Nobody has an edge in AI - except big tech

K5 Future Retail Conference Berlin 2026 | Image based on K5 Mediathek recording material.

At K5, I kept seeing the same promise in different packaging: an AI agent that searches, advises, decides, or buys. Many of the demos were impressive. Yet once I looked under the hood, one uncomfortable question kept coming back: if the intelligence is rented through an API and the solution can be copied, where is the competitive advantage? In my opinion, it is not in the agent. It is in proprietary data and the knowledge of how your business actually works.

A view from the K5 stage

In June, thanks to Emporix, I had the pleasure of speaking at the K5 Future Retail Conference in Berlin. Together with Mark Holenstein, we presented the joint journey of Emporix and Vaillant Group towards an agentic future.

Our presentation was titled 150 Years of Evolution. How Vaillant Adapts Its Core Values for the Age of Agentic Commerce. We talked about what happens when the buyer is no longer a human browsing a website, but an agent searching, comparing, advising, and eventually making decisions on the customer's behalf. The recording is available in the K5 Mediathek.

I enjoyed the presentation, but one thought followed me off the stage. Nearly every offer I had analysed included some form of an AI agent, advisor, assistant, or copilot. Agentic everything.

The names differed. The demos differed. The promise was usually the same: let AI understand the customer and do part of the work for them. When everybody offers a variation of the same capability, however, it stops being a differentiator. It becomes the ticket to enter the race, not the ticket to win it.

So where is the edge?

The wrapper is not a moat

The majority of these solutions use off-the-shelf models. That is not a criticism. To be honest, I would probably make the same choice. Using a strong existing model is usually the most sensible engineering and business decision. Training a foundation model is expensive, risky, and unnecessary for almost every company.

The problem starts when a company treats access to that model as its competitive advantage.

Your competitors can use the same API. They can apply a similar system prompt, connect comparable tools, and reproduce the visible part of the experience. For a well-funded big-tech company, copying such a solution is even easier.

A thin wrapper over a model can still become a useful product. It just is not automatically a defensible one. A prompt is not a moat. A chat interface is not a moat. Calling a workflow an agent is definitely not a moat.

Today, the large technology companies are still fighting over the biggest part of the market: models, infrastructure, cloud capacity, developer ecosystems, and distribution. My assumption is that once this race becomes more settled and the winners become clearer, they will move further up the value chain. The smaller pieces of the cake they ignore today may suddenly become attractive.

If your entire product is a thin layer over somebody else's model, what will remain when the model provider adds the same feature?

This does not mean that companies cannot create value with AI. It means they need to be honest about which layer of the solution creates that value.

Algolia gets the positioning right

Algolia was the positive exception in my analysis. This may sound like a strange compliment, but I liked what they did not claim. They did not try to sell me another mysteriously intelligent black box. And yes, I like their approach - especially how professionally they explain the underlying concepts.

Their proposition does not depend on pretending that they have discovered a uniquely intelligent model. Algolia's Agent Studio lets customers connect their preferred LLM provider, choose the model, and keep control over data governance and model costs. Their tooling also makes the prompting and retrieval behavior configurable. Even practical details, such as reducing how much context is sent to the model, are explained in terms of the trade-off between quality and token cost.

In other words, Algolia does not position a proprietary foundation model as its edge - at least not in this field. Their value lies in search, retrieval, relevance, and orchestration. The model remains an interchangeable component.

I find that far more convincing than another serving of secret AI sauce. It is also a useful clue as to where the customer's own advantage can be found.

Your data is the advantage

In the K5 presentation, I argued that the most defensible AI-specific advantage for most companies is proprietary data - especially data describing how their business actually operates. I still believe that.

A general-purpose model knows a lot about the world, but it does not know your customer relationships, product constraints, service history, negotiated contracts, or the outcomes of past decisions. It also does not know how work actually gets done inside your company. Your competitors should not have access to this context either.

There is an important catch. Proprietary is not a synonym for valuable. Most companies own plenty of proprietary rubbish. The data has to be:

• Relevant to the problem
• Clean and accessible
• Governed and used legally
• Connected to current business processes
• Improved through real feedback loops

As a product person, I understand the temptation. A friendly chatbot looks good on a slide. Cleaning data, agreeing on its meaning, and connecting it across organizational silos does not. Unfortunately, the less glamorous work is also where copying becomes much harder. A competitor may reproduce your user interface in a few weeks, but they cannot reproduce years of high-quality operational data and the organizational knowledge behind it.

