AI value or AI glitter: How to decide which ecommerce AI is worth investing in

A decade ago, every SaaS service was cloud-based. It was everywhere. These days, “AI” has taken the crown. Every service has features that use AI to varying degrees. Some fundamentally change how a business makes money. Others are just decades-old tech with a new fancy name. How do you tell the difference?

That question drove the closing panel of the online ecommerce conference RARE 2026, "AI in ecommerce: Which technologies are worth investing in for 2027." Three different perspectives on one stage: a consultant who sees every vendor pitch, a CTO who builds AI products, and a business leader who invests in technology in one of Europe's most regulated verticals.

The panelists:

• Jordie van Rijn, Email marketing specialist & MarTech consultant, founder of Email Vendor Selection
• Dmytro Mukha, CTO at Claspo, a conversion optimization engine
• Dmytro Shapovalov, Chief Business Architect at Liki24, a cross-border health marketplace

Dmytro Kudrenko, founder and CEO of Stripo, and Alex Danchenko, founder and COO at Yespo, moderated the discussion.

The moderators set one rule up front: disagree out loud and explain why. The panelists took it seriously. Below, we've grouped the debate into five decisions every ecommerce team faces when it budgets for AI, followed by a checklist to run before your next vendor demo.

Key takeaways

1. The AI label tells you nothing about a tool. Judge it by its business effect and by how reliably it repeats a task.
2. Keep the domain knowledge, data rules, and guardrails in-house. Rent the models.
3. Before you experiment with AI, make sure your basic offer works.
4. Speed pays off only once you know the direction. Otherwise, AI burns your budget faster.

Ignore the AI label and look underneath

Jordie opened with a contrarian take. Vendors now stamp "AI" on almost everything, including plain rules and marketing automation, and he thinks that's a good thing: the hype pushes people to read up on the field. Still, he was clear that the label won't help you choose a tool. Calling send-time optimization "AI" is questionable at best.

Dmytro Shapovalov agreed and went further:

Dmytro Shapovalov, Chief Business Architect at Liki24

 

"From the business side, I don't care what label a vendor puts on a tool, AI or automation. I only care about its business effect. It was exactly the same ten years ago with CRM automation."

Dmytro Mukha brought a historical lens. When factories first replaced steam engines with electric motors in existing plants, productivity barely moved. According to economic historian Paul David, US manufacturing productivity took off only in the 1920s, about four decades after electricity was commercialized, once factories were redesigned around the new power source.

Dmytro Mukha, CTO at Claspo

 

"When factories plugged electricity into their old plants, it didn't help. It took about 40 years to show up in the productivity statistics. We're at that point with AI: we've only just started to rebuild everything around it."

Dmytro Shapovalov pushed back on the idea of a grand rebuild. Nobody redesigns the factory in one go. You add small pieces, check that each one works, and move to the next. For him, the open question is reliability: whether AI agents can be trusted to run a full decision loop.

Maturity differs by department

The moderators framed AI adoption as three levels. The panelists' answers showed that one company can sit on all three at once.

Logistics at Liki24 stays at Level 1 because AI can't negotiate better rates with carriers or unstick a delayed shipment. Jordie added a deliberate Level 0: his team keeps AI out of software reviews entirely so the reviews don't read like everyone else's.

Decision 1: Who owns AI in your company

Dmytro Shapovalov splits ownership into two layers. The company owns the rules: which tools are allowed, how data is secured, and which decisions AI must never make. Liki24 works in a regulated market, so AI is never allowed to suggest treatment or offer prescription drugs over the counter. For him, the company owns the rules and departments own the business outcomes and decide how AI helps them get there. 

Jordie added a stage that comes first in less mature companies: a cross-functional task force. It owns the AI agenda, pushes adoption across teams, then hands the work over to departments and dissolves. He ran this setup with a data science group at a large home-improvement retailer. He also warned that rules only work if they match how people actually behave:

Jordie van Rijn, founder of Email Vendor Selection

 

"You can set rules, but they have to stay close to reality. People will tell you, 'Of course we only use the approved tool,' or 'No, we don't use ChatGPT.' Meanwhile, they're using it on their phones."

Dmytro Mukha drew the line between what to own and what to rent:

Dmytro Mukha, CTO at Claspo

 

"The most important part is the brain of the project: your domain knowledge, your expertise, your access to data. You should own that completely. For the models, the most productive approach is to rely on vendors like OpenAI or Anthropic."

At Claspo, that "brain" takes the form of skills built on top of other skills, chained into automated workflows. When two people run the same workflow, they get the same process, which makes the work easier to standardize and scale.

Use your own AI agent with a CDP

The Yespo plugin connects Claude and OpenAI Codex to your Yespo account through MCP. Describe a task in plain language, review the proposed actions, and let the agent work with your templates, contacts, workflows, and analytics. The agent acts only within the permissions of the user who connected it, and ready-made skills cover email editing, deliverability audits, and campaign analytics. 

