Two Years of AI Commerce: An Honest Assessment

Two years ago, the coverage coming out of NRF 2024 recalibrated an industry's calendar. The world's biggest retail conference had one dominant theme. AI, specifically generative AI. What came back from New York made clear something had shifted. An industry that had spent 2023 cautiously circling the subject had collectively decided, in the space of a week, that this was the thing now. Not in the future. Now. The announcements were enthusiastic. The demos were impressive. The questions from practitioners were thoughtful but tentative, nobody wanting to be the person who failed to grasp the significance of the moment.

I wrote a piece at the time, watching it unfold from here, trying to capture something I was already feeling: the genuine excitement was real, but the gap between capability demonstrations and production deployments was going to take longer to close than the conference energy implied. That piece was the first thing I published on this site.

Two years later feels like a reasonable moment to take stock honestly.

What Actually Moved

The customer service story is the clearest win. Klarna's 2024 announcement looked dramatic at the time (2.3 million AI-handled conversations, claimed equivalent of 700 human agents; Klarna's original press release), and in hindsight it was an early signal of something real rather than a vendor overpromise. Klarna subsequently reversed course and re-hired human agents, which complicates the triumph narrative, but the underlying capability was genuine. AI customer service, properly implemented on top of clean knowledge bases, now handles a significant and growing share of tier-1 enquiries across the retail sector. The ROI is measurable. The customer satisfaction data, where it has been published honestly, shows results at parity or better for routine query types. The Klarna story in full is instructive: it worked, then they overcorrected, then they found the right balance. That's not failure. That's how implementation actually goes.

Payment infrastructure for agentic commerce is built. Mastercard, Visa, PayPal, and Stripe all put the foundational frameworks in place in 2025. The Universal Commerce Protocol, announced by Google with 60-plus endorsers, is the emerging open standard for agentic transactions. This is the kind of infrastructure investment that takes years and is easy to underestimate while it is happening. It is not visible to consumers. But it is the thing that makes everything else possible.

The AI traffic quality story is also real. Adobe Analytics data shows AI-referred traffic to US retail sites grew 393% year-on-year in Q1 2026; as of March, that traffic converted 42% better than non-AI visitors. Two years ago this was a concern on the horizon. It is now a measurable revenue line for the retailers who were ready for it.

And supply chain AI, in the specific context of tariff disruption and scenario planning, has demonstrated value in conditions that were not manufactured for a demo. The retailers who had invested in it responded faster. That matters.

What Didn't Move the Way I Expected

The agentic purchasing thesis — that consumers would meaningfully delegate discretionary purchases to AI agents within two years — has moved more slowly than the infrastructure timeline suggested. The trust data tells you why: a Quad/Harris Poll from February 2026 found that only 39% of Americans trust AI to make everyday purchases on their behalf, and that drops further for anything that isn't commodity or routine. Agentic purchasing will happen, is happening at small scale, but consumer habit formation is the long pole, and it is genuinely long.

Cashier-less physical retail at grocery scale has not happened. Amazon's 2024 removal of Just Walk Out technology from its Fresh grocery stores was a telling signal: the technology is real, but the economics at general grocery scale are not there yet. Smart carts in specific contexts, computer vision for inventory intelligence, yes. The frictionless full-store experience of the original vision, not yet. I do not think this vision is dead, but I was wrong about the timeline.

And the UK picture specifically has been more sobering than the global headline numbers. Research commissioned by Quickfire Digital in November 2025 found that 77% of UK ecommerce retailers admit their AI initiatives are falling short, with roughly £92 million of investment at risk of underperformance. The access gap between enterprise and mid-market is widening rather than narrowing. I keep coming back to these findings because they are a corrective to a narrative told almost exclusively through the lens of Amazon, Walmart, and Salesforce enterprise deployments. The median UK ecommerce business is in a different position to all of those.

What Surprised Me

The zero-click search crisis arrived faster and harder than I expected. Depending on which study you read, between 58% and 68% of Google searches now end without a click, and AI Overviews are showing up on 48% of searches. Two years ago this was a concern on the horizon. It is now a measurable constraint on organic traffic that requires strategic responses rather than planning exercises.

The AI content quality problem also surprised me, not that it exists but how many organisations are running faster than their quality controls. The confidence with which AI generates plausible-but-wrong product specifications, at scale, in production environments that have not built sufficient verification layers. That is a problem that will create a long tail of customer service issues and returns that have not fully shown up in aggregate data yet.

And the retailer backlash is real and is being driven by legitimate structural concerns rather than inertia. The margin erosion from AI-negotiated commodity deals, the customer relationship question, the platform dependency concerns: these are sophisticated objections from people who understand the technology, not resistance from people who do not.

The Honest Score

Customer service AI: delivered, with caveats. Payment infrastructure: in place. Search and discovery disruption: arrived ahead of schedule, in more disruptive form than expected. Supply chain AI: delivering for those who invested early. Physical retail AI: partial, context-dependent, slower than advertised. Agentic consumer purchasing: early, building, not yet at scale. UK mid-market access: still a serious problem.

If I had to characterise the two years overall: the infrastructure bets were right and have been built faster than I expected. Consumer adoption has been slower and more category-specific than the infrastructure timeline implied. And the distributional picture between enterprise and mid-market, between the US and UK, between retailers with strong data foundations and those without, is more unequal than the industry headlines suggest.

Looking Forward Without Looking Too Far

I have learned enough about the pace of change in this space to be cautious about firm predictions, but a few things feel reasonably probable.

The Universal Commerce Protocol will drive meaningful agentic commerce volume within 18 months if it achieves the adoption trajectory its backers are targeting. Product discovery is migrating inside AI interfaces, the transaction infrastructure is in place, and the consumer habit formation is the variable I cannot model precisely.

The UK regulatory environment will tighten around AI personalisation, dynamic pricing, and agentic checkout in ways that most retailers are not yet prepared for. The EU AI Act's phased implementation, with full provisions applicable from August 2026, will continue to create de facto compliance requirements for UK retailers serving EU markets, regardless of what the UK government does or does not legislate domestically.

And the access gap will remain a serious problem unless platform providers, Shopify primarily, make genuinely capable agentic tools available to mid-market and independent retailers in a way that does not require significant technical capability to implement. Alibaba's Accio Work, which the company reports has been adopted by 230,000 businesses globally since its March 2026 launch, is a signal that someone is trying to solve this for SMEs. Whether a Shopify-native equivalent arrives within 12 months is the question I would most like to be able to answer.

---

The piece I wrote after NRF 2024 ended with something like "the experiments are becoming infrastructure." I think that turned out to be right. The question I am sitting with in May 2026 is whether the infrastructure is becoming reality at the speed and equity the industry needs, or whether we are in a phase of concentrated advantage that looks like broad progress from the outside.

I genuinely do not know the answer to that yet. The data is still arriving.

---

Data sources: Klarna AI assistant press release, Feb 2024 · Adobe Analytics Q1 2026 AI traffic data via TechCrunch · Quad/Harris Poll AI shopping trust survey, Feb 2026 · Quickfire Digital UK ecommerce AI research via Modern Retail, Nov 2025 · Zero-click search statistics, 2026 · EU AI Act implementation timeline · Alibaba Accio Work global deployment press release, Apr 2026