Aetsoft attends DMEXCO 2026 in Cologne on 23 and 24 September. The subject on Aetsoft’s agenda is what AI already does inside loyalty, commerce and fintech products, and what that is worth to the business running them.
This year’s DMEXCO motto is . Most of the AI sessions in the programme share one question: what happens to a retailer, brand or payment provider when an AI agent discovers, compares and, soon, buys.
Three sessions answer it from different ends of the same system. Microsoft’s opening keynote puts it as a change in what brands compete on, from being found by a customer to being recommended by a model.
The CEO of idealo works backwards from the agent’s inputs. Without accurate product data, current prices and a clear reading of intent, it cannot buy reliably. The Transaction Horizon panel takes the question to the checkout, where agents weigh price, delivery time and availability, and loyalty-based payment starts to separate one retailer from another.
Aetsoft’s answer comes from systems that already run. In loyalty, commerce and fintech, AI ranks what a viewer or customer sees, turns a plain-language request into a configured loyalty offer and scores transactions for risk. In each case the model is only as useful as its surroundings.
When the customer is an agent, those surroundings are the product. The agent sees the data, the offer rules, the price and the payment path, and nothing else.
Each stage of digitalisation opens a different lever
Companies arrive at AI from different starting points. A retailer may still be joining customer data across its online shop, app, CRM, stores and payments. Another business already has a shared customer view and wants a better recommendation engine. A third needs to run hundreds of campaigns without more manual set-up.
The order matters. Personalisation stays limited while each channel holds only part of the customer record. A recommendation engine gets more useful once it can combine customer behaviour with product, pricing and inventory data. An AI assistant becomes useful once it can act through defined business rules and approved system interfaces.
Scale changes the economics. Response time, permissions, monitoring and the cost of each interaction start to matter as much as model quality once a system handles requests at volume.
METRO AG describes the same sequence on the Commerce Stage. The retailer replaced a 14-year-old legacy set-up and harmonised customer data across 20 markets. It now runs more than 60 automated campaign processes on that data. The campaign automation became possible because the data work came first, not alongside it.
“Companies digitalise at different speeds, and that is normal. Each stage opens a different economic lever, so no single solution fits every business.
What grows in value over time is the data a company has built around its own customers and products: preferences, behaviour, transactions, pricing and inventory. Any company can buy access to the same model. None of them can buy another company’s years of customer knowledge. That accumulated knowledge is the real differentiator. It lets a business build recommendations and AI workflows around how its own customers actually buy.”
Artem Kirylin, Commercial Director at Aetsoft
Recommendation engines turn connected data into product decisions
For a company that already has a shared customer view, a recommendation engine is one way to put that data to work. It ranks content, products or offers for one user in one context. How useful it is depends on the signals it can see, and on whether the item it picks can be shown, bought or delivered. Aetsoft’s guide to explains the data, model and product decisions behind this type of system.
Aetsoft has worked on this problem inside a live enterprise product. From 2019 to 2022, a dedicated Aetsoft team developed, tested and supported product changes for Kaltura’s Cloud TV platform, including work on its AI recommendation engine. Read the for the scope and delivery context.
The case also shows why recommendation work cannot be separated from the wider product. Each recommendation reaches the viewer through search, content rights, device support, playback and the rest of the customer experience. Improving the model changes nothing for the viewer if one of those steps drops the recommendation on the way.
AI assistants can become the operating interface
Recommendation engines help decide what may be relevant. A team that has to run hundreds of campaigns faces a different problem: how a person works with the system. That is where an AI assistant fits.
For a North American loyalty technology provider, Aetsoft built an AI assistant inside the business dashboard. A business owner describes a campaign in plain language. The assistant asks for whatever the request leaves out, pulls data from that owner’s loyalty programme and creates the offer through the platform’s existing APIs. Platform rules and permissions still decide what the assistant may access and what counts as a valid offer.
The shows how it was built. It covers the authenticated tools the assistant uses, the context it holds for each tenant and the way every action runs through the platform’s APIs.
Nothing in that pattern is specific to loyalty. An assistant can sit above an existing product or back-office workflow, but it still needs structured data, clear authority and controlled actions. Conversation is only the interface. Existing systems still decide what the assistant may do.
The loyalty assistant was built around exactly that: context held for each tenant, coordination through the platform’s own rules, and actions that run only through approved APIs. Two Tech Stage sessions arrive at them from the marketing side. pi-automate benchmarked autonomous agents on real ad-account tasks. Supplying the context they needed ate the time they saved. Twilio lists four traps for AI in customer experience, starting with agents that run without context or coordination.
The next step depends on what already exists
A company that has not yet chosen its use case needs to weigh the options before it builds anything. Aetsoft’s work covers the product, data and architecture decisions at that stage.
A team with a defined problem and no product yet needs a first working version to test it against. That is the job of Aetsoft’s : define the first version, build it and test it against real use.
Payment providers, brokers and other financial businesses have a further constraint, because AI has to fit the controls and integrations of the financial product around it. In fintech, Aetsoft builds AI for exactly that setting: AML monitoring, KYC workflows and risk scoring. Each comes with the permissions, audit trails and reporting that let a compliance team use the model’s output in daily operations.
Those same controls decide how far a payment provider can let an agent act for a customer. The Transaction Horizon panel puts that question to PayPal, eBay and OTTO.
Ecommerce Germany for that AML, KYC and risk-scoring work. It follows the same rule as the rest of Aetsoft’s . The team designs product logic, transaction workflows, data, providers and back-office operations together, so an AI feature has to live inside that structure.
Some established platforms need none of these. Their next step is better data access, orchestration or cost control. A regulated product may need permissions, review steps and records built into the workflow before an assistant can act.
Artem Kirylin is taking this argument to DMEXCO
, Commercial Director at Aetsoft, will represent the company in Cologne on 23 and 24 September 2026.
Artem has written up his working thesis: . At DMEXCO he wants to test that order against what teams in retail, e-commerce, loyalty, payments and fintech are working on now.
