A case without magical conversion claims

Liechtenstein's GB Marketing & Solutions GmbH and its partners launched the expert marketplace expert-hub.li. According to the European Digital Innovation Hubs (EDIH) network, it offers expert profiles and service packages, while AI helps match a customer's request with relevant expertise. The marketplace is publicly available: a user compares offers and describes the need during booking, after which the adviser decides whether the request fits their area of expertise.

This is more useful to a small business than a vague promise of “AI in sales.” The narrow job is to shorten the path from an unstructured request to a suitable provider without routing someone to an expert merely because one word in a profile matches. The site also shows appointment booking and service packages. EDIH lists chat, document exchange, video meetings and invoicing among the platform's functions. It does not, however, publish verified figures for increased sales or matching accuracy: metrics were promised after the first months of operation, but are absent from the available description.

The comic failure mode will be familiar to anyone who has sought a contractor: ask for one expert and receive a stack of forty profiles. A useful AI is therefore not the one that produces the longest ranking, but the one that helps ask a clarifying question and leaves two or three explainable options. A human adviser retains the right to decline a request, while the customer can choose and clarify terms.

What the sources actually confirm

EDIH describes a marketplace for finding experts and processing consultations, adapted to the local market. The technical operation is handled by Austria's HalloSophia by finothek GmbH; GB Marketing & Solutions develops the business model and marketing. The live expert-hub.li page confirms that offers and booking are available, and that an adviser accepts a request only after assessing the fit with their expertise. The page separately warns users not to share personal information in the AI chat.

The model used, ranking method, conversation-storage architecture, automatic-match rate, financial effect and comparison with manual routing have not been published. In particular, there is no basis to call the live platform locally deployed or to claim that it uses RAG. These limits matter: a small company should not copy an unknown technical stack just because the interface looks convincing.

How to build a similar process

Imagine a Russian service company or trade association receiving repair, audit, training or engineering-consultation requests by email and through a website. The first layer does not require a model. It needs structured provider records: verified skills, industries, service area, language, minimum price, availability, required qualifications and tasks the provider will not accept. Every field needs an owner and a review date. Otherwise, AI will confidently recommend someone who no longer offers the service.

Next, normalize each request into a limited set of attributes: task type, object, urgency, location, budget range and missing information. Rules first exclude candidates on hard constraints such as service area, license, availability and conflicts of interest. Only the remainder should be ranked using semantic search or a small model. For rare or ambiguous requests, the model can suggest one clarifying question. It is better to leave an unknown item unknown than replace it with a polished guess.

A local model is appropriate when requests contain confidential technical details, volume is steady and a secure infrastructure already exists. But local inference is not mandatory for a catalog of a few dozen experts: filters, full-text search and human confirmation may be cheaper and more dependable. If an external API is used, specify which fields and documents may be sent to it, and avoid personal data unless necessary.

Integrations and control

The minimum operational loop is an intake form, provider catalog, availability calendar, CRM request status and recommendation log. The model should not independently create commitments to customers. The user can see why a candidate was suggested; the provider accepts or declines; a manager reviews disputed cases. If a service is regulated or requires a qualification, qualification checks remain a separate deterministic step.

For a pilot, record the profile revision and reason for every suggestion: “experience with this task,” “available in this region,” or “fits the budget.” Do not present internal scores as an objective assessment of a person. Ranking depends on profile quality and request wording; an expert with an incomplete profile might otherwise become invisible. Catalog owners should therefore review profiles, duplicates and requests for which no suitable provider was found.

The public marketplace explicitly warns users not to enter personal information in the AI chat. For a Russian deployment, this is a good reason to separate free text used for discovery from contact details used for booking. Contact information and contract files should be available only to the appropriate staff; request logs need retention limits; exports to outside services must be checked separately. This is a practical access design, not a claim of compliance with any specific law.

Economics: count completed engagements, not “smart” answers

Because EDIH provides no verified platform KPIs, another business should begin with its own baseline. For two weeks, record incoming requests, time to the first suitable offer, the share accepted by advisers, reassignment rates, completed consultations and complaints. Compare against the current manual route for comparable tasks. When volume is low, operator time saved may not pay for maintaining a new service; convenience and availability may still have value, but they too need measurement.

The following is an illustrative calculation, not an expert-hub.li result: at 80 requests a month and four minutes saved on initial triage, gross savings equal 320 minutes, or five hours and 20 minutes. Subtract profile moderation, error handling, integration maintenance and compute costs. If wrong assignments increase, the net effect can be negative even when the first response arrives faster. Buying a dedicated GPU solely for matching at this volume is unlikely to be worthwhile unless it supports other workloads too.

A manager's next step

Start with one service category and 20–30 current profiles. Take 50–100 anonymized past requests and mark acceptable providers as well as cases requiring a clarifying question or no recommendation. Compare three modes: manual routing, rules plus search, and rules plus AI ranking. Show a recommendation on the site only after checking permissions, current status and an explanation of the match. If AI merely makes the list longer, it is not solving the problem yet.

The illustration was created with AI for VnutrII; it does not depict expert-hub.li staff or its interface.