Not always a person will be a person: five hours in Warsaw that reframed retail AI

It is a little after nine on a Friday morning at Google Campus Warsaw, and the room fills the way industry rooms do: retailers in one cluster, investors in another, a few people still working their phones. Cenk Karacaev, co-founder of Replenit, opens by thanking the room, and notes that the Turkish ambassador to Poland is among the guests. Then he hands the stage to Google.

An easy yes, twenty years later

Michal Kramarz, Head of Incubation & Acceleration at Google Cloud, does not start with AI. He starts in 2006, when he was one of Google's first employees in Poland and retail was worried about entirely different things. Allegro was fighting eBay for the market. Its staff were handing out flyers outside a newly opened Warsaw shopping centre, telling shoppers they could find it cheaper online. And Kramarz was sitting with the founders of Media Expert, then a small company, trying to convince them that launching a website would not kill their stores.

"Which you can see now has kind of changed a bit," he says, and the laugh that follows is the laugh of people who were there.

His argument is that retail's job description keeps being rewritten. In 2006 the work was getting the product online. Two decades later:

Now we're in the moment that retail is about intelligence. Retail is about utilizing artificial intelligence to predict what the customer wants.

He offers evidence in the form of a shopping anecdote rather than a chart: a package ordered on eBay in the US over the summer took seven days, while an order placed at Media Expert two days before the summit arrived the same day. Polish logistics, he suggests, has quietly overtaken markets that like to call themselves developed.

The opening also has a local subplot. This campus space opened in 2015 and has trained roughly 200,000 people since. In 2025, a group of founders from Turkey applied for a residency here with an idea about retail intelligence. "For us it was an easy yes," Kramarz says, and adds the detail that gives the morning its shape: they took their first Polish lessons in a classroom one floor below, and they turned up to everything, always in the front row.

That group was Replenit. This morning, they are the ones hosting.

Two halves of a market that are about to collide

Olaf Piotrowski, who spent thirteen years running data, operations and AI at Allegro before moving to Google Cloud, takes the stage and rules out a product pitch. What he wants to do is connect two halves of global retail that most people in the room only see one of.

The East, now 55% of global e-commerce revenue, did not build better websites. It collapsed the whole journey into single ecosystems the customer never leaves. Live commerce moves hundreds of products in one broadcast and is worth over $900B in China alone, roughly three quarters of total US e-commerce. Quick commerce delivers in 10 to 30 minutes from some 10,000 dark stores, the equivalent of 90,000 convenience shops. In South Korea, 75% of purchases happen on a phone and 70% of shoppers find the product on social media first. The interesting detail is the motive: those models are not optimised for speed, Piotrowski says, but for certainty. Chinese shoppers pay more when they are sure the half hour holds.

The West spent its energy elsewhere: measurement, monetisation, retail media networks that now take over 20% of total media spend. His verdict on that is blunt. There are more than 270 retail media networks in the region, brands typically run eight or more at once, and cross-channel incrementality remains the biggest shared headache of retailers and brands alike.

Then he puts up the sentence that quiets the room.

Not always a person will be a person. People will not in every situation start the e-commerce journey, and they will not in every situation finalize the deal.

The numbers behind it are not speculative. McKinsey projects $3 trillion in agent-triggered transaction revenue by 2030. Morgan Stanley's more conservative case is 10 to 20% of all retail transactions by the same year. Adobe research puts agent-initiated purchases at roughly 30% higher conversion probability. Alipay already clears more than 120 million agent-driven transactions a week.

For anyone who buys media, that changes the shape of the job. "The funnel is compressing," Piotrowski says: impression and conversion land in the same moment, and the ad moves out of the search results and into the conversation. The plumbing is arriving in the form of protocols, first OpenAI and Stripe's payment-focused effort, then Google's broader Unified Commerce Protocol covering discovery, live inventory, checkout and refunds. His expectation is that retailers will end up speaking several of them, because nothing has consolidated yet.

There is a quieter line in the keynote that operations people write down. Overnight batch processing, the thing that produces tomorrow's demand and inventory picture, does not survive in a world where demand is created live on a stream and inventory has to answer in real time.

Asked implicitly what to do with all this, he answers with fundamentals rather than strategy. Structured product data, real-time inventory signals, attributes written so a machine can read them.

Structured product data is the gate to engage with the agentic infrastructure of commerce successfully.

He closes with a correction that follows the day around: stop calling the destination personalisation. "The new reality is not personalization anymore, but a very individual purchase experience." No buckets, no general messaging.

