According to Gartner, 40% of enterprise apps will embed AI agents by the end of this year, yet most retail infrastructure is “still built for human interaction, and not for systems operating at machine speed”. That was the analysis from one commerce software company recently, which went on to suggest that this supposed gap is “becoming a commercial risk as AI starts executing pricing, inventory and promotions decisions in real time”.
So, how to catch up? In the view of Dirk Hoerig, co-founder, Managing Director, CIO and former CEO of Munich, Germany-based commercetools, a complete overhaul of e-commerce – and of shopping generally – is necessary, so it can become autonomous/headless to use his term. In this new model, agentic AI will step up to run entire retail operations for both buyer and seller, rather than merely assist humans.
But in the same week, research arrived from Economist Enterprise, the consultancy wing of publisher the Economist Group, which suggests that agentic AI is running riot in most organizations. According to a new white paper Power Without Control – Rethinking Cybersecurity for the Age of Agentic AI, a staggering 98% of enterprises have experienced a disruptive agent-related incident, with 90% saying they are deploying agents faster than their security teams can evaluate or govern them.
That doesn’t sound like a world of instant wins. So, why are business and IT leaders acting so rashly? The answer appears to be hype, FOMO, and commercial, competitive pressure. The Economist Group goes as far as saying that, in the agentic era, “failure is inevitable”, with the threat no longer at the perimeter, but brought into the heart of the enterprise as agents break business processes as much as automate them. Such findings surely beg the question: should we be rushing into headless, autonomous commerce, as Hoerig suggests, when agents are doing such damage elsewhere, in enterprises that were keen to secure productivity gains, but now find themselves battling chaos?
Context
As background it’s worth considering why Hoerig founded his firm in the first place. He explains:
We founded commercetools in 2010 to provide brands and retailers with a more modern [software] solution, as we saw the market shifting, and we believe that the traditional commerce platform stack had been solving an old purpose. With more touchpoints and faster-evolving markets, brands and retailers need to adapt more quickly, and be faster and more flexible. We wanted to have something that’s future proof, cloud native, easy to adapt, easy to integrate, and can empower any touchpoint for sales, so more like a commerce operating system than a pure, Web-only shop.
Today, commercetools is the largest pure-play in its market in Europe and second only to Shopify on the world stage, with over $110 billion in gross merchandise value (GMV) running over its platform, via brands like BMW, Chanel, John Lewis, Screwfix, Burberry, JD Sports, Lululemon, and Sephora.
So, why the deep dive into agentic AI? Rather than a me-too strategy, Hoerig stakes a claim to being an early adopter:
We had been at the forefront of agentic commerce, even before it became a topic. We thought, ‘OK, how is the customer journey going to change if humans are using AI for the discovery journey via their favourite LLM (Large Language Model)?’ You, know, ‘I go camping, help me find the right equipment’. Those kinds of things are impacting the journey, so what does it require from a technology standpoint? You need to start thinking about where things are going and the purpose of your commerce tools. And that was why we built a new platform after creating a leading one out there. We saw, as we did at founding, that the initial pass wasn’t sufficient anymore.
The requirements of brands and retailers are changing. Our mission is to always ensure that whatever is going to happen, our customers can easily adapt. And we believe now is the time where we need to figure out a better way to work with a computer. So, how do we work with AI agents to drive significantly better outcomes? That’s the question for us.
But better outcomes for whom, and how? He says:
At one end, every company has this pressure to improve efficiency and productivity: to get more done, ideally with less. But the other challenge is in the day-to-day operations of any brand. Pick any brand and see what’s happening, what the teams, the Marketers, the Product Managers, and the category managers are doing. There are a lot of manual workflows that are lacking real-time data, and decisions that must be made under time pressure, and there are constant changes in the market.
Most promotions are still being planned with spreadsheets, and it’s just humans trying to collect the data around what they believe customers are looking for. So, it’s, ‘What do we have? Where do we have overstock? What worked last year at this time?’ But what we do now is never really based on real-time information and a sophisticated approach. And five minutes later, they need to do the next thing, and the next.
Then look at how much time it takes for somebody to create a sophisticated campaign across their commerce platform and their marketing tools. It can take multiple hours. So, what we envisioned is, ‘Can’t we help our customers make better decisions?’ At the same time optimizing product data, inventory, pricing, campaigns, and so on, and suggesting new campaigns. Could we not just automate it instead of letting the team do it manually? We believe there’s a big shift there.
Autonomous
In this world of agents running all these processes autonomously, various possible problems present themselves. First, what’s wrong with a human team spending hours creating a new campaign? Second, enterprises are, as Hoerig says, primarily driven by the desire to do more with less and save money, rather than by making smarter decisions.
But the Productivity Paradox reveals they are not being more productive, and growth is not magically appearing from AI, largely because token costs are soaring, agentic chaos is spreading – as the Economist Group suggests – and AIs themselves are hallucinating. Left unchecked, that undermines businesses, but when AI’s outputs are checked, all those time-savings vanish.
