Agentic transformation in Retail, Consumer and Media

In high-volume, reversible-decision workflows: merchandising and demand signals, supply chain exception handling, customer service triage with human resolution, and content production at volume. These have short feedback loops, contained risk, and produce measurable results inside a single trading cycle.

Yes. The unfair commercial practices rules apply to the trader, not to the author, so a misleading description is a misleading description whether a copywriter or a model produced it. The rules in the Digital Markets, Competition and Consumers Act apply to practices from 6 April 2025, and the CMA can now decide for itself that they have been broken rather than going to court first, with penalties of up to 10% of global turnover and the power to direct redress. The practical design response is to ground every factual claim about a product in a field in a system of record, and to hold anything the model asserts that cannot be matched to one. The same logic covers price display, where mandatory fees have to appear in the total up front, and reviews, where presenting incentivised reviews as genuine is a banned practice.

Almost always, and you need it more than once. A DPIA is required where processing is likely to result in high risk, which covers most personalisation, profiling and large-scale use of customer data. The failure we see is not a missing DPIA, it is one completed for a narrow pilot and never revisited when autonomy widened, which is the change that altered the risk. Two other things get missed. A retrieval corpus needs its own retention and deletion behaviour, because honouring a deletion request in the database and not in the index is not honouring it. And a decision with significant effect on a person needs a real route to human intervention and to contest the outcome. Tenhaw is not your DPO and gives no legal advice here.

It depends on the tool's terms, the licences behind its training data and the warranties you were able to negotiate. UK law still has no broad commercial text-and-data-mining exception to fall back on, and the government's copyright and AI report of 18 March 2026 dropped a broad exception with an opt-out as its preferred approach, saying it would gather further evidence instead. The engineering answer is more useful than the legal one, so record provenance at the moment of generation, which tool, which version, which inputs, which licence, and keep it with the asset. Most rights problems in media are discovered eighteen months later, and the difference between an hour of work and a month of work is whether anyone can say where the asset came from.

By treating the change freeze as a design constraint from the start. Delivery is sequenced so that pilots ship and stabilise before freeze, the freeze period is used for adoption, measurement and operating-model work that requires no deployment, and the next build window is planned against the trading calendar rather than a generic quarterly plan.

Almost everything. Head office knowledge workers adopt tools that make their own work easier and have discretion over how they work. Frontline staff work to fixed processes, often on shared devices, with little discretion and immediate customer pressure. The two require separate adoption designs, separate measurement, and usually separate sequencing.

Yes, and the control that matters is knowing which reviews fed the summary. Presenting fake or incentivised reviews as genuine is a banned practice under the unfair commercial practices rules, and the trader carries a duty to take reasonable steps to stop them appearing, so a summariser that quietly folds incentivised reviews into a headline average produces a number nobody can defend. Build it so every claim the summary makes traces back to reviews you can identify, and hold anything the model asserts that the underlying reviews do not support. We test the claim path rather than the tone. Where puffery ends and a misleading claim begins is your legal team's call, not ours.

Only behind a rules layer it cannot talk its way around. The total price including mandatory fees has to be presented up front rather than assembled during checkout. Price display is regulated in its own right, so a promotion agent that splits a mandatory fee off the headline price has produced a misleading practice on the trader's behalf. Since 6 April 2025 the CMA decides for itself whether consumer law has been infringed rather than litigating first, with penalties of up to 10% of global turnover, or £300,000 if that is greater, and the power to direct redress. Put the pricing rules in code the agent cannot override, and test what the display claims rather than how it reads.

Not automatically. UK GDPR asks for a lawful basis and a purpose the data was actually collected for, and data gathered to fulfil an order was not obviously collected to train a recommendation model, which is where retail personalisation gets uncomfortable. That does not make it impossible, it makes purpose and lawful basis inputs to the design rather than a gate at the end. The Data (Use and Access) Act 2025 reworked parts of the regime, and it is the ICO's guidance rather than the headlines that a DPO holds a programme to. Our default in pilots is synthetic or mocked data, which takes the hardest approval out of the fastest-moving phase. We work to what your DPO decides.