Maestro, and what it means to hire software

Ilyas Kurklu, Replenit's co-founder and CEO, comes on to launch the product the summit was built around, and he begins with the gap it is meant to close. A year ago retail leaders said 54% of their AI investment was working. This year the same question returns 38%.

So where did these 16 points go, and why?

His diagnosis is short. Assistants wait to be prompted and depend on your team's bandwidth, and are never accountable for the outcome. Agents mostly are not agents: of roughly 7,400 AI startups funded since ChatGPT's launch, he cites Gartner counting about 130 that deliver a genuinely agentic experience. The rest are "just cosplay of the assistant model," without real memory, judgment or the right to execute. Meanwhile the stack keeps growing, with 91% of retailers adding tools, about a third of paid capability actually used, and teams losing up to 10% of their time in between. "Those tools were never designed to own the outcome," he says.

Which is the setup for what Maestro is: not another platform in the stack, but a manager placed above it.

The launch film makes the distinction in a run of short lines. Not build segments: decide for one. Not automate: act. Not analyse: enrich. Not store: remember. Not follow rules: bring skills.

In practice, Replenit describes Maestro, the world's first AI CRM Manager, through four verbs: it reasons, remembers, decides and executes. Each one maps to something the CRM stack currently pushes back onto people.

It owns whole workflows, not campaigns. Cross-sell, replenishment, engagement, churn and win-back are handled end to end, rather than delivered as fifty disconnected placements chasing the same objective. Kurklu's example of the current state is one every retailer recognises: everyone agrees cross-sell matters, and then cross-sell lives in dozens of unconnected campaigns and rules.

It decides at the level of one customer. For each person, Maestro determines what happens next, why now, which product, in which tone, and toward which commercial outcome. The company's phrase for it is a segment of one, and it is positioned against the practice of putting thousands of people in a bucket and hoping the same message produces the same impulse.

It has living memory. Not a database of events, but a working picture of the customer and the catalogue together: what someone likely is, what drives their purchases, what they reorder and on what cycle, what they will never buy, what was decided last time and what happened, and what each product replaces when it runs out.

It runs on retail skills rather than rules you maintain. A beauty retailer gets something like a skincare routine builder; a fashion retailer gets seasonal skills instead. The skills are the part a team builds and extends, in place of writing and babysitting rule sets.

It executes inside the tools already in place. Nothing gets replaced, nothing gets migrated, which is also the answer to the change-management objection that usually kills projects like this.

The framing is a hiring decision. Today, Kurklu argues, AI is the co-pilot and the human is the driver, carrying the outcome personally. Maestro swaps the seats: the machine drives execution, the team sets direction, priorities and judgment calls, and the work is measured against a commercial number rather than activity. If it does not deliver, the accountability sits with the hire.

Maestro is already hired. It works with the stack you already have. Don't buy more software. Hire Maestro.

The results he puts behind it come from live customers. L'Occitane lifted post-purchase performance by 235%. Faith in Nature attributes 12.7% of total company revenue to Maestro. One of Europe's largest daily deal platforms, operating in eight countries, reports $16.6 returned for every dollar invested. The system runs on Google Cloud, is built with two academics working on the underlying models, and counts ElevenLabs CEO Mati Staniszewski among Replenit's backers.

The room breaks for twenty minutes. There are bags at the back with the retail magazine Replenit publishes, and Kurklu points people toward the Walmart piece in it on the way out.

The panel, where the agreement broke

After the break, Bartosz Lipnicki, Managing Director of Endeavor Poland, takes the moderator's chair, discloses that he is a small angel investor in Replenit, and refuses the premise of most AI headlines: search is not dead, marketplaces still hold the traffic, and what has actually moved is discovery and decision support.

He opens with a forced choice. Is AI in retail overhyped, underused, or misdirected? One word each, no preparation.

Kurklu says underused, and used wrong, which is where the missing returns live. Damian Zaplata of Empik Group answers from the retailer's chair: "super overhyped right now," because you spend more to stand still, though he concedes underused over the medium term. Piotrowski says misguided, and goes further than either: the technology is ready, the execution inside organisations is not, and "investing in activities that are checking if AI works doesn't make sense anymore."

The numbers Lipnicki puts up frame the next 45 minutes. In the US, 39% of consumers have used AI for online shopping and 85% of them say it improved the experience. AI-referred traffic to retail sites grew almost fivefold year on year in Q1 and converted 42% better. And Allegro still holds around 60% of Polish marketplace traffic. Discovery has moved; transactions largely have not.

Zaplata describes what that feels like from inside a retailer, and it is not a growth story. Discovery is shifting into chat, so there is a new channel to serve, new infrastructure and structured data to build, and at the same time a collapse in search traffic.

"You have a dramatic drop in SEO. Super dramatic. If you're not fast, you're basically a loser."

Is it incremental revenue? Probably not, he says. Same products, more competition, more transparency.

Piotrowski, who spent years on exactly this problem, sees more upside. Product discovery was the most painful part of the journey at Allegro, thousands of near-identical offers between the customer and a decision. Conversation compresses that, and it can surface alternatives a shopper never considered, which is the one place he expects genuine incrementality to come from.

Kurklu names the risk sitting underneath the opportunity. Every conversation a customer has inside an AI interface builds intelligence about that customer, and it accrues to whoever hosts the conversation.

"This intelligence stays with them, not yours. The cost of doing nothing and not building your own intelligence layer about those customers will be really hurtful."

Zaplata takes the same worry into strategy: hand over too much data and you wake up with a partner who owns the relationship and monetises it while you do not.

Then he moves the discussion off traffic entirely, which turns out to be the most quoted turn of the panel. "Focus on what you really control. You control the supply chain, you control the product." His example is procurement, not marketing: ordering Adidas ten months ahead, with no idea what the world will look like at delivery, versus fashion players who test demand online and reorder in days. Whoever gets that right, he argues, builds the next generation of retail leaders. He also marks the boundary an AI platform will not cross soon: Empik's subscription loyalty programme shows incrementality quickly, and ChatGPT is not going to launch a loyalty programme tomorrow.

Piotrowski's version is an accounting point. Technology becomes commodity; data and the customer relationship do not. The gap he sees is the willingness to treat AI as a general capability with a line in the P&L, rather than a token bill to be minimised.

On what retailers should not build, Kurklu is direct: do not build a vendor. He describes a customer certain that their device-level data mattered, until Replenit's lab tested it and found no effect on predicting the next purchase. "Are you buying all of the cars? Most of the cases, you are renting."

The handover question gets calmer answers than the framing suggests. Zaplata is comfortable letting AI spend a budget, with a limit, starting in the thousands and scaling step by step, because a named person still carries responsibility. Kurklu adds the compliance floor: machine decisions are recorded and auditable, with regulation arriving next year. Piotrowski holds the line on guardrails, and concedes the human exception, that some purchases people simply enjoy making themselves.

And then, unplanned, the morning produces its most repeated moment, and it is about cat food. Kurklu has two cats, 4.5 kilograms each, on the same food, ordered every month for three years from a billion-dollar retailer that still cannot anticipate the reorder.

"This is ridiculous from my point of view, and I don't want to accept this."

Zaplata matches it immediately. His monthly coffee delivery ships nothing at all when one selected item is out of stock, no substitute offered, and it happens every second month. Two anecdotes, one diagnosis: neither retailer lacked data, both lacked a system willing to make a decision with it.

What the room was asked to do on Monday

Lipnicki closes by asking for one action rather than a strategy. It is Friday, he points out. What changes on Monday morning?

Kurklu says stop postponing, because the customers who moved first are earning more, holding better margins and managing stock better, and the cost of waiting compounds. Zaplata says take five people out of their day jobs and put them in a room with nothing but this problem, because bolting "do a bit of AI" onto an existing job description does not work; he ran the experiment, and six months later that team came back with ideas nobody else in the company had. Piotrowski says pick one low-risk, time-consuming, uncomplicated task and delegate it end to end.

"And don't push the human-in-the-loop button."

The audience adds the last two beats. Someone asks whether this is really a headcount story, and Kurklu contrasts Klarna, which he says cut hundreds of call centre roles and then started rehiring because the handover was never designed, with IKEA, which kept its people alongside a design platform and grew revenue. Zaplata reports the same question arriving from his own board, in reverse: are we going to reduce IT costs now? "Actually not. But we have a lot more output and a lot of new ideas."

Someone from an ad tech firm gets the final question in, and it is the most grounding one of the morning: 80% of purchases still happen in physical stores, so what happens there? Kurklu's answer is honest about the state of the art. It is doable with unified customer profiles and in-store signals, an American retailer has asked about using blurred camera data to see which products a shopper touches, and it works, and it is expensive. Early days, deliberately.

The panel ran fifteen minutes over. Lipnicki sends the room into networking with an instruction that fits the morning: three meaningful conversations, and the day pays for itself.

What Warsaw arrived at, said out loud by people with nothing to sell each other, is that the technology is ready, the protocols are forming and the consumer has already moved. What is left is an organisational decision: whether anyone will let AI own an outcome and put their name next to it.

In a building where the hard question, twenty years ago, was whether to launch a website at all.