Show me how to connect my AI agent →

Decision 2: Which data stays inside

Dmytro Shapovalov sorts data by sensitivity. In countries where Liki24 handles prescriptions, customers can upload them to the marketplace, and that data stays inside with no exceptions. Data used to optimize the business is different: funnel analysis or competitor prices that feed Liki24's pricing model can go to external tools without a problem.

He added a question that comes before any data decision: why do you need this model at all? For Liki24, the answer is always the fastest route to solving a specific business problem. Building a tool just because the technology exists doesn't qualify.

Alex Danchenko pushed the debate further. He described a talk by the head of a major marketing platform whose team gave a model sandboxed access to all internal APIs. They asked it to optimize onboarding or reactivation campaigns, let it choose which APIs to call, and even asked which data and APIs it was missing. Mukha admitted his heart rate went up just hearing it. In a multi-tenant platform, an unrestricted model could spot a conversion trick on one brand's website and reuse it for another.

Dmytro Mukha, CTO at Claspo

 

"Of course AI can access all the APIs. But we must build guardrails and anonymize the data first. Only then should the AI decide."

Jordie's rule of thumb: contain any data the task doesn't need. You don't know where the weak point is until you hit it. Personal data is regulated, and even metadata, such as who emailed whom and when, can be valuable in the wrong context. The chance of a leak may be small, but the consequences are huge.

Alex closed the loop from the vendor side. Customer data stored in a CDP belongs to the platform's clients, and Yespo doesn't use it across accounts, even though an AI would love to copy the best-performing campaign from one brand to another.

AI recommendations that follow your compliance rules. Liki24 runs Yespo AI recommendations in one of the most regulated ecommerce verticals. Prescription drugs and sensitive categories are excluded from recommendation blocks, and the Liki24 team added its own category relationships on top of the model. 

I want to see how Liki24 set it up →

Decision 3: Build or buy, and what to ask vendors

Jordie refused to answer with "it depends." If good tooling already does the job, don't reinvent the wheel. Build only when AI is the core of what your company does. If it's a side function, buy. The one exception he joked about is a CTO who loves building things, in which case it depends on how many shares the CTO owns.

Buying means vetting, and the panel named two questions vendors tend to dodge.

Security. According to Jordie, most AI vendors either avoid the painful security questions in a demo or simply answer "yes." If you can't check the answer, it's worth little. Sales prospecting, data enrichment, and scraping agents sit in the greyest area. Email and CRM tools are easier to vet. For enterprise purchases, specialized testing companies and security certifications tell you more than any selection committee can.

Reliability. Dmytro Shapovalov pointed to τ-bench, a benchmark from Sierra that tests AI agents on simulated retail customer-service tasks. GPT-4o completed about 61% of tasks on the first try. When it had to succeed at the same task in eight out of eight attempts, its score fell to about 25%. The benchmark dates from mid-2024 and models have improved since, but the gap between a one-off success and consistent performance is what a buyer should ask about.

Dmytro Shapovalov, Chief Business Architect at Liki24

 

"If I know an AI agent's reliability drops to 20–30% when it repeats the same task, the business question is simple: what kinds of tasks should I delegate to it?"

Jordie reframed it as a question of utility. An agent that successfully handles 25% of customer service tasks may still be worth it. In a medical setting, it isn't. He also warned that AI tools are SaaS products running on models that change. Whatever you test today can behave differently tomorrow, so testing has to be ongoing.

Decision 4: How you'll measure what AI changes

Dmytro Mukha came to the panel with a sobering number. A meta-analysis of more than 6,700 ecommerce experiments run on the Qubit platform found that 90% of them changed revenue by less than 1.2%, positive or negative. AI makes experiments cheaper to run, which doesn't make them more likely to matter.

His diagnosis: people love running experiments and love breaking the rules of statistics even more. They peek at results before a test finishes and stop the moment it first hits the confidence level. His fix is a strict system. Define the hypothesis and acceptance criteria up front, build small metrics that roll up to one North Star metric, and record what each experiment taught you.

Dmytro Mukha, CTO at Claspo

 

"For every metric, you need two criteria: uplift and harm, because it's always a trade-off. These rules should be deterministic, and humans or models should make decisions based only on them."

The moderators then put an audience question to the panel: with purchase history, website behavior, and email data in hand, which AI use cases should you test first?

• Dmytro Mukha believes AI systems can now design and run experiments themselves if they have rules, domain knowledge, and guardrails, with humans controlling what customers see. Claspo is building a team of agents for this. In one client project, the agents found that half of the website's traffic was landing on a sold-out product page. No one on the team had noticed.
• Jordie recommends testing "close to the fire," meaning as close to your main revenue KPI as possible: next best product, next best offer, cross-sell and upsell, and the welcome flow.
• Dmytro Shapovalov disagreed with both. Cheap experiments create more noise and fewer results. Five years ago, every test cost real resources, so teams thought harder about which ones to run.