Third, many customers don’t appreciate AI being forced down their throats across every platform; a good technology should never require infantilizing users. And fourth, we live in a world of dumb algorithms, even on the biggest, most cutting-edge commerce platforms. AI has made many outlets clumsier and pushier, and customers are an afterthought in an industrialized, lowest-common-denominator process.
Someone once observed that, if you asked AI to build a coffee business, you would end up with a million Starbucks clones, because Starbucks works, is global, and people seem to like it. Thus, it becomes the most significant token, if you will, and this principle applies in every kind of retail. But that denies humans’ capacity for original thought: in general, AI does not (and cannot) innovate, it merely reproduces the average at scale, based on scraped precedent.
Oasis’ songwriter Noel Gallagher once famously said that people didn’t know they wanted the Sex Pistols. As any good marketer knows, your job is to tell customers what they want, even if they have no idea they will like it. But that takes intuition, imagination and a willingness to be wrong, the very things that some enterprises now regard as an unnecessary cost center. By contrast, agentic commerce is industrialized vanilla – every day, forever, just more of the same.
Hoerig appears to concur, saying:
From our perspective it’s the wrong approach. But from a technical use case, because you may have interfaces that you want to use manually, a smart assistant may be the right approach. Most experiences aren’t there yet. I think technically with AI, it could get much better, but you run down the thread and optimize to the average. And for a true personalized experience, most companies don’t have the data for it, and lack the capabilities. So, AI will only be half of the equation [for them], or maybe only a third.
So, what does he suggest? He explains:
So, we said, ‘OK, we have all these interfaces, but let’s create a new one that works differently to any other business user interface that has been created out there.’ And that’s what we have just announced with MosAIc.
This is the new “autonomous operations layer for commercetools Sphere”, says the vendor, where “intent replaces interfaces, AI navigates complexity, and commerce executes within the rules you define.” Join the wait list now, it adds.
The pitch here is this, according to the vendor’s website – MosAIc “puts your commerce operations on autopilot”, with “promotions, pricing, approvals, and more, handled by AI, within the rules you set. Hoerig continues:
[With MosAIc, you can say], ‘Help me analyze across all the sales of the last month where we lost gross margin, then provide some recommendations of what we should improve’. Now, you can argue if you personally want this or not, but it can connect with any agent you give it access to, and not just to the ones we are providing within the commerce platform.
So, you might have an ERP system that has an agent, and they can talk to each other, or a fulfilment system that has an agent, and they can talk with each other too, or with a procurement platform. It can talk to all of them, collect all the data, structure it, and generate a user interface on the fly, and it will provide you with the data and give you suggestions. And that is the autonomous commerce approach.
But not everything should be automated, he admits, and not everybody wants that:
It’s not a goal of you and I sitting on the beach, while [brand redacted] is running all the sales on its own.”
Hallucinations
But the potential for any platform to do that is clearly there. And in a world of autonomous agents talking to other agents and making decisions, the Economist Group suggests that enterprises may be losing control of their businesses, rather than putting themselves in the driving seat. All of which raises another risk of autonomous commerce in general terms – hallucination, bias, and – at an extreme – potentially fraud and anti-competitive practise.
First, how do humans know the AI is surfacing accurate, trustworthy data from within the enterprise or from other agents? And second, how can the end customer know what commercial relationships may be lurking behind the scenes? For example, why might the AI recommend Product A, rather than B or C? What if B is better, but the platform is incentivized to recommend A? Potentially, agentic retail could mean rigged commerce that syphons revenue to preferred partners, invisibly and autonomously?
Hoerig states:
I believe transparency is super important – why has the recommendation been made, and on what data? That needs to be made visible. In conversations I have had with Anthropic and the Google Gemini team, that’s why they don’t have advertising in it yet, because they fear losing the trust layer. I believe from the user standpoint this is very much required.
At the moment, agents are not really shopping. But once we get into a purchase flow, once they are shopping, trust becomes much more of a thing, right? And there’s also the risk on the fraud side. Transparency across the whole chain is critical, both for the buyers and for the sellers.
My take
Easy to say, of course. Which brings us back to sovereignty versus agentic chaos. Enterprises say they want sovereignty over data and technology, yet at the same time are handing control to AI agents. Hoerig says:
Sovereignty, I fully agree. I’m in discussions with politicians and entrepreneurs here in Europe on that topic. We are operating globally, but obviously we are founded in Europe, but we still rely, to a certain extent, on data centers that are owned by US hyperscalers, like everybody else. At the moment, there are not enough alternatives out there. On the second point, you cannot let agents just access your infrastructure without any kind of governance and protection layer, because that’s like giving the root key to a hundred new employees that you haven’t done any onboarding with yet! You need to have security in place on an infrastructure level.
But the challenge is that many companies just give out OpenAI, Claude, or whatever, licences, but don’t have the underlying infrastructure in place. Our way is the opposite. We started last year with MCP and AI gateways on top of our APIs. It was important for us that we are not creating a space where this can’t be under control.
Time to look a little closer at The Economist Group report, next.