assumptions or a timetable that forces a choice before the evidence has been assessed.

AI and analytics make this more visible, not less. They increase the volume of analysis, and volume is not clarity.

Three patterns most leaders have seen:

The investment case with friendly assumptions. The calculation is correct, but the assumptions were chosen by the people who want the project. Nobody’s role is to question them, so nobody does.
The decision between headquarters and the site. The data on a plant restructuring is complete. What is unclear is who decides what: the group on the portfolio, the site on execution. The decision drifts, and by the time it is taken, the options have narrowed.
The steering committee of fifteen. Everyone is informed, so everyone feels involved, and nobody decides. The real decision happens in a side conversation afterwards.

Better decisions require usable evidence, clear authority and deliberate challenge. A dashboard can support those habits; it cannot create them.

What leaders can do?

Name the decision owner before the analysis begins. Where a board or committee must decide collectively, clarify its mandate and who owns the recommendation, evidence and follow-through.
Separate recommending from deciding, and make clear who does which.
Assign a challenger. Run a pre-mortem: assume the decision failed in two years and ask why.
Write down the key assumptions and what evidence would change your mind.
Review the decision after six to twelve months, to learn from the outcome rather than defend it.

A question to take back to your team: think of the last major decision in your organisation. Can you name who owned it, who challenged it, and what would have changed the outcome?

07 PERSPECTIVE - ADVISORY

Everyone talks about talent shortages. So why is hiring still so inefficient?

Companies say they cannot find talent, and candidates say they cannot get a reply. Both are often true, which suggests that the bottleneck is not only the labour market. It is how organisations describe work, evaluate people and make decisions.

Four examples:

The unicorn profile. A job ad lists twelve must-have requirements, copied from the last person in the role. Only three of them matter for the job. The perfect candidate does not exist, and good ones do not apply.
The slow process. Five interview rounds over three months, each step waiting for someone’s calendar. The best candidate has accepted another offer by week four.
The like-me decision. Hiring managers prefer people who resemble them, without noticing. Unstructured interviews tend to favour similarity over performance.
The talent next door. Roles are filled externally while people inside, with relevant skills but a different title, are never considered.

Technology can speed up administration, but it will not repair a profile nobody has questioned. AI used to analyse applications or evaluate candidates can fall within the EU AI Act’s high-risk framework; assess the use case rather than assuming every recruiting feature has the same status.

What leaders can do?

Define the work before the profile. What should this person have achieved after twelve months?
Cut the must-haves to those that truly decide success, and develop the rest.
Set a decision timeline and an owner. Speed is part of your employer brand.
Structure the interviews, with the same questions and criteria for every candidate.
Look inside first, and make internal skills visible.
Use AI for administration, such as scheduling and drafting, and be cautious about letting it select people.

A question to take back to your team: how long does your best candidate wait between first contact and a decision, and what does that tell them about you?

08 PERSPECTIVE - ACADEMY

AI transformation without co-determination? A risky shortcut.

In many companies, the works council appears on the AI project plan late, as an approval to obtain shortly before go-live. That can be a costly shortcut.

In Germany, technical systems capable of monitoring employee behaviour or performance can engage co-determination rights under section 87(1)(6) of the Works Constitution Act (BetrVG).

Further provisions on informing the works council about AI use and on its right to call in an AI expert point in the same direction. The assessment depends on the functionality and context. A rollout that skips the conversation may move it to the point where changing the design is most expensive.

Three patterns many organisations will recognise:

The assistant rollout. A company introduces an AI assistant for all staff. The usage logs would allow analysis by person. Nobody asked, the works council notices, and the rollout may be put on hold until an agreement exists, and weeks of goodwill are lost.
The pilot that cannot scale. A pilot in one department works well. For the wider rollout, questions on data, monitoring and training come up for the first time, and the sponsor has no answers.
The framework agreement first. A company agrees principles with its works council before the first rollout: purpose, no individual performance evaluation from AI data, transparency, training. Later projects move faster, because the basic questions are settled.

The legal side is one reason to involve employee representatives early. The practical reason is stronger: they know where work really happens and where a tool will fail. Trust also decides whether people use a tool honestly or work around it. Sustainable AI implementation requires dialogue, trust and structured participation.

What leaders can do?

