SAP Connect 2025 - inside an early Ask my Payslip rollout: how PostNL balances fast AI adoption with risk management

SAP Connect 2025 brought line of business leaders together, from across the rebranded Business Suite (on-demand replays are available).

It's hard to disagree with SAP's forward strategy for customers, where data, apps, and embedded AI spark each other, as SAP brings end-to-end visibility to bear. Today's business realities demand no less - the trick, of course, is in the execution: 

• Can SAP justify the value across these customer investments, given that much of the vision involves licensing products beyond S/4 ERP?
• Can SAP partners support real business model change, and move past custom projects that add to a customer's legacy code base, instead of paring it down?
• Can SAP make good on its commitment to integrate third party software (and agents?) in the end-to-end view, or will SAP downplay these necessary integrations? 

How will innovation be consumed by customers? SAP responds

SAP Connect's ambitious scope is the right place for these questions, as I wrote in SAP Connect 2025 - can we (finally) break the silos that block end-to-end thinking? Heartland Dental says yes. But I had two more goals for the show: 

1. Hear from customers that have embarked on SAP generative/agentic AI projects, and hopefully document early lessons.
2. Revisit my tougher questions for SAP, not the least of which is access to new innovation for customers on older SAP releases. Too often in the past, SAP's messaging implied that customers need to embark on a RISE/cloud upgrade, if they want to quality for new features.

Before I get to AI customer highlights, let's rehash my pesky question to SAP executive board members during the Q/A: "The problem that I see in the install base: when it feels overwhelming to say, 'I have to move to private cloud first before I can start taking advantage of these things. And then you have some very nimble AI shops that want to go into those customers right now and get started on AI today. Muhammad, you just criticized why that would be problematic [without the data context provided by SAP], but it's still appealing, right? Even if I start on a big SAP cloud project, I still have to go to my board and say, 'Here's what I did this month'... Are you going to be able to go to those older customers and say, 'Here's ways of engaging right now, build a cool app right now - and then you can still have your big [SAP cloud] project that's unfolding over time? '" 

Alam's response emphasized BDC and BTP, regardless of ERP release: 

That was the design principle with BDC, particularly with our BW and ECC install base, we said, hey, BDC is available to all. Of course, with private cloud, you get more things. The value is higher if you're in. But we didn't want to stop the value from accruing as you work, and as you were in your journey. 

More importantly, on BTP, with Joule Studio, you can go do exactly that as well. And we're looking to see how we can even make that simpler for customers to go, because we don't want to leave them behind. 

Alam also noted that making SAP cloud migrations easier is a key piece of this topic: 

We're also now looking at investing heavily to say, 'Hey, how do we help the migration process to be a lot quicker?' Because that process is ripe for disruption as well, particularly with the agentic development tools that are out there as well. We need to be able to deliver that for our customers.

Why PostNL is an early Ask my Payslip adopter

One customer that seems to have found the right innovation/risk balance is PostNL, the national mail and parcel service of the Netherlands. PostNL is a privatized e-commerce and courier company with a government-mandated role to provide universal mail and parcel delivery service to the Netherlands. The company also offers logistics services, such as delivery within the Benelux region. 

At SAP Connect, I spoke with Jack Naudé, SAP HCM Solution Consultant at PostNL, about PostNL's progress adopting SAP's "Ask my Payslip" AI solution (PostNL is one of the early adopters, with plans to adopt this at scale). Currently, Naudé's team is leading up a vigorous testing phase with a smaller group of users, though it's important to note: this testing phase is still live in production, an important characteristic of most AI projects (AI with live data can behave differently, so at some point, you need to get AI solutions live, and start iterating). 

At the end of the October, "Ask my Payslip", which Naudé calls "Explain my pay statement," will be rolled out to 300 users, with the hopes of a full rollout to all 32,000+ PostNL employees in the next four months. 

I take a hard look at the unique pros/cons of gen AI projects - and the potential costs of the technology. So when I run into an ambitious AI implementation, I always want to know: why? What problems can this uniquely solve? Naudé's answer: 

1. The human service desk isn't always available, whereas Ask my Payslip will be available 24/7.
2. People don't necessarily want to talk with other humans about sensitive pay/compensation questions, and
3. A controlled AI solution from SAP can generate trustworthy, explainable results, pulling from PostNL's own payroll data. 

Naudé also pointed out that human service reps can make mistakes also. He firmly believes, based on his testing, that Ask my Payslip is going to be highly accurate and reliable. After speaking with the service department, Naudé' says the business case was clearly there: 

One of the main ticketing flows is regarding payslips. Up to this stage, we've already had 7,500 requests into anything related to pay slips, and we expect about 10,000 by the end of the year. So if we could have a reduction in that - if it's just down by 50 percent or 80 percent, that will be a huge cost savings.

Naudé's team is prepping thoroughly. Two main goals: make sure the source documentation is airtight, and make sure those using the system will be able to get what they need. Naudé explains: 

We're rolling out, by end of the month, to about 300 people in the company. We were actually surprised by the speed of which it was enabled - and now we have to ensure that the documentation behind this is factually correct, to ensure that our diverse workforce gets the information they want, on their level. We have blue collar and white collar in the system, and everybody has a different way of communicating to the system. They expect their results being communicated to them on a level they can understand and interpret. So we're putting a lot of effort into getting that correct now.

