When Users Start Building the Interface Themselves
Students already use general-purpose AI to help them learn. Restaurant diners are beginning to use it to interpret menus and wine lists. In both cases, users are building an informal interface around services that were never designed for conversational intelligence. What is missing may not be more AI, but better orchestration.
A recent Ravenous article on AI, menus, wine lists and restaurant ordering describes a behaviour that is becoming increasingly natural: diners photograph a menu or wine list and ask a general-purpose AI to help them understand it, compare choices, find value or identify something that matches their preferences.
At first sight, this is an encouraging signal for one of the verticals currently being developed within Ministerium.
But the more interesting point may be that we are observing essentially the same phenomenon in another, completely different field.
The service already exists — informally
Students did not wait for educational platforms to introduce AI.
They already use ChatGPT and other general-purpose assistants to ask questions, request explanations, solve exercises, translate concepts and explore subjects.
Restaurant customers may now be starting to do something similar. Instead of navigating a complex menu or wine list themselves, they ask an AI to interpret it for them.
In both cases, users are effectively constructing their own interface around an existing service.
And in both cases, something important is missing.
From general intelligence to service orchestration
A general-purpose AI can explain mathematics.
But it does not automatically know what a particular student already understands, where misconceptions are forming, what should be practised next, when an explanation is appropriate, or when giving the answer would actually undermine learning.
That is the role of the Ministerium tutoring vertical.
For its first launch in Switzerland, we currently call it Swiss Tuto. The generic working designation is Socratic Tutoric — conveniently preserving the same initials, ST. Its definitive product name will come later.
ST is not intended to be another chat window around an LLM. It is intended to orchestrate tutoring: maintaining learning context, identifying gaps, adapting explanations and exercises, deciding when to explain and when to question, following progress over time, and selecting the most appropriate conversational or structured interface for the task.
The benefits are straightforward: greater continuity than isolated AI conversations, more individualized pacing, earlier detection of misunderstanding, preservation of the cognitive effort required for learning, and a learning path that can evolve rather than restart with every prompt.
A tutor does not merely know things.
A tutor knows how, when and why to use that knowledge with a particular learner.
Catering exposes the same missing layer
A general-purpose AI can analyse a photograph of a menu and recommend a dish or a bottle of wine.
But it may not know whether that item is still available, its current price, the restaurant's actual options, allergens, preparation rules, combinations, service constraints or what can really be ordered at that moment.
And after making a recommendation, it still does not necessarily turn the decision into an authoritative transaction.
This is where our second vertical begins.
It currently carries the development codename Ctwo, or C2. Like ST, it already has a prospective definitive name, but we prefer to keep that private until the project is further validated.
C2 combines natural interaction with the catering provider's own authoritative menu, prices, availability and operating rules. A customer might simply say:
Something light, vegetarian, without mushrooms, around CHF 30.
The system can interpret the intent, ask only what still needs clarification, present appropriate choices and move naturally from conversation to structured selection, confirmation, ordering and payment.
The conversation is the doorway, not the constraint.
Two verticals, one underlying problem
Socratic tutoring and catering self-ordering could hardly look more different.
Yet the same structural problem appears in both.
General-purpose AI provides intelligence, but intelligence alone does not constitute a service.
A useful service also requires context, continuity, authoritative data, rules, state, permissions, appropriate interfaces and reliable execution.
That orchestration layer is one of the central ideas behind Ministerium.
Perhaps it is therefore not entirely accidental that ST and C2 are being developed concurrently. They allow us to test the same underlying service hypothesis in two environments with almost nothing else in common.
If the principle works in both, the interesting result will not simply be that we have built a tutoring service and a catering service.
It will suggest that a more general service architecture can sit between human intention, increasingly capable AI and the authoritative systems required to make something actually happen.
Sophisticated for the user, almost trivial for the operator
C2 also tests another part of this hypothesis.
The sophistication experienced by the customer should not translate into technological complexity for the service provider.
Our objective is that virtually any café, bar, restaurant, temporary venue or other catering operator can configure a functioning conversational self-ordering service in just a few minutes, using equipment already available and without requiring IT expertise.
No dedicated kiosk infrastructure should be necessary merely to discover whether self-ordering creates value in a particular environment.
That combination matters: natural interaction for the customer, operational efficiency for the provider, and as little technology as possible to deploy and manage.
The behaviour described by Ravenous therefore interests us for more than its connection with restaurant ordering.
It may be another early indication of a broader transition already visible in education and elsewhere:
users increasingly expect to express what they want naturally; the service layer must increasingly take responsibility for understanding how to deliver it.
ST and C2 are two very different experiments in that direction.
And C2 still has another dimension we have deliberately said very little about so far: its business model.
We will come back to that shortly. It may prove even more attractive than the interface itself.