Involve the works council at concept stage, before tools are selected.
Agree a framework on AI covering purpose, data access, transparency, training and safeguards against individual performance monitoring. Assess the requirements of each use case within that framework.
Be open about what the tool logs, and who can see it.
Offer joint learning, so that all sides assess AI from the same knowledge base.
Treat the pilot as a shared experiment, with clear criteria for scaling or stopping.

A question to take back to your team: at what stage of your last AI project did the works council first hear about it, and what would have been different if it had been the first conversation?

09 PERSPECTIVE - DIAGNOSTIC

Does this problem really need AI?

Most AI initiatives start with a tool and look for a place to use it. A better starting point is the P&L: where are the biggest economic levers today, and which of them have anything to do with AI? Some problems are not AI problems at all.

Take a quotation that needs seven days. A closer look shows four approval levels, poor master data, an unclear pricing logic, Excel interfaces between systems and nobody who owns the whole process. An AI tool on top would produce faster drafts for a slow process. The better recommendation may be: fix the process first, and do not buy AI.

That conclusion is not a failure of an AI assessment. It is among its most valuable results, because it saves the investment and points to the real lever.

What leaders can do?

Start with the economics, not the technology.
Look at revenue by product and customer, gross margin, personnel cost, scrap and rework, lead times and working capital. Ask where the largest levers are, not where AI could be used.
Map the work in five to seven core processes, from sales to finance.
Separate repetitive work, knowledge work, analysis, decisions, coordination, documentation and manual handovers.
Turn findings into hypotheses with numbers. Eight sales colleagues who spend 20% of their time on research, quotations and follow-up, with a 35% improvement, release about 0.56 FTE (8 × 20% × 35%). The decision is then whether that capacity goes into growth or into cost reduction.
Rate every use case on four dimensions: P&L impact, feasibility, risk and time to value. A table of five to eight use cases is something a CEO understands immediately.
Be ready to conclude that AI is not the answer, and keep the result short: where AI matters, where it does not, the P&L potential as a range, the impact on work, the governance implications and a 100-day plan.

A question to take back to your team: which of your current AI ideas would survive if the first question were not “where can we use AI?” but “where are our biggest economic levers?”

10 PERSPECTIVE - ACADEMY

How to lead a hybrid team of people and AI.

Many teams already work in a hybrid form that has nothing to do with the home office: people and AI systems share the work.

An assistant drafts the report, an agent prepares the order, a model proposes the forecast. Whether to call that a colleague is a matter of taste.

For leaders, the more important point is that it does not behave like one. It does not carry responsibility, it often gives no signal when it is unsure, and it can be wrong with great confidence.

Leading such a team is therefore less a question of technology than of three questions that were often left implicit:

what is delegated,
who checks, and
who answers for the result.

Three situations most leaders will recognise:

The delegated first draft. In a finance team, AI prepares the monthly commentary. One analyst checks every figure, another forwards it after a glance. Same tool, same output, different quality. The difference is not skill with the tool, but whether the team has agreed what “checked” means.
The silent gap. A young colleague uses AI for the first analysis and delivers within an hour. The manager is pleased and does not notice that nobody can explain how the result was reached. Speed has hidden a gap in understanding.
The team that stops arguing. When a plausible answer arrives without a name attached, people may stop challenging it. Disagreement is a source of quality in good teams, and with AI in the team it has to be protected deliberately.

A further point is specific to teams with several generations. Younger colleagues often move faster with the tools; experienced colleagues are better at judging whether the result makes sense. Neither group has the full picture, and a leader who turns this into a contest loses both.

What leaders can do?

Define what is delegated to AI and what stays with people. Write it down per task or process, and review it as the tools improve.
Make verification part of the job: who checks what, how, and what “checked” means.
Keep people in the learning loop. Let them do the first version themselves in selected tasks, and ask them to explain results, not only to deliver them.
Protect challenge. Ask routinely what would make this answer wrong, and recognise the colleague who finds the error.
Pair experience with fluency. Mixed-age tandems on concrete tasks, where both sides teach, work better than abstract mentoring.
Keep accountability with a person. A name stands behind every result, regardless of who or what wrote the first draft.

A question to take back to your team: if AI answered your team’s questions with the same confidence for the next month, would anyone still say “I think this is wrong”?