The agent is what looks good in a demo. The context is what makes it work on a rainy Monday morning, when a real customer arrives with a case the demo never covered.

Processes are proprietary data in motion

For me, one particularly valuable type of context is hidden in business processes.

Anyone who has worked in a large company knows that the official process diagram is, at best, an optimistic version of reality. Most companies openly describe what they sell, but rarely explain in detail how they operate. The real process is the path an order, invoice, service request, or approval actually takes across systems and teams - including all the exceptions, manual decisions, workarounds, and dependencies accumulated over the years.

This knowledge is difficult to copy because much of it is never written down. Nevertheless, the process leaves a digital footprint. Orders are created, approvals are granted, deliveries are delayed, invoices are corrected, and payments are received. Every such event adds another piece to the history of how the company works.

Celonis is a good example of a specialist building value in this area. Its process-mining approach uses event data from transactional systems to reconstruct how processes actually run, including their variants, bottlenecks, and deviations. Its object-centric approach can represent orders, invoices, deliveries, events, and their relationships instead of forcing a complex operation into one flat, linear view.

What I like about this example is that the result is more than another collection of documents for an LLM to search. It is structured process context: what happened, in which order, to which business objects, under which conditions, and with what outcome. Add business rules, KPIs, and organizational knowledge, and AI gains a much better foundation for reasoning about the operation.

An agent grounded in this context can move beyond answering What does our policy say? It can start addressing more valuable questions: Why is this order blocked? Which deviation caused the delay? What should happen next? Has a similar exception occurred before?

This is an area where a genuine edge can exist. Your competitor may use the same model and even the same process-mining platform, but they do not have your process history, exception patterns, business rules, or feedback from past decisions. The technology can be bought. The operational memory cannot be copied overnight.

A basic home-grown RAG is not a moat

There is another trap here. Once companies recognize the value of their data and processes, the temptation is to build the entire technical stack around them. This brings us to RAG.

I can already hear software architects objecting: there are valid reasons to build RAG in-house. They are right. Security, control, or a highly specific domain may justify owning it. My problem is not with RAG itself. My problem is with treating the first working vector-search demo as a competitive advantage.

Consider the two-tower recommendation model described by Databricks. One tower processes information about the user, while another processes information about the item. Their representations can then be compared efficiently to retrieve thousands of relevant candidates. More complex models can rerank those candidates using a richer set of features. Databricks combines this with distributed training, large embedding tables, experiment tracking, and infrastructure designed to work at scale.

A two-tower recommender is not a one-to-one replacement for every RAG system. It is, however, a good illustration of how deep the retrieval problem can become once relevance, personalization, scale, and cost matter.

I bet it is hard for a generalist product team to beat specialists who work on these problems every day. The same applies to vector search, reranking, evaluation, observability, and inference optimization. A simple proof of concept is easy. A reliable production system is not.

The sensible choice is usually to buy or integrate the commodity parts and invest your own engineering capacity in making proprietary data and process knowledge usable.

What should companies own?

If you want an AI advantage tomorrow, start producing the right data today. Record decisions, outcomes, exceptions, human overrides, and the context in which they occurred. Documents explain how a process should work. Operational history reveals how it really works - and whether a decision was successful.

Is this less exciting than unveiling another friendly advisor on a conference stage? Absolutely. It is also much harder to copy.

Before starting another AI project, I would ask four questions:

• What proprietary data do we have that the model providers and our competitors do not?
• Which processes and exception paths contain knowledge unique to the way we operate?
• How can we turn it into reliable context without losing control of it?
• Which parts of the solution are truly specific to our business, and which should be left to specialists?

There is no shame in having no edge in models. Almost nobody has one. The danger is spending time and money while pretending otherwise.

Big tech owns the models. Specialists own the plumbing. You must own the context.

My conclusion is less exciting than most AI keynotes: do not try to win by adding yet another agent. Build the data and process memory that makes whichever agent you use noticeably better.