Two developments are currently emerging in online retail that belong together.On the operator side, AI agents – i.e. software components that plan and execute tasks independently – are taking over operational routines in the shop.On the customer side, the same mechanisms appear as shopping advisors that compare products and, in perspective, also handle the purchase itself.

The German startup Shop Agentic stands for the first direction.According to its own account, a large part of the daily work in online retail is to be handled by around 60 AI agents.The company was founded by four people with many years of experience in the e-commerce environment.Which tasks the individual agents cover in detail and how far the automation actually carries in everyday life cannot be conclusively assessed from the reporting to date.

The second direction concerns the purchasing process itself.Agentic AI can in future take over a range of tasks in online retail that people in front of the screen have previously handled: from product comparison to purchase processing.The technological prerequisites for this are largely in place – but components are still missing for fully autonomous online shopping.This very gap is the relevant quantity for shop operators, because it determines how much time remains for preparation.

Why both developments need the same technical basis

Whether an agent on the operator side maintains stock or an agent on the customer side compares offers: both do not access what people see, but what a system provides in a machine-readable form.A visually successful product page is of little use to an agent if variants, availabilities, delivery times or return conditions are available only as running text or graphics.

This shifts the focus of many shop projects.For years the focus was on conversion optimisation at the surface – layout, checkout path, purchase arguments.Agent readiness, by contrast, arises one level below: in data quality, in clean interfaces and in traceable rules for prices, discounts and shipping costs.

What usually causes failure in everyday life

Experience shows that it is less the AI models than the grown system landscapes that put the brakes on such projects.Typical problem areas are:

  • Scattered product data: Attributes exist in parallel in the shop, in the ERP and in spreadsheets, without a source being defined as the leading one.
  • Incomplete attributes: Dimensions, materials or compatibilities are missing precisely where a comparison agent would evaluate them.
  • Interfaces without contracts: APIs exist but are not documented, not versioned and change without announcement.
  • Unclear distribution of rights: It is not regulated which automation may trigger which action – for example changing prices or cancelling orders.
  • Missing logging: Without clean logs it cannot be clarified afterwards which process triggered a change.

These points are not AI topics.They are classic architecture and data questions that are merely made more painfully visible by agents.

Agents in shop operations: realistic expectations

That a startup largely hands over everyday shop operations to agents is a signal, but not an operating model for every company.A staged approach makes sense: first narrowly defined, recurring tasks with clear success monitoring, such as enriching product texts, pre-sorting service enquiries or checking data sets for gaps.

The decisive question here is where a human must approve.Suggesting agents whose results are released before publication are considerably easier to justify than agents with direct write access to prices or stock.This approval step can be reduced later once quality and error patterns are known.

When the customer is an agent

Should agentic shopping prevail, this changes the logic of visibility.An agent evaluates offers based on structured criteria and not based on design or brand experience in the classic sense.For shop operators this raises questions that should be answered today:

  1. Which data does the shop provide in machine-readable form – and is it correct and up to date?
  2. How are automated accesses detected and distinguished: desired shopping agents versus scraping or misuse?
  3. How can orders be secured that are not triggered by a human in the checkout – from authentication through payment approval to revocation processing?
  4. Which information does an agent need in order to be able to evaluate additional services, guarantees or service quality at all?

The point of securing in particular is likely to be among the components still missing for fully autonomous shopping.Much is technically possible; legally, contractually and organisationally, many processes are still designed for a person acting.

A pragmatic roadmap

For medium-sized retailers, a sequence is more advisable than a major project.First the data stock: a leading source per attribute, measurable completeness, regular checks.Then the interfaces: documented, versioned, with clear access rights and logging.Only then the agents – initially internally, in delimited task areas, with defined abort criteria.

This path has a pleasant side effect: clean product data and robust interfaces pay off even if agentic shopping develops more slowly than currently expected.

You improve search, filters, feeds and marketplace integrations independently of any AI strategy.

Classification for web and software projects

For web and software projects, the investment focus shifts from the interface to the structure: data model, interfaces, permissions and traceability determine whether automation is even possible later on.Anyone who now invests in leading data sources and documented APIs keeps both options open – the internal use of agents and the connection of external purchasing agents.Projects should therefore be structured in such a way that every step delivers measurable benefits even without AI.

Sources