Naudé is confident in the technology. As I've covered in detail, SAP's approach to AI is tied to providing the appropriate models with the right data context (via RAG, tool/database calls, etc) - with the compliance/privacy vigor you would expect from a European software vendor, one that is well-versed with tech regulations e.g. GDPR, etc. Naudé's pre-launch prep is therefore about data and system interaction: 

Technology-wise, we can do it today, but we want to be really sure the way that we present information is correct, not just because now it's factual information, but now I've got to break that down into the language part of it, and the perception of people on the language... We had some documentation before that was split on different departments and levels, and we're just incorporating all that information now into the system.

A gentle debate - on LLM agents, "faithfulness," and sourcing factual data for users

I thought about leaving this out, but I think it's useful for customers to consider a minor disagreement Naudé and I have about the technology. Naudé believes that with the architecture SAP is providing here, as long as the context data is high quality, there will be no hallucinations, and all the information will be factual to the context of the question the employee asks. 

However, my research indicates that LLMs still like to revert to their own training data (aka parametric knowledge) on occasion - yes, even when strictly instructed to use the "context knowledge" (which can be tracked as a "context adherence failure," or lack of "faithfulness" to the RAG context). 

The good news for this particular use case: I doubt this will happen often, as LLM agents are unlikely to have much existing training on the types of payroll questions being asked. (Sidenote: of course, the reverse can also be true, and LLMs can be unduly influenced by context windows. In the wild, this can introduce security risks, but in a controlled environment like this PostNL, it's exactly what you want to happen). 

When it comes to LLMs sometimes pulling the wrong factual information, Naudé and I agree: this might happen at times - including LLMs pulling something from the wrong pay period, for example (time and chronology are not where LLMs excel). But here, we are on firmer ground, because this type of scenario involves information being pulled from context.

In other words, the LLM is behaving as this scenario intends, e.g. pulling from context, but perhaps the context window is not serving up the exact right information from the database. But as Naudé correctly asserts, it is still serving up factual data from validated systems (one example of this is a data ranking issue, which can often be solved by re-ranking, or other "context engineering" techniques. Once the data is 'ranked' properly, the LLM serves up the best answer, e.g. pulling from the most likely pay period for the query). This can be improved as we go: any data provided via context/tool calls allows for the sourcing of that information, to show where it came from. I believe that as Naudé's team bears down, they will get into the 95 to 99 percent output accuracy range. Whether there is hallucination or not is semantical; accuracy for the user is the most important thing. 

Naudé's work on testing LLM interactions will also help here, as his team can propose the best prompts, and even embed some of them automatically. In this case, if users have a complex issue Ask my Payslip cannot answer, they will be notified to contact the human support teams. 

Okay, geeky discussion over. To be fair, I thought Naudé prevailed in our back-and-forth, with this comment: 

It's mainly for the initial contact, right? Say we now had 7500 tickets. If you can service 80% on an initial contact, and you can have a first contact with AI or a machine, that's already a huge savings, right? From that, if you have 5% that comes back with a deeper question or bigger problem, that's still a massive saving.

Hard to argue that one. Getting this right is mostly about use case design - and, as Naudé says, quality data pipelines.

My take - balancing AI innovation and risk is no small thing

The only downside I see to these kinds of AI efficiency plays? if you're not careful, you will create a culture of job loss fear, and lose much of the gain you aspire to, via deteriorating employee morale. But that's an issue Naudé and team have worked on as well. As he explained to me, PostNL has a history of pursuing automations, but they also work hard to redeploy employees who want to stay on into different roles: 

Employees that think if there's an agent, what's happening to my job? There, hopefully we can re-utilize them in other departments, right? And that's already happening to a certain extent. We also had a process where we just ran payroll, and then we started implementing payroll control center. There was a lot of discussion about reporting. There were people just running report after report, and suddenly they were made redundant on the reporting that they were doing. But then we could utilize their deep knowledge of the process in other sectors.

That sets the right tone for change - and the change won't stop anytime soon. PostNL uses virtually all of the SuccessFactors solutions. The transition to a skills-driven organization on SuccessFactors is well underway, though Naudé says there is plenty of organizational change still ahead. He sees "Explain my pay" as just the next step in their AI pursuits: 

Looking at the offering they have now with the AI agents being delivered, we are the pioneers of an age where everything is going to be different. What we have now with Explain my pay statement is just a teaser into the possibilities... I foresee that by the end of next year, it's going to be totally different. There's going to be a whole interaction section with AI, and all different contact opportunities there, and service deliveries as well.

This fall, I'm seeing a difference with customers and generative AI. Though some of the use cases are still at smaller scale, I'm hearing a clearer philosophy of how to approach AI projects. The exact method varies by customer - but I'm hearing much clearer narratives of how customers define "AI first" mandates, or talk to employees about change and job transitions. I don't always agree with these narratives, but that's not important at all. What matters is the organizational clarity. 

PostNL is a particularly strong example. Balancing an early adopter mindset with careful risk management/use case design takes project savvy. Naudé: 

Well, we've been pushing the envelope for quite a while. We were one of the first companies that went from on-prem payroll to Employee Central Payroll...  We're at this stage now where we say the benefit of having the new technology outweighs the caution, to a certain extent. The backup was there. We had the fine line. We had the the good feeling that it is going to be a success. We believed in what we did. We saw the factual proof that it is going to work, and that's why we did it.

Naudé makes it sound easy, but it is anything but - another worthwhile lesson. Stay tuned, more SAP Connect coverage coming, with podcasts up next.