AI platform: building blocks, selection criteria and the trade-off behind them

A chatbot subscription is access for one person. An AI platform is access for an organisation. That sounds like semantics and in practice decides three things: whether IT knows who has access, whether there is one contract instead of many, and whether the usage can be evidenced afterwards.

Technically, the difference is the one between an end product and a layer that administers what sits on either side of it. A chatbot wraps a model in an interface. A platform slots in between the organisation and the models and governs who may use which model with which data, where that data resides, and who receives the invoice. The term operating layer is also common for this intermediate layer; where it turns up in an offer, this is what it means. Which is why the model name is the least important selection criterion. It changes every few months, the operating model around it stays.

At the start comes user management. New joiners get access, leavers lose it, and in larger organisations that runs through the existing staff directory. Two standards govern it: signing in with the account someone already holds rather than with a new password (the technical term is SAML SSO), and creating and revoking access automatically out of the directory (SCIM). Langdock documents both, plus IP restrictions and session management.

Model access is where the products part company. Model-agnostic platforms serve several vendors through one interface; Langdock lists Google, OpenAI, Anthropic, Meta, Mistral and DeepSeek on its models page and puts the number of models on its API page at over 40 (langdock.com, retrieved 11 August 2026). An outside count by gewusst:KI on 4 August 2026 came to 35 listed models, seven of them with no fixed EU region, among them GPT-5 Pro, o3 Pro, DeepSeek v3.1 and Llama 4 Maverick. