Retailers are being told they need to move faster on AI, but Satalia founder Daniel Hulme argues the biggest gains will not come from shiny generative tools, but from using the right algorithms to solve the operational problems that quietly drain margin, capacity and customer experience.
For many retailers, the conversation around artificial intelligence still begins and ends with generative AI.
Chatbots, content tools and customer-facing assistants have dominated the discussion since the arrival of large language models in the mainstream. They are highly visible, easy to understand and simple to demo in the boardroom.
But according to Dr Daniel Hulme, CEO and founder of enterprise AI business Satalia and chief AI officer at WPP, that narrow view risks distracting retailers from where AI can make the most meaningful difference.
“Most people think AI is generative AI,” he says. “The reality is there are many different types of algorithms out there that we bundle in terms of what we call AI, from machine learning to optimisation.”
Hulme has spent more than two decades researching, building and implementing AI systems. Satalia, founded in 2008 and later acquired by WPP, has worked with companies including Tesco, DFS, Waitrose and The Coca-Cola Company. Its expertise sits in the less glamorous but commercially critical world of optimisation, operations research, machine learning and decision intelligence.
In Hulme’s view, generative AI is only capable of addressing a relatively small slice of the friction inside a retail business.
“If I’m being totally honest, I think generative AI probably can address about 10 per cent of the frictions across the retailer supply chain,” he says. “You get the biggest bang for your buck if you use optimisation algorithms.”
For retailers under pressure to reduce cost, improve availability, protect margin and serve customers faster, that should be a wake-up call. AI’s greatest value may not be in the front-end experiences everyone can see. It may be in the invisible systems deciding how vans are routed, how stores are staffed, how engineers are allocated, how stock moves and how capacity is unlocked.
Where AI actually creates operational value
One of Hulme’s clearest examples comes from Tesco’s last-mile delivery operation.
Around 12 years ago, Satalia built Tesco’s last-mile delivery solution after the retailer concluded that no existing system could deliver the level of performance it needed. The challenge was not a chatbot problem. It was a series of optimisation problems, each with its own technical complexity.
When a customer enters their address and asks to see delivery slots, the system has to calculate how long it would take to get from that address to existing stops in the schedule. That means producing what Hulme calls a travel matrix in milliseconds.
“If you go and ask Google Maps, give me 1,000 routes, it’s going to take several minutes,” he says. “To be able to create 1,000 or even 3,000 routes that are accurate in 100 milliseconds required us to build bleeding-edge routing algorithms.”
The system then has to decide which slots can be offered, optimise the schedule between one customer booking and the next arriving, then later rebalance routes across drivers so workloads are fair, vans are staggered and capacity is maximised.
The impact was substantial. Hulme says the work helped Tesco save around 20 million miles a year, reducing carbon emissions while unlocking greater delivery capacity.
Satalia has since applied similar thinking to Waitrose, as well as middle-mile challenges such as moving goods from depots to stores. Hulme says a middle-mile project for Tesco took seven years to solve because the problem had never been cracked in that way before. Once solved, however, the underlying innovation could be deployed elsewhere far faster.
“The reason why it took seven years is because that problem had never been solved before,” he says. “But by solving that problem, we can now deploy that new innovation to another company in three months.”
That is a critical point for retailers. Bespoke AI doesn’t always mean starting from zero. Some underlying algorithms can be repurposed across organisations, while others must be tuned to the specific shape of the business.
A grocery drop may have relatively predictable delivery characteristics. A sofa delivery may take anything from 10 minutes to three hours, depending on access, installation and the reality of the customer’s home. The model has to understand the difference.
The hidden value in workforce and store optimisation
The same logic applies beyond vans and warehouses.
Hulme points to work with UK Telco, where machine learning was used to predict the nature of infrastructure faults, the skills of engineers and how long each engineer might take to solve a specific issue. The aim was to stop sending people to jobs they were not best equipped to complete.
“That project had a 200 times return on the investment,” he says.
In retail, a similar approach can be applied to store labour. Hulme cites DFS, where Satalia used machine learning to predict store footfall and customer demographics, then optimisation to allocate the right staff against the shape of demand.
Many retailers still roster teams in ways that are too blunt. A store may be staffed in similar patterns across the week, despite customer behaviour changing sharply by day, hour and demographic. Matching labour to demand is not just an efficiency play. It can directly influence sales, service quality and employee experience.
That wider view is important. Hulme argues AI should not be used to chase a single KPI in isolation. When it is, businesses can end up improving one metric while creating problems elsewhere.
He points to work with a leading accountancy firm, where Satalia built an algorithm to allocate thousands of auditors to jobs. The goal was not simply to increase utilisation. It was also to reduce travel time, improve employee happiness and strengthen client continuity.
“AI can improve all of your KPIs, not just one of them,” Hulme says. “If you focus on just one KPI, it can massively overachieve that goal, and by overachieving that goal, it can then actually cause harm elsewhere in the supply chain.”
For retail leaders, this may be one of the most important lessons. AI cannot be treated as a bolt-on efficiency project. It has to be understood as an operating model issue.
Start with frictions, not technology
If retailers want to move beyond hype, Hulme believes they should begin by listing the frictions across the organisation.
That means identifying where work is slow, repetitive, costly, unpredictable or constrained. Some of those problems may be solved with simple automation. Some may be addressed by buying mature third-party software. Others may require specialist AI expertise and custom-built solutions.
He describes it as a three-part strategy.
First, employees should be given access to tools that help them innovate at the edge of the business. Second, companies need to identify the hard problems where deep specialist expertise can create competitive advantage. Third, they should use partners and existing AI products for back-office tasks they do not need to build themselves.
“Start with listing all your frictions and then start knocking them off one by one by either building them, co-creating them or buying them,” he says.
This is also where the data conversation needs to mature.
For years, retailers have been told to build data lakes, unify everything and wait until their data is ready. Hulme is blunt about that approach.
“Don’t wait for your data lake to be ready. Your data will never be ready,” he says.
Instead, he argues that businesses should start with a clearly defined problem. Once the objective and constraints are understood, the necessary data becomes clearer.
In the accountancy firm example, the problem was how to allocate staff more effectively to jobs. That meant defining the objective function, such as maximising utilisation and minimising travel time, then mapping the constraints, such as availability, skills and client requirements. Only then does the data challenge become practical.
“A problem well defined is half solved,” Hulme says.
That does not mean data quality is unimportant. Poor data can produce poor decisions. But retailers should not confuse imperfect data with unusable data. In many cases, data issues only surface once a system is tested in the real world.
From optimisation to digital twins
The longer-term opportunity is not just solving individual operational problems. It is connecting those solutions together.
Hulme believes retailers should ultimately be working towards digital twins of their organisations, allowing them to model how one decision affects the wider system.
For example, a marketing campaign may increase demand by 10 per cent. But can suppliers cope? Is there enough warehouse space? Are there enough drivers? Can stores fulfil demand? Will the customer promise hold?
“Most retailers, because they are siloed, can’t project those questions across their supply chain,” Hulme says.
The promise of AI, he argues, is to create a simulation layer that allows retailers to test those scenarios before they become operational problems.
DFS offers a clear example of where this can lead. Satalia worked with the retailer on last-mile and middle-mile delivery, then helped build towards a broader digital twin. Hulme says DFS later platformised some of that delivery innovation through The Sofa Delivery Company, turning what began as an internal capability into a revenue-generating opportunity.
“That is the opportunity with AI,” he says. “If you build something that is genuinely differentiated, you can turn it into a revenue generator.”
Why quick wins can be a trap
Retailers are understandably under pressure to show progress. Boards want AI strategies. Shareholders want evidence of adoption. Leadership teams want quick wins.
But Hulme warns that quick wins can be misleading.
Most low-hanging fruit, he says, can be solved by third-party tools at a fraction of the cost. The problems that create true differentiation are usually not quick or easy.
“You need to focus on the problems that are going to differentiate your supply chain,” he says. “And those problems are not quick and they’re not easy.”
That creates a difficult challenge for C-suite leaders. They are being bombarded with AI vendors, consultancies and platforms. Many are being told to build internal teams, launch pilots or adopt the latest agentic tools, often without a clear view of what problem they are actually solving.
Hulme is sympathetic to that pressure but firm on the risk.
“Organisations can’t afford over the next three to five years to place the wrong bets,” he says.
The agentic AI ‘reality check’
No AI discussion in 2026 can avoid agents.
Agentic AI has rapidly become the industry’s latest obsession, promising autonomous systems that can take actions, complete tasks and collaborate with other tools or agents. Hulme believes agents will become hugely important, but he is clear that the market is still immature.
He compares the current agentic moment to the big data boom.
“People think everybody’s doing it, nobody’s doing it, but if they are doing it, they’re doing it badly,” he says.
The difference, he adds, is that agents will eventually drive real value. The danger is that companies deploy them before they know how to verify whether they work.
Large language models, he argues, are still like “intoxicated graduates”. They can be impressive, but giving them agency across a business without proper testing could create serious harm.
The issue is not just security or performance. It is functional verification. If an agent is built to optimise a media plan, for example, can the business prove it will do that well? If it is given a £1m budget, can the business be confident it will not spend it badly?
Hulme believes this will become one of the defining governance questions of the next phase of AI adoption.
The message is becoming increasingly clear. AI agents may be coming, but they need structure, accountability and verification before they are trusted with meaningful business decisions.
The real competitive divide
Hulme’s view of AI in retail is both optimistic and cautionary.
The technology can unlock capacity, reduce waste, improve service, support employees and create new revenue streams. But only when retailers understand the difference between shiny tools and strategic capability.
Understanding where those key frictions are, which problems are worth solving, which capabilities are differentiating, and which experts they need around the table, is fundamental to building AI-centric solutions that have longevity.
But that requires a shift in mindset. AI is a way to rethink how the business allocates resources, predicts demand, responds to complexity and makes decisions at scale.
That work is harder than launching a generative AI pilot, but it’s also where the real value lies.
Click here to sign up to Retail Gazette‘s free daily email newsletter
