Retailers are currently obsessed with the wrong side of the screen.

The industry is watching a race between Google, Shopify, and Amazon to build the next great interface, the AI agent that can search, recommend, and eventually execute a transaction without a human ever clicking a “buy” button.

But while the market focuses on how these agents will talk to customers, a much more dangerous gap is opening up behind the scenes.

The hard truth is that most retail organizations are structurally incapable of being operated by a machine. We are moving from AI that suggests to AI that acts, from conversational shopping to delegated execution.

In this transition, if a business has not mapped its decision ownership, internal data flows, and operational accountability, an agent will not reduce complexity. It will simply scale confusion at a speed the business cannot handle.

From suggestions to delegated execution

Most people still treat AI agents as smarter chatbots, or as a conversational layer designed for product discovery.

This is a fundamental misunderstanding of the technology’s trajectory.

An agent is software with permissions to take action. Platforms like Shopify are already leaning into this reality, having recently released integrations that move beyond discovery to allow for direct agent-led checkouts.

Realistically, we are ready to delegate low-risk, high-volume tasks that are easily reversible, such as product data enrichment or internal data preparation. We are not yet ready to delegate high-stakes commercial decisions, and the same can be said about legal contract approvals or final pricing strategies.

The risk profile changes entirely when you move from an AI that tells a customer which shirt to buy to an AI that is authorized to spend that customer’s money.

Preparing the surface vs. the operating system

Retailers are spending millions to ensure their product catalogs are machine-readable so they show up in agent-led searches. However, a machine-readable catalog is not the same as a machine-operatable business.

There is a gap between hype and reality. On the one hand, Gartner recently predicted that 60% of brands will use agentic AI by 2028. On the other, according to Deloitte, only 11% of organizations have actually deployed agents with success. This highlights a massive disconnect between interest and actual infrastructure readiness.

In this regard, the biggest operational gaps exist in the “boring” middle layers, such as real-time inventory accuracy across twenty different markets, pricing consistency between channels, and warehouse logistics.

Currently, these answers live in a fragmented mess of ERPs, spreadsheets, and employees’ knowledge. When an agent asks, “Is this item actually in stock?” it needs a definite answer. And if your internal systems are in conflict, the agent cannot function.

You cannot run an agent on fragmented data. Instead, you need a dynamic “digital twin” of your day-to-day operations in the form of a single, living data layer that reflects the true state of your business in real time. In a nutshell, you cannot build a working complex system until you have a working simple system.

The decision architecture bottleneck

The real bottleneck in retail today is decision architecture. An AI agent cannot improve a process the business itself does not understand.

In my experience, very few companies can actually map how a decision is made across teams. They still rely on “Slack-based” or “Email-based” approvals that leave no digital trace for a machine to follow.

Before automating, a company must map who owns a decision, what data is trusted for that decision, and what the thresholds for human escalation are. This mapping, combined with your operational data, forms the context layer, the digital twin that the agent uses to ground its judgment.

The warning sign that you are automating confusion is when your teams spend four days a week cleaning data and only one day making decisions. If a human cannot explain the logic, it is impossible for an agent to execute it.

When automation scales complexity

There is a pervasive myth that adding AI will automatically streamline a business. In reality, automation often makes systems more complex. When agents act quickly across weak or fragmented data, errors scale faster than a human team could ever manage.

This is why Gartner predicts that 40% of agentic AI projects will be canceled by 2027 due to a lack of clear business value or the absence of these essential risk controls.

For instance, if your inventory systems disagree, a human might catch the discrepancy during a manual check. An agent, operating on a “junior employee” level of judgment, will simply place the wrong order or promise a delivery that likely won’t be fulfilled.

Certain areas, such as sensitive brand topics, high-margin pricing strategy, and complex customer compensation, should never be fully automated. These require human critical thinking and accountability, which machines lack.

The new meaning of trust

Trust, today, goes beyond a customer’s perception of the brand. It encompasses the technical and operational trust between the customer, the agent, and the merchant.

We are already seeing this friction play out in the courts. A recent 2026 legal battle between Amazon and Perplexity over shopping agents browsing on a human’s behalf has forced a debate on whether agents should be treated as transparent digital identities or human replicas.

Retailers are bracing for this by updating their fundamental terms of service. Target, for example, recently updated its terms in collaboration with Google to address third-party agent behavior and consumer liability.

To maintain control, retailers must treat agents like employees on a payroll. They need their own digital identities and clear permission scopes, as well as rigorous audit trails.

How to keep control? By defining exactly what an agent can read, buy, and change, and by ensuring every action is traceable and reversible by a human.

The five levels of readiness

In this landscape, the companies that thrive will be those whose operations are clean enough for agents to act safely. This requires five forms of readiness:

• Data readiness: One reliable operational truth across all systems
• Decision readiness: Clear ownership of what is automated and what is escalated
• Process readiness: Redesigning workflows for an agent-first world, instead of patching old ones
• Governance readiness: Full audit trails and transparent human accountability
• Commercial readiness: A deep understanding of how agents improve margins

Here’s an example. If a retailer had 90 days to prepare, they should start by picking one high-volume, narrow workflow where the ROI is obvious. Build a “digital twin slice” of just that process by tracing the inputs, the approvals, and the outcomes. Prove it works in a small, measurable way before expanding.

The most uncomfortable question executives must ask is this. If we disappeared tomorrow, is our data foundation and decision-making process documented well enough into a clear digital twin that a machine could replicate it, or does our business context only exist in our employees’ heads?

If the answer is the latter, I’m afraid to say no amount of AI investment will save you.

However, for those willing to do the hard work of operational cleanup today, the coming agentic era offers the first genuine opportunity to scale the best part of their business without scaling the chaos.

We've featured the best AI tool.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit