AI is the accelerator. The transformation is the hard part.
AI is on the agenda of every board in European grocery, and the technology is ready. But AI only accelerates what it is pointed at. The value sits in transforming processes and redesigning ways of working. And that redesign is the hard part.
September 2026 · 14 minutes read
AI is the accelerator
in brief
There is no shortage of AI projects in European grocery, and that energy is exactly right. The
opportunity is to make each project worth more. A lot of pilots today get reported as a success,
yet the impact on profit is hardly visible — and that is not a ceiling. It is a starting
point.
The good news: this is not a technology problem, and it never was. The pilots succeed inside ways
of working that were never redesigned around them, so the gain stays on the one step the tool sits
on. That is the quick-fix trap: one tool, for one problem, in one step. Escaping it starts
with seeing AI for what it is — an accelerator of transformation, and not the transformation
itself. The real work is redesigning the process and the way of working, together with the people
who do it, into something built around AI from the start. That redesign is the hard part, and it
is exactly where the value sits, because AI accelerates whatever it is pointed at. None of this
asks for a big bang: it starts small, proves the value early, and lets the results carry the
argument.
AI is not new in grocery. It has been running for years in the systems that forecast demand,
reorder stock and recommend products. What is new is what the software does.
It used to analyse. Now it works with messy input, makes decisions and takes action: shoppers let
an assistant fill a ready-to-buy basket, and AI agents solve ordering problems without a human
checking every step. A survey of 330 retail executives worldwide puts AI at the heart of the
changes now reshaping how retailers sell, talk to customers and run their
operations.1
Analysing and acting are two very different jobs. Software that analyses produces a report, a
person reads it and decides, and the process does not have to change. Software that acts sits
inside the process. It has to understand the situation well enough to judge whether an
action makes sense here, in this store, this week; it needs a clear mandate for what it may settle
on its own; and it needs people who trust it enough not to redo the work. The model accelerates,
and the process around it decides what gets accelerated. That is the opportunity: the hard part is
within retailers’ own control. The industry’s own leadership already puts it near the top of the
agenda, and by its own account has yet to see it pay.
of grocery CEOs put AI and automation in their top three priorities
The ambition is there, and only cost and margin pressure ranks higher.
70%
see no measurable profit impact from AI yet, or say it is too early to tell
Which means the prize is still on the table.
Figures unchanged from McKinsey & Company / EuroCommerce, The State of Grocery Retail
Europe 2026.2
Two numbers from the same room of executives: the ambition is near the top of the agenda, and the
profit is still unclaimed. That is not a warning, it is an invitation. The technology works, and
the organisation around it is the one thing every retailer can change itself.
What has been built, almost everywhere, is stand-alone tools: technology aimed at one
specific problem in one specific step. A tool for the search bar, a tool for checking shelves, a
tool for finding suppliers. Each does its job well, but the process it landed in stayed exactly as
it was, and so the gain stops at that step. That is the quick-fix trap, the AI version of
the standard-solution trap we have always rejected. This piece is about the value that opens up
once a retailer steps out of it.
Our position is simple: transformation is the goal, and AI is the most powerful accelerator we
have today. An accelerator only helps if there is something to accelerate — a process being
redesigned and a way of working being rethought. That redesign is the hard part, and the hard part
is where the profit lives.
Five lessons that unlock the profit
Count the AI projects running in a European grocery group and the energy is real. The next step is
to make that energy earn. Most projects today are defined around a single step, and value that
stays on one step of one process, in one chain, stays too small to see. Widen the frame and the
same effort starts to show.
We see five lessons — five places where the value of a project gets settled early and settled
low today, and where a different choice settles it high. Each one is practical, and each one adds
to the profit.
lesson
Solve the real problem, not the perceived one
Find the true bottleneck first and the same technology suddenly pays.
Most AI projects are defined in a meeting room, from a process diagram, by people who are
responsible for the process but do not do the work themselves. The specification that comes
out makes sense on paper, but it does not necessarily hit the biggest pain point. The
technology then performs perfectly against it, and that is exactly why the failure is so hard
to see: nothing broke.
from our work
A European grocery group asked us to look at a supplier development agent: a tool to
find, profile and vet alternative suppliers for its own-brand products. Own brands are
around 40 percent of what European grocery sells, at gross margins up to 35
percent,3 and they are the only part of
the range where the retailer writes the product specification and carries the risk. In one
market, three quarters of the own-brand products had no approved alternative supplier.
The first research turned up a question nobody could answer. Some products already had six
approved suppliers, and the buyer still had not switched. If six alternatives did not
trigger a move, a seventh found faster would not either.
