Burberry: can technology tailor a turnaround?

An awards page says a three-person team at Burberry built an AI tool for client advisers, and that average transaction value in customer-service channels rose 24 per cent. If you run a small fashion brand and someone is selling you an AI tool for customer service, how far can that number be trusted? This post uses three public sources (as of 3 November 2026) and keeps facts, company claims and my own inference apart. See also the Burberry case.

What do the three public sources say about Burberry?

The AP report of 15 July 2024 (the copy archived on the Wayback Machine) says Burberry appointed Schulman, 52, to replace Jonathan Akeroyd. The same day the company said first-quarter revenue fell 21 per cent, with declines in every region except Japan. Chair Gerry Murphy said that if the trend persisted through the second quarter, Burberry expected a first-half operating loss, and that the company had moved quickly with its creative transition in a luxury market proving harder than expected (Associated Press, 2024). The report does not mention technology.

Maghan McDowell's Vogue Business article of 30 July 2020 covers Burberry's first "social retail" store, in Shenzhen, built with Tencent and run on a WeChat mini-app. According to Mark Morris, then Burberry's senior VP of digital commerce, a customer-service message sent through it connects automatically to a second mini-app for store associates, who can see the customer's information (McDowell, 2020).

The DataIQ page describes Penguin, a generative-AI platform for luxury client advisers. Burberry's Data Science team, "just three specialists strong", worked with Customer Service on three functions: search in the client's own words, recognising products in a photo and matching them to the catalogue, and suggesting outfits from a single item, fine-tuned on looks curated by Burberry. The outcomes section lists more than 100 consultants, a 24 per cent uplift in average transaction value within customer-service channels, and higher productivity and satisfaction, none of it quantified (DataIQ, 2025).

What does the 24 per cent rest on?

A bare number. The page does not say against what (before and after Penguin, or advisers using it against advisers not using it), over how long, across how many orders, or how average transaction value was calculated (DataIQ, 2025). The text names no author either. The library classes it as an entry the company submitted, and the page's structure fits that; the post could not verify it.

Average transaction value is a ratio: revenue divided by the number of orders. It rises when customers buy more per order, but also when small orders disappear. A hypothetical example, with numbers I made up and no connection to Burberry:

Average order value is up exactly 24 per cent while revenue is down 0.8 per cent. The DataIQ page gives no order counts, so nobody outside Burberry knows which row they are in. Nor does it rule out other readings: early adopters among the advisers may simply sell better, and customers who seek out the advice channel may differ from the rest. Lemon and Verhoef (2016) raise a related problem for customers who switch channels: they differ inherently from those who do not, so the selection effect has to be handled (Lemon and Verhoef, 2016, p. 82). Carrying that over to an advice channel is my analogy. See also selection bias.

Which statements are fact, which are claims, which are my inference?

Facts are what the press reported, claims are what the company says about itself, inference is mine. The figures in the AP report are numbers Burberry announced, relayed by AP; I did not compare them with the company's own release.

So the straight answer to the title: the public sources used here do not support saying technology tailored the turnaround.

Where does Penguin sit among the three types of analytics?

In my reading, somewhere in the middle, and that is not a weakness. The three types of analytics are split by question: what happened, what could happen, what should we do. The abstract of Lepenioti et al. (2020) says prescriptive analytics seeks the best course of action and is often seen as the next step in analytics maturity, after descriptive and predictive analytics (Lepenioti et al., 2020).

The DataIQ page names the starting problem as advisers working with systems built for consumers, so finding products was manual and slow (DataIQ, 2025). Searching in a client's own words and matching photos to the catalogue is lookup, closer to the descriptive layer than the predictive. Only the outfit suggestions resemble advice, and the page says they were fine-tuned on curated looks, so they follow the house's taste; it never says the tool optimises for sales. I would not call it prescriptive in the optimising sense, with the limit that the post uses only the Lepenioti abstract.

Public sources do not say whether Burberry had descriptive reports or forecasting models before Penguin, so nobody can claim it climbed the ladder in order. And the order is a convention, not a law; it is useful in one respect: without average order value by channel and by adviser from before, you have nothing to compare when someone offers "24 per cent".

Which signals along the path to purchase are worth measuring first?

Those at the points where customers meet you, not inside the tool. The library draws the customer journey tied to data as five stages:

Lemon and Verhoef (2016), the source the library gives for this framework, do not have those five stages. They use three: prepurchase, purchase and postpurchase; postpurchase covers use, engagement and service requests (Lemon and Verhoef, 2016, p. 76). They also sort touchpoints into four groups (brand-owned, partner-owned, customer-owned, social or external) and say a customer may meet all four at every stage (Lemon and Verhoef, 2016, p. 76). So treat the five-stage version as the library's practical split, not the paper's.

The Burberry material shows two joins. A client's words in the DataIQ example, "the coat I saw in store last week", carry a memory from the shop into the advice channel (DataIQ, 2025). In Shenzhen the customer-service channel connected straight into the associates' mini-app, yet Morris conceded that elsewhere Burberry has to work with multiple parts because no ecosystem is as widely used as WeChat (McDowell, 2020).

For a small brand I read three measurable places: what customers type when they ask (before purchase); close rate and order value by channel (purchase); returns and after-sale questions (after purchase). That is my inference, not drawn from any source.

In what order should a small brand proceed?

Measure first, run a controlled trial second, buy last. This is my own usage, not something the sources say Burberry did.

1. Build a baseline in a spreadsheet: average order value, order count and close rate for the advice channel, by adviser and by month, for several months before any trial.
2. Pick one question the tool must improve, such as the time to find the right item for a customer, rather than "raise revenue".
3. Trial it with some advisers or some of the customer flow, keep the rest as a comparison, and compare both order count and order value.
4. Think about prediction only when your order history is long enough; I have no source for how long that is.
5. Put machine-suggested outfits last, with the adviser still working with the customer.

This suits you if you have a few advisers, a catalogue with consistent names and photos, and patience for a few months of testing. It does not suit you if you handle a few chats a day: then a searchable catalogue and a sheet of model answers cost less and are easier to measure. If the tool talks to customers directly, the risk changes; see Air Canada's chatbot.

Going deeper: what does each source cover, and what is unchecked?

The DataIQ page was published on 5 November 2025 according to its metadata. Its text names no author (its metadata lists only a site author account), and the post relies on the library's classification of it as an entry the company submitted. AP's own site blocked automated access, so the post uses the copy archived on the Wayback Machine. It uses no document published by Burberry itself.

For Lemon and Verhoef (2016) the post uses the University of Groningen repository copy, and only the passages on stages, touchpoints and selection effects. For Lepenioti et al. (2020) it uses only the abstract on RePEc. It does not use Burberry's financial reports, so it says nothing about results after July 2024.