Dmytro Shapovalov, Chief Business Architect at Liki24

 

"Before you jump into AI experiments, check your basic offer and your product-market fit. If something is wrong there, fix it first. Only then move on to marketing, experiments, and the nice tools."

His example: Liki24's AI personalization in marketing campaigns, built with Yespo, showed great results. A plain experiment with the free delivery threshold, run in parallel, also delivered a strong business impact.

"Too boring!" Jordie joked, but ultimately agreed. Fundamentals come first, but a team also needs some "learning money" for exploratory projects, because experimentation works like a pinball machine: you try something in one place and learn something in another. Dmytro Kudrenko added a warning about metrics that look good and earn nothing. One recommendation engine he'd seen sold piles of low-margin batteries while another sold mostly socks. The numbers went up, and the business barely felt it.

Continuous A/B testing with a human in the loop.

The One from Many block tests a pool of message variants inside a Yespo workflow. It starts by sending variants at random, then shifts toward the ones with higher CTR and automatically excludes variants that keep losing. With Auto-optimize variants enabled, AI drafts new variants for your team to review before they join the pool. 

Show me how One from Many works →

Decision 5: Speed or performance

Dmytro Kudrenko asked the panel to pick one of three ways ecommerce teams adopt AI: producing faster, doing more (for example, personalization at scale), or improving a performance indicator. "It depends" was banned as an answer.

Jordie picked speed. If your team spends a lot of time producing marketing materials and campaigns, cut that time first. It gives your current people breathing room, and it's usually the easiest thing to implement.

Dmytro Shapovalov picked performance and brought a cautionary tale. In one project, a team used generative AI for Meta ad creatives. Before, they could produce 5 to 10 creatives. With AI, they produced hundreds and started testing them. Then they ran the numbers: customer acquisition cost was too high, and customer lifetime value couldn't cover it. Two or three experiments would have been enough to see that.

Dmytro Shapovalov, Chief Business Architect at Liki24

 

"If you speed up in the wrong direction, you'll just lose your money. You don't need speed if it leads you the wrong way."

Dmytro Kudrenko called it a chicken-and-egg problem: speed frees up time to work on strategy, but you need a strategy to know what's worth speeding up.

Dmytro Mukha sided with performance too, with one addition. Invest in your own skills and toolchain, whichever vendor you use, because the environment changes too fast to wait. And plan for scale: AI is no longer just an assistant. Agents can now run in loops and iterate over whole workflows.

AI investment checklist

Jordie's closing advice sets up the list. Teams often rush into buying a CRM, an ESP, or another big system without separating what they need from what would be nice to have, and only find out afterwards that it was the wrong fit.

Jordie van Rijn, founder of Email Vendor Selection

 

"Get to know the core of the tools you use and will use. Give them a little scratch with your fingernail and find out what is AI glitter, the stuff around the core that you don't really need."

Before you sign off on your next AI investment, ask yourself:

1. What business effect will this tool have, apart from its AI label? If the vendor can't name the metric it moves, treat the label as marketing.
2. Who owns the rules, and who owns the results? The company sets policies on tools, security, and decisions AI may never make. Departments own outcomes. Make sure the rules match how people actually work, or you'll get shadow AI.
3. What data will the tool touch, and does it need all of it? Keep sensitive data inside, anonymize what goes out, and contain anything the task doesn't require.
4. Is AI the core of your business? If not, buy rather than build. Either way, own your domain knowledge and rent the model.
5. Can the vendor back up its security and reliability claims? Ask for certifications and independent testing. Ask whether the agent can complete the same task consistently across many attempts.
6. Have you defined uplift and harm for each metric before the test starts? Set acceptance criteria up front, don't stop a test the moment it first looks significant, and check margin alongside volume.
7. Have you tested your basic offer? Pricing, delivery thresholds, and product-market fit can move the business more than any AI feature.
8. Do you know the direction before you speed up? Run the unit economics on a few experiments before AI multiplies them by a hundred.

The panelists didn’t agree on everything, but the common ground was obvious. Jordie van Rijn put it best: strategy is now more about what you choose not to do. AI gives a creative marketer something new to build every morning, so choosing is the hard part. He also expects more marketers to open GitHub accounts, write some Python, and ship "ugly software that works."

Alex Danchenko closed the conference with a prediction. AI will lift novice marketers and weak campaigns to a new normal that sits far above today's. What gets you above that new normal is a human understanding of why customers come back.

Recordings of all sessions are available on the RARE 2026 website. 

Special Request Inline