On the retail side, under the Tenhaw banner: five agile teams coordinated through YOOX NET-A-PORTER's £1bn e-commerce re-platform, the mobile app and integration squads at Greggs made predictable and trusted again with data showing that clearing tech debt sped delivery up, and three newly merged teams at Colart turned into one unit that shipped an e-commerce site. On the media side, James Rooney ran the visual rebrand team on the Discovery+ and Eurosport launch, six teams against a date the CEO had already announced, and held delivery roles at Sky. That work was delivery transformation rather than agentic build, and we label it that way because the two are not the same claim.

A retail AI operating model comes down to four choices: which workflows agents touch, where the human sits in each one, how the build is sequenced against the trading calendar, and who owns what an agent says to a customer. The last one is the trap. Anything an agent generates that reaches a customer is a commercial practice by the trader, so the claim path belongs in the design rather than in a review at the end. The other structural point is that head office and store or contact-centre operations are two different transformations, and running them as one programme is a reliable way to fail at both. Peak freezes make sequencing a design constraint, not a scheduling detail.

Because seasonal and promotional data makes naive forecasting agents unreliable, and retail history is mostly seasons and promotions. The model reads a promotional spike as demand rather than as a lever somebody pulled, then carries that shape into a week when nobody is pulling it. The version that works feeds the promotional calendar and price changes in as explicit inputs instead of leaving the agent to infer them, and keeps a buyer between the recommendation and the purchase order. Merchandising and demand signals are still one of the fastest places in retail to get a return, because the decisions are frequent and reversible. What fails is giving the forecast to an agent and the promotional calendar to nobody.

Yes, and it is one of the first places we would look, because that is variance humans currently absorb by hand and the volume is high enough to show a result inside one trading cycle. Every time, the agent assembles the exception, a late shipment, a short delivery, a substitution, pulls the related records and the history, proposes a resolution and hands a person a decision instead of a search. Two boundaries hold. It does not commit money or change a customer's order on its own, and anything it produces that reaches the customer, a revised delivery promise included, is a commercial practice by the trader and needs grounding in a system of record.

By putting the boundary in a different place from where you would put it in the back office. Get a customer interaction wrong in retail and it is a social media post the same afternoon, which is a different risk profile from mislabelling an invoice, so the agent drafts and a person resolves rather than the other way round. Start where a mistake is recoverable: triage, summarisation, routing, and internal drafting that a colleague sends. Widen autonomy only where the evidence from the narrow version supports it. The failure mode is rarely a rude reply. It is a confident factual assertion about a product, a price or a delivery date that nobody checked.

By buying it in pieces small enough to prove themselves inside a trading cycle. The commercial case for agents is strongest where margins are thinnest, which is also where the appetite for multi-year consultancy spend is lowest, so the shape of the spend has to match. An agentic proof of concept is £20k–£55k fixed over 2–4 weeks. The AI readiness audit is £44,000 fixed over four weeks and ends in working prototypes, and a recommendation to stop is a valid outcome. After that, engagements are retainer-shaped rather than milestone-shaped, and a month that delivers no measurable value gets reported as a failed month. That is the honest version of return inside a trading cycle.

Yes, and it is usually the least contentious place to start, because what gets replaced is manual consolidation rather than judgement. Store and field reporting in most retailers is someone pulling figures out of several systems by hand every week, which is high volume, low risk and easy to check, because the numbers already exist to reconcile against. The design work is not the model. It is deciding what the report is for, cutting the fields nobody acts on, and giving the frontline something back rather than only feeding head office. Adoption is the risk here rather than the build. A store team will use a report that saves them an evening and ignore one written for head office.

A trading one with engineering inside it. If IT owns it, you get something merchandising did not ask for, store operations cannot use, and a frontline half of the programme that quietly never happens. The sponsor should be whoever owns the trading outcome, with the merchandising or category lead, whoever runs stores or the contact centre, brand or legal for the line between puffery and a misleading claim, and your DPO in the design sessions rather than in the approval queue. Engineering reports into that group instead of running alongside it. If nobody in the room can approve a change to a customer-facing message, the wrong people are in the room.