A thin supplier base is a real cause, and widening it is real value, but it is not the only
cause. Other things stood in the way too: certification, capacity, delivery rules per
country, product testing, and a purchasing manager with no room in the week to prepare a
supplier auction. The agent is still worth building, and it is being built, but it was aimed
at one cause rather than the full set.
Nobody was careless. What was missing is a proper search for the root cause, and specifically
the simple technique of the five whys: ask why the problem occurs, then ask why of that
answer, about five times, until the answer stops describing a symptom. Done properly it rarely
ends at one cause, and the discipline is to map them all before choosing which one to build
against. Here the first plausible answer arrived with a business case attached, and so the
questioning stopped. The lesson is simple and energising: five good questions, asked before
the build, multiply what the build is worth.
lesson
Give the data the strength to carry it
Fix the picture of the business and every model on top gets better.
An AI system does not act on the world. It acts on the picture of the world it is given. In
grocery that picture is product catalogues, stock records and sales forecasts, and none of
them were built to be read by software that makes decisions. That picture is the ceiling on
everything above it, and no choice of model raises it.
from our work
A European grocery retailer we work with had an online search problem that looked
like a search problem: type “dishwashing liquid” into the webshop and the top results came
back as chicken. The search engine was not broken. The product data behind it was incomplete
and getting worse, because a large part of the catalogue changes regularly, and enriching
it by hand could never keep up.
The answer was not a smarter search engine. AI was pointed at the data itself: image models
read the product photos for labels, packaging text and brand, and a language model wrote the
descriptions, keywords and categories that search needs. The result came fast. In a live
test where part of the visitors got the improved data:
+6%
orders
+5%
revenue per visitor
No new search technology was bought, and that is the exciting part: the value was already in
the building, waiting in the data.
lesson
Redesign the process, do not just speed it up
The biggest gains sit in the steps that can be removed, not accelerated.
McKinsey estimates that 40 to 50 percent of the routine work in grocery head offices is within
reach of AI agents.4 That is an enormous
amount of time to give back to better work. The gain depends on where the agent is pointed:
aimed at the existing process — one built around a weekly meeting, a monthly cycle and a
reporting chain — it makes one step faster, and that is worth something, but not much.
Aimed at the process itself, it changes what the work is.
from our work, and one step beyond it
A grocer checked whether products were actually on the shelf by sending someone to look:
once a month, in 50 of more than 260 stores, on a sample of products. That rhythm was never
a design decision. It is what a retailer can afford when detection means a person walking an
aisle.
The automation answer for keeping shelves filled, in the trade called
on-shelf availability (OSA), was a better audit: a faster scanner, a smarter route,
the same monthly sample. What was built instead removes the sampling altogether. Three
signals from data the business already had — stock, daily sales and the daily forecast
— flag where the shelf is probably empty, for every product in every store, every day.
1x/month
50 of 260+ stores, sampled
≥2%
of sales lost to empty shelves today
Tested in a second market against a manual process that has been running since 2019, it gave
comparable results without the walk. But it is a better detector, not a redesigned process:
the sequence is still detect, report, send someone to look.
Carrefour shows how exciting the next step is: in its connected-store pilot the shelf
reports on itself, through smart rails, cameras and electronic price
labels.5 Detecting is then no longer a
task anyone performs, and what reaches the store employee is not a report but the next
action in the aisle.
the process as it wasManual audit
automationDetection from data signals
redesignThe shelf reports itself
What detects the gap
A store employee walking the aisle, in a sample of stores
Stock, daily sales and the daily forecast — every product, every store
Nobody. Smart rails, cameras and electronic labels — the shelf reports itself
How often
Monthly
Daily
Continuous
Shape of the process
detect→report→check
detect→report→checkIdentical to the column on its left
detect→actThe detect step is gone, not accelerated
Exhibit — three answers to the same question: is the product on the shelf? The first
two approaches share a sequence, and only the third removes a step from
it.5
On-shelf availability (OSA) is not the same as having stock. A product can sit in the
back room all day and still be, for the customer, out of stock. In-store availability
(ISA) is stock in the building, and OSA is stock the customer can actually reach. On
sector benchmarks, one extra point of OSA is worth roughly a third to half a percent of
sales.6
Both are AI, and only one is a redesign. The test is not how much of the work the technology
does, but whether the sequence survives it. A monthly audit made continuous is still an audit.
A shelf that reports on itself has no audit left in it, and every process around it can be
rethought. That is not harder technically. It is harder organisationally, and that is good
news, because the organisation is something a retailer owns.
lesson
Launch above the threshold of trust
A solution users can rely on gets adopted, and adoption is where the value starts.
A solution that is right most of the time sounds like a success and behaves like a risk,
because users cannot tell which answers are the wrong ones. So they check all of them, and the
tool becomes an extra step in a process it was meant to lighten.
from our work
A tool that analyses customer reviews reads app store reviews across markets and
languages, works out what customers feel and what they talk about, and lets product, design
and customer care teams question that data in plain language. Asked about a drop in ratings
after a particular release, it explained the drop. There had been no drop after that
release.
The answer was fluent, specific and plausible, and that is exactly what made it damaging: a
product manager cannot tell it apart from a correct answer without going back to the
reviews, and that is the work the tool was meant to remove. One proven made-up answer is
enough: after it, every answer has to be checked, and a source that has to be checked is not
a source.
It fell below the bar at which a busy person stops double-checking. Where that bar sits is for
the business to decide, not the data team, and the fix is refreshingly simple: write it down
before the build starts. We do, and we call it minimum viable accuracy, and it protects
the most valuable asset an AI solution has — the trust of its users.
lesson
Build once, use everywhere
Shared foundations turn every local win into a group-wide one.
AI in a large grocery group rarely has one home. It has several: a central expert team at
group level, a data team inside each chain, an innovation unit, and whichever department moved
first on its own budget. Each behaves sensibly on its own, and together they build the same
capability several times over. The flip side is where the opportunity sits: a group that
shares what it builds turns every local win into a group-wide one, at a fraction of the cost.
from our work
In a grocery group with several chains, nobody could answer a simple question: what AI is
already running, and where? Many parts of the organisation were building on their own.
We are working with the group on three things at once: a clear division of
AI roles, so the centre and the local chains each know what they own; a structured
canvas that scores use cases on value, feasibility and risk; and one shared
list of all live AI projects, each checked against the group’s own AI risk rules.
That list makes the argument by itself: what exists becomes visible, compliance becomes
visible, and every chain can reuse what its neighbour already built.
In a sector where margins have sat at 3 to 4 percent for close to a
decade,7 building once and reusing
everywhere is one of the cheapest competitive advantages available, and it makes the other
four lessons compound, because every insight travels across the group.
The opportunity — out of the quick-fix trap
Five lessons, one message: AI is the accelerator, and the transformation of processes and ways of
working is the hard part. Retailers who do that hard work will find profit the quick fixes never
touched, and they can start tomorrow, because none of it requires a technology breakthrough.
The way of working is proven. At HazelHeartwood we work by
five design principles we never compromise on, and
three of them turn these lessons into daily practice.
principle 1
Gemba-first, technology second
The problem lives where the real work happens: the shop floor, the planning desk, the service
line. That is where we start, understanding the real pain point before any tool is considered.
Workshops help, provided the people who do the work are in the room, but they cannot replace
‘going to see’. If the value cannot be defined, we do not start the build. The technology is
usually the last 20 percent of the work, and understanding why things are the way they are,
and whether they should stay that way, is the other 80.
Gemba is the lean word for the place where the real work happens: the shop floor, the
planning desk, the picking line. It is where the work is really done, as opposed to where it is
described in a diagram. Gemba-first means the problem gets defined there, by the people living it,
before anyone develops or chooses a tool.
That answers lessons 1, 2 and 3. A pilot defined far from the work solves a symptom that was never
the real bottleneck, builds on data whose weaknesses only the people using it could have
described, and inherits a way of working nobody at the desk would have defended. No amount of
model quality rescues any of the three.
principle 2
Tangible impact (ROI) is the only metric
A solution nobody uses is pure waste of money. We link every initiative to one or more tangible
value KPIs, and we measure the return at every step. End users are involved early, behaviour
change is measured, and we do not sign off until the new way of working is the standard way of
working.
Counting live tools and launched pilots is the wrong measure: that number climbs all year while
the profit stays more or less unchanged. The figures that decide are the ones the business already
steers by: margin, waste, product availability, hours freed up and where those hours went. In a
survey by BCG and the Consumer Goods Forum among senior retail and consumer goods executives, more
than half of companies do not formally measure the return on their AI investments at
all.8
Minimum viable accuracy (MVA): an AI solution that is almost right is not right. MVA is the
accuracy level a solution must reach before it goes anywhere near real operations, agreed up front
with the business, for each use case. Only solutions that have reached their MVA are
operationalised, and only solutions with proven demand and business pull get that far. This
protects adoption and trust, because end users only work with solutions they can rely on. That is
what answers lesson 4.
principle 3
Staircase approach, not a big-bang transformation
Tangible outcomes at every step, not a promise at the end. Each step delivers value on its own
and builds the technical foundation the next one needs. We combine rapid experimentation with
regular feedback, and we only scale what has proven business pull. No long-term theoretical
plans.
‘Going to see’ first defines the right problem, and tangible impact and MVA define when a solution
is ready. What remains is how far the ambition goes, and this is where the staircase gets
exciting: every step is available today, each one is worth more than the last, and the steps are
not delivery milestones. They are levels of how much the work itself changes, each resting on a
deeper technical foundation than the last.
The staircase: value rises with how much of the process changes; below the line, the technical
foundation each step needs. In our experience, most grocery AI sits on the first two steps.
The first two steps are worth having, but neither touches the shape of the work, and that is why
they can be counted in the hundreds without any noticeable effect on the profit. The third is a
real break, the system stops proposing and starts acting. The fourth is the destination worth
aiming for, because it changes what a department is: people move to genuinely new work that adds
value — setting the rules the agents run on, judging the exceptions they pass up, and owning
the outcome instead of executing the steps. That is not people being replaced. It is people moving
up to the work that only people can do.
The bottom row says why the fourth step is worth the climb: it is gated by foundation, not by
ambition or budget. A process built around agents needs the same data across markets, a shared
layer of business context, coordination between agents and clear rules at group level. No retailer
can buy its way onto the fourth step, and that is exactly why it is defensible: the foundation
built today is an advantage competitors cannot copy overnight. Scaling is a strategic decision and
not a technical progression.
Which brings the story back to where it started. AI accelerates whatever it is pointed at. Point
it at an unchanged process and it accelerates the status quo. Point it at a redesigned way of
working and it accelerates the transformation, and that is where the profit lives. The rethink
starts where the work happens, because a process can only be redesigned by the people who know why
it looks the way it does, and every round puts something real into operations, at an accuracy the
business agreed in advance. Fewer isolated experiments, more processes and ways of working rebuilt
with AI as the accelerator, and one shared foundation that makes reuse the default: the retailers
who take on the hard part today are the ones who will be reporting profit impact in 2027, and it
is a choice, not a lottery.
Bring us a process, not a tool idea.
Most conversations about data and AI start with an idea for a tool. We love the enthusiasm, and we
start one step earlier: at a place where the result is stuck and nobody is quite sure why, because
that is where the biggest wins hide.
One process where the numbers have not moved despite the effort: we go to where the work happens
and find the root cause together with the team.
Or one project stuck in pilot: we walk through the MVA check and say honestly whether it should
be scaled up or stopped.
We do not pretend to know the answers up front, but we are very good at finding the real problem
and the solutions that fit the reality on the ground. A first conversation costs nothing but an
hour.
Deloitte, 2026 Retail Industry Outlook, a survey of 330 global retail executives,
October–November 2025.
deloitte.com↩
McKinsey & Company / EuroCommerce, The State of Grocery Retail Europe 2026, CEO
Survey 2026 (n = 36): 47 percent rank AI and automation top-three, and 70 percent report no
measurable profit (EBIT) impact yet, or say it is too early to tell.
mckinsey.com↩a↩b
Own-brand (private-label) value share of 40 percent across the EU-11 in 2025: McKinsey /
EuroCommerce, as above. Gross margins up to 35 percent: Strategy& / PwC,
Grocery Retail Outlook 2026.
mckinsey.comstrategyand.pwc.com↩
McKinsey / EuroCommerce, as above: “40 to 50 percent of routine activities in grocery headquarters
could be automated by AI agents.”
mckinsey.com↩
Carrefour and VusionGroup, EdgeSense connected-store pilot, press release, 12 June 2025. Villabé
hypermarket: ~70,000 electronic shelf labels, 500 cameras, 7,000 rails, and shelf cameras live in
35 stores.
vusion.com↩a↩b
Colin Peacock, Managing Shelf Out of Stocks — The Role of Loss Prevention, ECR
Retail Loss (ECR Community a.s.b.l., Brussels), 18 August 2020. Modelled on a European
out-of-stock baseline of 8.6 percent (Corsten and Gruen, 2003): eliminating out-of-stocks
entirely would lift sales by 3.7 percent and halving them by 1.8 percent — about 0.4
percentage points of sales per point of availability.
ecrloss.com↩
Strategy& / PwC, Grocery Retail Outlook 2026: European grocery EBITDA margins at 3 to
4 percent for close to a decade.
strategyand.pwc.com↩
BCG and The Consumer Goods Forum, AI in CPG and Retail: How Winners Are Pulling Ahead,
June 2026, a survey of 39 senior retail and consumer goods executives.
bcg.com↩