To find out what that looks like on the ground, we spoke with two people helping to lead that shift at PwC: Andreas Braun, Managing Director for Artificial Intelligence, and Julien Jacqué, Partner for Digital Transformations and GenAI Champion for Industry

 

Agentic AI is everywhere in the headlines. What makes it different from the chatbots we already know?

A.B: Agentic AI is far more advanced in terms of autonomy than the typical chatbot. A chatbot answers a question. An agentic system can plan a sequence of steps, execute that plan, access  different tools and data sources, and produce real business outcomes. The progress has been dramatic. One year ago, AI could complete a task that would normally take a human about 15 minutes. Today, it already automates tasks that take four hours or more.

We see companies moving through three stages when working towards agentic AI. Most start at stage one, where employees use AI assistants like Microsoft Copilot or other enterprise AI solutions to speed up everyday work, such as drafting, summarising, or preparing meetings. It is the entry point to builds the skills, habits, and confidence of teams to use AI solutions.

The real step change comes at stage two, where you create dedicated AI agents that take over specific, labour-intensive tasks within a workflow. You are not just typing faster, you are delegating the least value-added work to your AI. This is where the most advanced organisations in Luxembourg are right now. 

Stage three is transformative: you redesign the entire process around what AI does well. Humans set direction; agents run the execution. The quality of the process is assured through this human-in-theloop approach. This is what many organisations aspire to, but it has not arrived in Luxembourg yet, and to be transparent, not really anywhere else.

J.J: I agree. Most Luxembourg companies today are firmly in between stage one and two. Copilot or other AI enterprise licences are easy to buy, and the productivity gains are immediate, better inbox management, faster document drafting, quicker meeting preparation. But over the last six to nine months, we have seen a strong push into stage two, particularly in customer-facing and support functions: customer service, marketing, finance, HR, IT. When stage two works well, you start to see hard advantages over competitors who have not made the shift. Response times drop, sales cycles shorten, and forecasts become more accurate. Days sales outstanding come down. These are not theoretical gains, they show up in the numbers.

 

From the outside, Luxembourg still looks quite cautious… Would you agree? 

A.B: That is fair. The big disruption has not happened yet. But the pace is accelerating. The government has released its AI strategy including investment tax credits and programmes for  companies to create their AI strategy with support of dedicated consultants. Major IT vendors are also offering incentive schemes to speed up the adoption at their clients. The infrastructure is there. What is needed now is for industries to pick up the pace, because their global competitors already are.

J.J: A year ago, most of the conversations I had about AI were still exploratory. Companies wanted to understand the landscape, assess the options, and build internal alignment before committing. That was a reasonable approach at the time. What has changed is the pace. Today, those same companies are buying licences, running pilots, training teams, and bringing in firms like ours to deploy real use cases. There is genuine momentum, and the companies moving now are building advantages that will structurally change how they compete going forward in terms of speed, cost, and quality.

 

 

Case Study 1: PwC Luxembourg / ‘Client Zero ’

Can you give us a concrete example?

A.B: Well, we started with ourselves. PwC is our own ‘Client Zero.’ We gave everyone access to Copilot and let them experiment with clear boundaries, while also building the governance infrastructure for AI. As a regulated entity, we need to be fully in control of whatever AI does in our environment. That meant clear security access procedures, sovereign data hosting, and defined boundaries for what AI can and cannot do.

That governance-first approach is something I would recommend  to any Luxembourg company. The risk people worry about with any AI solution of accidentally accessing information it shouldn’t, can be controlled that way. 

We asked our employees and leadership, using our AI tools, what  ideas do you have that could genuinely transform how we deliver services? We collected over 100 ideas, many of which have been put into production already. We now have many active stage-two tools in production for individual teams, plus what we call ‘AAA agents’. Systems that have been thoroughly tested and usable by entire business units, like Advisory.

We also have our first real transformative wins. Tasks we previously outsourced are now handled end-to-end by a tool we built ourselves. It paid back the initial investment within a year. Whenever we solve a problem internally, we look at whether it can be transposed to clients. That is how our client work keeps getting better.

 

 

Case Study 2: Enterprise Copilot Deployment / Measuring the Business Case

What about implementing this for other organisations?

J.J: We supported a large organisation in Luxembourg in deploying Microsoft 365 Copilot across multiple departments. The headline result: users saved roughly an hour a day, about 180 hours per year per person, and unanimously recommended the tool. They also reported a noticeable improvement in the quality of their deliverables. How we got there was deliberately pragmatic. We started by identifying the most valuable daily scenarios by finding and synthesising information, drafting documents, role-specific workflows, and then upskilled a focused pilot group with small, structured sessions that included IT and HR representation. The pilot ran on real work, not demos. We captured feedback continuously and translated it into quick wins and a clear list of improvements, including opportunities better suited for Copilot Studio agents.

This company had already invested in a substantial number of licences and wanted a structured evaluation of the return. So we modelled what the time savings could mean in workload terms, assessed their ability to absorb more work without additional hires, and benchmarked the gains against licence costs. The business case was clear even at stage one.

But we also documented what limits value at scale, the part that is easy to overlook. Data access was uneven. A lot of content still sat on legacy file servers. Document organisation was inconsistent. These are not AI problems, they are data hygiene problems. If you do not address them, you hit a ceiling on what Copilot or any AI solution can do for you.

 

Does the client believe it was worth the investment?

A.B: In short, yes. The financial case was tangible even at stage one, and it only grows as you move to stage two. I would also emphasise that speed and quality are crucial non-financial KPIs. If an employee can suddenly handle more clients because of efficiency gains, that is when the business case is realised.

 

 

Case Study 3: Finance AI Agents / From Productivity to Decision Intelligence

Do you have a case where AI goes beyond general productivity – what you call ‘stage 2’?

J.J: Absolutely. We are supporting an international company in augmenting their finance team with specialised agents. Three examples.

First, a data quality agent for controllers. The team was working with production-cost files of over 5,000 lines – plants, products, markets – manually checking entries against master lists and  hunting for cost outliers. The kind of work that takes days to sanity-check. The agent now handles the data quality verification and flags anomalies automatically, cutting turnaround to a fraction of what it was.

Second, an insight agent for controlling. Instead of analysts spending hours manually hunting variances in budget-versus-forecast files and rewriting commentary, the agent highlights the key  movements and drivers, and packages them into a concise insight pack. The target: under 20 minutes from raw data to decision-ready insights, with a 25–35% reduction in back-and-forth refinement.

Third, a finance policy Q&A agent. This is deceptively simple but high impact. Finance teams constantly need quick answers on internal policies and procedures, and the typical approach is to dig through folders. We built a retrieval agent grounded on the organisation’s SharePoint policy repository. It answers based on what is actually documented, no guessing, no inventing rules, and with clear citations back to the source document. It cut search time from 20–30 minutes to under a minute. The value is not dramatic; it is cumulative. Multiply that friction removal across hundreds of queries per month and you free people for genuinely analytical work.

 

It sounds like firms already good at BI are halfway there? 

J.J: You would think so, but in practice most organisations are still in between. They have structured dashboards for certain reports, but alongside those, many workflows are still semi-manual: data extracted from one system, analysed in Excel, written up as a separate report. The process depends heavily on individual expertise and is rarely standardised.

That is exactly what makes these workflows ideal for agentic AI. Instead of waiting to rebuild everything as a dashboard, you deploy agents to analyse the data and surface insights now. And even when you eventually do have a fully digital workflow, there is a fundamental difference between clicking through filters and receiving a structured summary that tells you: here are your top growth markets, your biggest revenue declines, and your most significant cost variances. That is the shift from exploration to interpretation.

A.B: And consider the traditional process for adding a new KPI. A business user requests it, IT queues it, and weeks later the updated dashboard arrives. With AI, a finance professional can simply ask: ‘Show me margin evolution by product category, adjusted for raw material fluctuations, over the last six quarters.’ The system connects the data sources, runs the calculations, and delivers the answer. That is the direction enterprise applications and every major BI platform are heading towards; natural language queries embedded directly in the environment.

 

How do companies ensure AI is not making things up?

A.B: A very common question, and the answer has changed significantly from the early chatbot era. Modern enterprise AI does not guess. When it analyses financial data, it generates small pieces of code that query databases via APIs and perform deterministic calculations on the retrieved data. The queries can be logged, the code can be reviewed, and the outputs can be replicated. In many cases, this is actually more transparent than a manual Excel process where the logic is often not adequately documented. That said, safeguards remain essential. AI should not have unrestricted access to corporate tools. Write permissions must be carefully controlled, role-based, and auditable and human oversight must be embedded in the workflow. AI is another pair of eyes, not an autonomous decision-maker.

J.J: I want to reinforce one point: AI adoption does not mean accountability disappears. In most deployments, every output is linked back to its data source, so users can trace exactly where the information comes from. But that transparency does not replace professional judgement. Part of successful adoption is teaching teams when to trust the output, when to verify it, and when to set it aside entirely. Not every task should be automated, and not every output should be accepted without scrutiny. 

 

 

Case Study 4: Customer -Facing AI / From FAQ Bot to Agentic Service

Tell us about the final case.

A.B: This one is external-facing, and it shows nicely how a chatbot can evolve step by step. The company started with a simple FAQ chatbot. Customers asked a question, and if the answer was in the database, they got a response. Limited, but a solid first step. Then they integrated an existing online service-finder tool into the chatbot. Instead of clicking through structured forms, customers could describe what they needed in plain language, and the AI would match them to the right service in the background. The third step was back-office automation: if the chatbot could not fully resolve an issue, it created a support ticket automatically, with all the collected information structured and transferred into the ticketing system. No re-entry required.

Where they are moving now is genuinely agentic. The AI does not just respond, it structures information and initiates processes. The next step is preparing draft emails and even commercial offers for a human agent to review and approve before sending. But here is the critical point. This chatbot uses a powerful European language model that can handle French, German and English fluently. The capabilities are impressive and that is why guardrails are crucial. You do not want a system this capable drifting into topics outside its mandate or answering random questions from the internet. Think of it as a 500-horsepower engine: extraordinary when properly steered, dangerous when not. The guardrails ensure the intelligence stays focused, controlled, and aligned with business needs.

Four companies, four different starting points, one consistent message: AI adoption in Luxembourg has moved from conversation to implementation. The gains are starting to become real and measurable, whether that is an hour saved per employee per day, a five-fold reduction in analysis turnaround, or a customer service operation that no longer requires anyone to re-enter data. The firms making progress share a common approach: they started with governance, proved value at small scale, and then expanded. For those still watching from the sidelines, it’s time to get into the game.