Insights Playbook

What was said

Omar opened with evidence that agents have entered every stage of commerce, on the demand side through shopper agents and on the operations side through agents that prepare products, route orders and manage exceptions. Machine actors have grown from 6% to 10% of total web traffic in a year, agentic traffic reached a new monthly high in August, and it is growing six and a half times faster than human traffic. Retail's share of that traffic rose from 42% to 46% in a single month, with e-commerce traffic growing 38% month on month.

He argued that operations will see an even bigger impact than the demand side. Enterprise agent deployments are up nine times year to date, the average company has moved from five agents to 13, each agent now carries six skills rather than two, and Salesforce data shows deployed retail agents completing 18 times more work than a year ago. He also acknowledged the limits: ServiceNow found only around 9% of deployed agents run complex, multi-step workflows, and most companies sit between AI-assisted and agent-executed on the maturity curve.

His central point was that work will change task by task rather than role by role. AI will first rank exceptions, draft disputes and propose fulfilment plans, then execute routine workflows end to end within set boundaries. As authority is delegated, new human work appears: checking what agents did overnight, looking for drift and deciding which policy boundaries to move.

Key takeaways

• 77% of agentic e-commerce traffic goes to site search and product pages. Taxonomy, semantic context, pricing and availability determine whether agents land on yours.
• The operations side is moving faster than most expect. One Fortune 500 retailer cut five analysts' 100 hours a month to one person overseeing agents for two hours.
• Most deployed agents still perform narrow, supervised work. Fully autonomous multi-step workflows remain rare.
• Plan the shift task by task. Some tasks will be automated, some will become hybrid, and new oversight tasks will emerge.
• Operations depend on people who carry undocumented knowledge. Capturing it in systems and data is what allows agents to operate with context.

Think about what you had to do to grow to get to where you are. And is that what's going to get you to where you need to go next?

What was said

Consumers are moving faster on AI than companies. That is the fundamental issue. People, as well as companies, tend not to change unless they are pushed. Existential pressure drives transformation, and we are already seeing it in three sectors.

In automotive, Chinese manufacturers such as BYD are operating at a speed that is forcing traditional manufacturers to respond. In financial services, challengers such as Nubank are expanding beyond Latin America into North America and Europe. And in packaged goods, the pricing power companies enjoyed during Covid has disappeared, putting greater pressure on growth and efficiency. A downturn may actually accelerate AI adoption because it will force companies to address cost, productivity and bureaucracy.

At the same time, there is a legitimate question around the extraordinary level of investment going into AI infrastructure. Forecasts for hyperscaler capex have risen beyond $800 billion this year and towards $1.1 trillion next year. The major technology companies are moving from relatively capital-light to increasingly capital-intensive models, with significant amounts of borrowing involved. The comparison is with the development of the railroads. The infrastructure will fundamentally change the economy, but it is difficult to believe that investment on this scale will happen without mistakes along the way.

For brands and retailers, there are three priorities: agility, control and first-party data. Companies have to move faster. One global packaged-goods company takes around 200 days to approve an advertisement, while Chinese competitors can develop and launch products in roughly the same period. That is not a technology problem. It is an organisational problem.

Companies also need to reconsider capabilities that were outsourced following the global financial crisis. Data, technology and increasingly content production are too important to competitive advantage to sit entirely outside the organisation. First-party data becomes even more valuable. Retailers have enormous amounts of customer data, but owning the data is not enough. They need the capability to use it at scale, combined with signals from the major technology platforms.

AI agents will also change the role of brands. If an agent is making or influencing the purchasing decision, you want the consumer asking for Colgate, not toothpaste. Brand preference therefore becomes more important, not less. At the same time, brands need to make sure their products and product information are discoverable by the machines crawling for them. AI should also bring greater transparency to the media supply chain, just as technology is reshaping retail supply chains.

Key takeaways

• AI is about workflow, not technology.
• Speed is a fundamental competitive advantage.
• Existential pressure accelerates transformation.
• First-party data and the ability to activate it are core capabilities.
• Brands become more important as AI agents mediate purchasing decisions.
• Bureaucracy remains the biggest barrier to transformation.

AI is not about technology. It's about workflow. The first thing that we do in every piece of transformation is analyse the workflow.

What was said

Linda opened with the SNKRS app at Nike, where limited-edition launches moved from queues outside stores to digital draws entered by millions. Demand was hard to forecast and customers were frustrated at never winning. Her team built in-app features designed purely to gather zero-party data on customer tastes, cutting forecasting error by 44% in the first year and by up to 80% over time. A machine learning model using 55 data points then decided who should receive exclusive access, producing 90% redemption and statistically significant increases in spending over the following 30, 60 and 90 days.

She then set out what is changing. Brands now serve two audiences, people and bots. AI search queries average 24 words compared with four on Google, 89% of B2B buyers check AI before choosing a vendor, and the traditional funnel is collapsing. Awareness shrinks as the bot does the discovering, choice narrows to one or two recommendations, and loyalty becomes algorithmic. Whoever controls the agent layer between consumer and brand captures value, and brands risk losing margin and customer data even when they win the sale.

Just as important was what stays the same: innovation, pricing, great product and loyalty still decide purchases, while brand building, authenticity and trust become more valuable. Her recommendations were to reclaim the direct customer relationship through owned channels, stores and community, invest in product metadata, and understand how AI search actually works. A Stanford study she worked with found 63% of AI answers to shopping queries were factually wrong. Scarcity messaging such as "limited edition" cut bot recommendations 13-fold, while social proof and recent content raised them.

Key takeaways

• Start from a business problem with executive sponsorship. An "AI strategy" with no defined goal will not deliver ROI.
• Automation will become table stakes. Transformation starts from the outcome and redesigns the work.
• Watch the full cost of AI. Automating purely for savings often ignores compute, deployment and training costs.
• AI search is not set-and-forget. Every model and every refresh behaves differently, so it needs constant test and learn.
• Most brands will need a selectively open model with agent platforms, sharing enough data to transact while protecting the customer relationship.
• Tools account for around 30% of the value in AI. The other 70% comes from data, process and skilled people.

The highest ROI in AI is not because you have the best AI tool. It's because you have a tool appropriate for your use case, and you spend time wiring it into workflow, process and people.

What was said

Komal framed digital performance around three levers: bringing more visitors to the top of the funnel, converting more of them, and selling complete solutions to those who buy. She then walked through what AI has and has not changed in each. On a last-click basis, traffic from large language models is still a low percentage (albeit growing at pace), yet visitors who have researched off-site arrive with high intent and convert better, a benefit she cannot yet attribute. Google's AI Overviews and AI Mode are pushing organic listings further down the page, creating a zero-click problem where the brand is cited but not visited.

She was candid about AI tools needing training. Google's AI Max is a good example of tooling that helps automate campaign management but is not ready to use straight out of the box. Once trained, it helped the business buy media more effectively. On site, Currys is testing natural language search, though customers still mostly type generic terms, and catalogue attributes remain the limiting factor. For omnichannel retailers it's more important than ever to create seamless experiences between channels, as order and collect remains the preferred fulfilment method for customers looking for speed and convenience.

The second half of the session turned to peak trading, which runs from Black Friday through gifting to Boxing Day. November is now the biggest month, and Boxing Day matters far less than it used to. Her approach rests on predicting the four or five product areas or trends that will make or break the season, planning for failure as seriously as for success, and building a story that brings the whole team with you.

Key takeaways

• Digital performance still comes down to three levers: traffic, conversion and basket size.
• LLM traffic is still small (albeit growing) but high intent. Its biggest effect may be conversion that attribution cannot yet capture.
• Zero-click search is a growing threat to organic visibility as AI summaries push listings down the page.
• AI tools are not inherently smart when first deployed. They need training, specific goals and brand context.
• Fundamentals come first. A faster checkout and payment infrastructure ready for agentic commerce matter more than a flashy shopping agent.

If you change nothing, nothing will change. But if you don't change anything, you cannot get better.

What was said

Matthew described True as a vertically integrated investment and advisory firm with stakes in more than 100 early-stage companies, a private equity portfolio including the Cotswold Company and Zwift, and a view of around 6,000 technologies a year from 40 countries. He called the current environment the hardest in his 27-year career, with no economic growth, rising taxes on retail and online economics worsening as paid digital channels become more expensive. His advice was to control what you can, assume there is no tailwind and build a better operating model than anyone else in your category.

He challenged the industry's obsession with channel over customer. When he launched his first online business, advertising on Google cost a fraction of opening a store; today the maths has flipped. At the Cotswold Company, stores have proved a far cheaper way to acquire customers than paid search and social. Using drive-time catchment data to place each one, the business has grown from an online-led model into a profitable network of shops while keeping strong returns on capital.

He expects AI to follow a path similar to e-commerce, but broader. Just as Amazon reinvested its economic advantage for decades while others dismissed it as loss-making, AI will pervade every process and organisational structure, and businesses that capture the advantage early will outcompete. He also shared early thinking on Argos, where he sees a 54-year-old brand, highly engaged colleagues, 94% same-day capability and one of the most automated distribution centres in the country. He noted that many large stores are currently used as collection points, a gap he sees as an opportunity to rethink what physical space does for the customer.

Key takeaways

• Automate what the customer never sees first. Back-office savings in procurement, legal and finance can be reinvested into price and service.
• Allocate capital by customer, not by channel. Well-placed stores can acquire customers far more cheaply than paid digital.
• A fully autonomous operating model is already possible at small scale. Project Lemonade runs on a contextual knowledge graph with a single source of truth for data, product and compliance.
• Technology belongs at the centre of the organisation and at every board table.
• Established brands carry equity that is now very expensive to build from scratch, which makes them valuable platforms to reinvest in.

The automation will lead to humans and good old-fashioned emotional intelligence becoming the asset for retail.

What was said

Steve Collinge opened with the view that some parts of home and garden will never change: the desire to improve your home at the start of the journey, and someone fitting a kitchen or hanging a door at the end. Everything in between, from inspiration and search to purchase, advice and delivery, is changing fast. With price, range and experience as the only levers for competitive advantage, he argued that range expansion with limited investment is one of the few ways left to open new profit streams, and warned against a future where every marketplace carries the same 10 million products.

Simon Oakes said Toolbank's dropship volumes have more than doubled in five years, and the business keeps investing in the capability while choosing its partners carefully. He urged retailers to extend ranges in ways that fit their brand identity and to build curated ranges that fill real gaps. Kevin Morgan described how Crystal Direct grew its range of standard-size windows and doors from around 500 to around 1,400 SKUs, delivered by its own fleet to reduce touchpoints and keep bulky, fragile stock out of retailers' stores. Consistent communication at every stage has helped Crystal reach 98% to 99% on time and in full, and the business has now begun offering made-to-measure products online.

The panel discussed the friction that slows growth. Product data requirements differ between every retailer, listings get lost without marketing and visibility, and some retailers still apply in-store margin expectations to dropship despite its lower cost to serve. Collinge described getting a product live on wilko.com within four hours of meeting the team, proof of what is possible when retailers prioritise speed. Both suppliers are now expanding into Ireland, with Crystal entering Northern Ireland and the Republic through its Virtualstock relationship.

Key takeaways

• Curate rather than flood. Stay within the range your customers would allow you to sell, a principle from Tesco Direct that still holds.
• Dropship is growing. Toolbank's volumes have more than doubled in five years.
• Supplier data should be a foundation. Retailers get more from it by enriching it for their own customers and for AI search.
• Realistic margin expectations unlock more business. Dropship carries lower cost to serve and no working capital for the retailer.
• Ireland is a sizeable, underserved opportunity for UK suppliers and retailers.
• Suppliers still need retailers. Most cannot match the reach, trust and marketing budget of a major retailer, so partnership beats going direct.

The retailers who decide to broaden their range with limited investment are the ones that can open up a pipeline of cash into their business.

What was said

Raman described how digital has become an increasingly important part of Dunelm's customer journey, including for considered purchases such as furniture. Her philosophy is to approach innovation like a scientist: start with a clear problem and hypothesis rather than with the technology itself.

Retailers rarely suffer from a shortage of ideas, but experimentation can easily become a cycle of pilots and demonstrations without a clear objective or end user. Raman advocated starting with customer missions and asking what tangible value an initiative will create for customers and colleagues.

She shared how focused experimentation can improve the customer experience incrementally, including Dunelm's work exploring AI-powered search and content. Rather than attempting to transform an entire journey at once, the emphasis is on testing specific opportunities, learning from the evidence and building on what works.

That same discipline applies to prioritisation. Potential initiatives need to balance strategic value with the organisation's ability to deliver them effectively. And different customer missions require different approaches: someone making a considered furniture purchase may want a very different experience from someone making a simpler household purchase.

Governance is an important part of moving from experimentation to impact. Raman described the importance of regularly reviewing initiatives against evidence, establishing clear ownership and understanding what would be required to scale successfully. Crucially, not every experiment needs to progress. Pausing or stopping an initiative can be as valuable as scaling one if the evidence does not support further investment.

Key takeaways

• Build the foundations before chasing the latest technology.
• Start with a genuine customer problem rather than an AI use case.
• Focused improvements can create meaningful gains across the customer journey.
• Think across physical and digital channels rather than treating them independently.
• Stopping or pausing an experiment is a legitimate outcome when the evidence supports it.

Stopping and pausing is almost like a nasty word. Actually, I think it's a really clever thing. I always say to my team, I'd rather fail fast.

What was said

Steve opened with the question many attendees were asking: with AI pilots running across merchandising, marketing and operations, how do organisations move to a real operating model? Hariss described the most common structure as hub and spoke. Strategy is agreed at management level, a centre of excellence acts as the hub, technology provides the platforms centrally, and the business owns the use cases.

He pushed back on the idea that AI needs a single owner, comparing it to asking who owns Excel. The ownership question usually comes up after something has gone wrong, when people want someone to hold responsible. A better question is who owns the outcome the AI is meant to deliver. Strategy, budget, technology, data, use cases and outcomes each need clear responsibility and accountability, and they rarely sit with the same person or department.

On speed versus governance, he argued that governance is what enables speed. If guardrails, guidelines and brand voice are defined globally, individual use cases can move quickly without each one becoming a risk. When agents make the wrong decision, accountability sits with whoever designed the use case and its guardrails, and a human in the loop has to be real rather than a box ticked for the record. No business should let an agent automatically authorise large refunds. He also shared his own experience of deploying an AI customer service platform, where involving every affected employee from the start built the buy-in that made it work, and the team discovered an upsell benefit nobody had planned for.

Key takeaways

• Treat AI as a capability across the organisation, not a territory for one department.
• Ask who owns the outcome, not who owns the AI. Split responsibility and accountability across strategy, budget, technology, data and use cases.
• A centre of excellence keeps the organisation current as models change and spreads learning across teams.
• Define guardrails once, centrally, so individual teams can move faster with less risk.
• Change management is harder than deployment. Treat affected employees as stakeholders from day one.
• AI deployments do not have to be about cost cutting. They should reflect the organisation's values and mission.

Organisations need to be asking not what AI system they have in place, but what their AI deployment is saying about their organisation.

What was said

Shaje began by polling the room. Most attendees had been through an AI transformation, fewer were happy with the results, and almost none had measured its impact on their people as a KPI. He cited MIT research suggesting digital transformation is 70% about people, yet people are rarely built into how transformations are measured. At a recent workshop with a telecommunications company, he spent time with frontline customer service staff rather than managers and found 40% of their time went on chasing internal teams for order information, with another 10% spent updating a CRM for reports few people read.

The core of his talk was the Irony of Automation, a 1983 study showing that relying on a human in the loop as a safety net is one of the weakest forms of control. Organisations automate the easy work, leave the hard work to people, and those people lose the skills they need through lack of practice. He cited a 2025 Microsoft and Carnegie Mellon study of 319 workers linking greater reliance on AI with less critical thinking, and asked how businesses will develop future experts if junior staff stop learning the job and senior experts no longer have time to mentor.

He also warned about runaway costs and cognitive offloading, and argued that AI literacy is the first step in any transformation. In the Q&A, he made the case for responsible regulation, comparing AI to the gas network, which is dangerous but made safe through regulation and treated as a public utility. He also recommended open-weight models run locally as a viable, lower-cost option for smaller businesses.

Key takeaways

• Start with the people who do the work. Frontline staff know where time is really being lost.
• Human in the loop is not a safety net on its own. Skill atrophy undermines the people meant to catch errors.
• Protect the talent pipeline. Deciding what to automate is also deciding what the next generation of experts will learn.
• AI literacy comes first, covering what AI is, how to use it responsibly and what it costs.
• Reframe the question from efficiency to growth: how can the same people do more?
• Plan for failure. A business that removes human capability and relies on a single model has no fallback when that model or platform goes down.

Find all the vicious friction and automate it. Virtuous friction, keep it. Reinvest that time back into solving virtuous friction and do more with it.