Here is how to be an early AI adopter without getting burned
Over the last two months, I have met with retailers and brands, attended several industry events and observed the window for waiting to see how AI plays out has closed. Fast follower is still a respectable position. Early adopter is the better one. Remaining a laggard is now a decision to fall behind competitors who are already shipping faster than you.
This is what speakers are discussing and showcasing on stage. Retailers are releasing in weeks what used to take quarters. Telcos are serving AI agents through the same APIs as their stores. Enterprises are giving their staff access to AI within frameworks that let them build safely. The gap between these companies and those still debating how to proceed is widening every quarter and the effect compounds over time.
One speaker gave the best framing I heard: “AI is not new like mobile or the internet. It is new like fire or electricity.” Both changed everything and both are dangerous in the wrong hands. Nobody rolled out electricity by handing every household a live wire. They built standards, fuses and licensed electricians and then they moved fast. The companies I saw making progress are doing the same. Below are the ideas worth holding onto if you want to adopt early and still be standing in eighteen months.
Foundations turn AI into speed
The clearest example came from JD Sports. The company had gone twelve months without a major release, then shipped three significant features since August, including a one–page checkout built and tested in five weeks. The AI tooling was the visible part. The less visible factor was two years of platform modernisation before anyone opened a coding assistant.
That order matters. AI does not remove the need for strong engineering foundations. It makes their absence harder to hide. A team with clean APIs, useful test coverage and a release process it trusts can move from a model to a production feature in weeks. A team without those conditions creates a backlog of ideas it cannot safely release.
Foundations do not mean replacing everything. Virgin Media O2 did not remove its billing systems or wider legacy estate. It added a composable layer around them, with an API fabric serving the web, contact centres, retail stores and AI agents. Brownfield beat greenfield because it reduced risk and brought value sooner.
That is the pattern we see in order management as well. Retailers do not need to replace every existing system before they can improve execution. They need to connect the systems that govern availability, orders and fulfilment, then make those connections reliable.
Guardrails, not gates
Shadow IT used to mean a spreadsheet on a shared drive. Shadow AI can mean customer data pasted into a public chatbot by someone trying to hit a deadline. It is worse by every measure and the instinct to respond with a blanket ban is understandable but counterproductive. Bans do not stop usage. They stop visibility.
easyJet described the alternative in more detail than anyone else I heard. The approach included tiered risk models, so a tool that summarises public documents is treated differently from one that touches customer records. It included a central platform with federated development, so teams can move without each inventing their own stack. It included a model gateway and an AI registry, so the organisation knows what is running and where. Policies are coded into pipelines rather than left in documents that nobody reads. AI literacy comes before access, so people understand what a tool can and cannot do before they use it.
None of that is a gate. It is the fuse box. It lets people build and it means the organisation finds out what they built before a regulator does.
The same logic applies to security. Vercel‘s CTO put it bluntly: anything that can be hacked will be. His advice was to re-scan your own codebase every time a new frontier model is released. The model that finds your bugs for you is cheaper than the bounty you pay a stranger for finding them first. Security as a one-off audit no longer makes sense when the attacker’s tools improve every quarter.
And it applies to the agents themselves. AI is probabilistic. The same prompt can produce a different answer on Tuesday from the one it produced on Monday and a model upgrade can change behaviour without any change to the surrounding application.
So, define a standard for each agent, measure it against that standard and re-run the measurement whenever the model changes. Test it before you trust it and keep testing after.
It is an operating model problem wearing a technology costume
The two-pizza team is shrinking. One speaker called the new shape the “one-sausage-roll team” (loved this): three or four people often an AI–native junior working with an experienced senior, delivering what a larger team delivered two years ago. The middle of the engineering pyramid is where the pressure lands. Most organisations have not yet decided how those roles should develop, or how the work should be divided between people and software.
This is a commercial issue as well as a people issue. A five–year platform contract assumes that a buyer can predict its needs in year four. Nobody in those rooms claimed to know what they would need in year two. Shorter commitments, clearer exit terms and vendors who earn the renewal are a better fit for the next few years than lock-in dressed up as partnership.
Leadership has a separate responsibility. Agents can reduce the time between a signal in the business and a decision about it, from weeks of reporting cycles to hours, but only when they have the company’s own context and someone checks the assumptions behind their output. An agent trained on an existing leadership bias will produce a faster and more confident version of that bias.
Agentic commerce: discovery is here, purchase is not
In session after session, the show of hands was the same. Most people in the room had used an AI assistant to research a purchase. Almost nobody had allowed one to buy on their behalf. Discovery has moved into AI channels. Payment and transaction authority have not followed yet.
That creates two immediate implications. First, if your products are not visible and accurately described inside the assistants your customers are using, you are losing the shopper before your site ever loads. This is becoming a distribution issue, not simply a marketing experiment.
Second, agents do not behave like human visitors. They do not browse in the same way, respond to retargeting or get distracted by a banner. They read the information made available to them. Product data, inventory accuracy and order management therefore affect whether an agent can represent and transact on a retailer’s behalf.
The order management layer is increasingly important here. Platforms, such as Fluent Order Management can connect real time inventory availability with order orchestration and fulfilment decisions, helping retailers give digital experiences a more accurate view of what they can offer and deliver before a customer reaches the buy button.
If those foundations are weak, the agent will encounter the problem before the customer does.
Buy for the problem, not the feature list
The way retailers assess platforms also needs to change. Two years ago, many evaluations were built around a spreadsheet of features, each one scored and weighted. That method is now close to useless, as features change quickly and an agent can increasingly generate missing functionality.
What does not change is the problem you are trying to solve. An order cannot be fulfilled from the right location. Stock appears online but cannot be purchased. A return requires three systems and a phone call.
So, the evaluation question is now, “How will this platform, with AI in and around it, take this specific problem off my hands and what will it cost me to discover that it has failed?”
That shifts what you look at in a vendor. Ask how the vendor’s platform is changing because of AI, not only what features it has bolted on. A vendor whose own roadmap is still a feature backlog has not understood the shift. Ask whether your agents, or your customers’ agents, can read and act through clean APIs. That is the brownfield pattern that worked for Virgin Media O2 and it is the one that lets an AI assistant buy from you. Ask how the vendor tests its own software when the underlying models change. A vendor who cannot answer that is passing the probabilistic risk straight to you. Ask what happens to the commercial arrangement when the original problem is solved before the contract ends.
The vendors worth considering can answer those questions in terms of your operation. A feature demonstration without an answer to the business problem is still a feature demonstration.
Early adopter, fast follower or laggard
Commerce leaders make an informed choice about how they move.
A laggard waits for the market to settle before committing. The market will not settle. Models change, competitors shorten release cycles and AI agents conducting your customers’ discovery are reading their catalogue, not yours. Every quarter spent waiting is a quarter your competitors are banking.
A fast follower adopts once a pattern is proven. That still works. Brownfield composable architecture, risk–based controls, continuous testing and small teams working on modern foundations are already visible in practice. If you are starting now, copy those and move. Apply those patterns to a defined business problem. You will not be first, but you will not be left behind.
An early adopter builds the fuse box and then wires the house. That means putting guardrails, a registry, a gateway and a test harness in place, then building on a platform clean enough for a team of four to ship to it. That may be slow for the first release, but it makes each subsequent release faster and less expensive. The companies I saw in the room doing this are not reckless. They are disciplined enough to move because they understood the conditions that let them go fast.
AI is new like electricity. The companies that benefited from electricity were not the ones that installed it first. They changed the way the factory operated around it, then built an advantage their competitors found difficult to copy.
Note: Major references to named organisations in this article reflect information those organisations shared publicly from the stage at industry events. None of it was shared under the Chatham House Rule.
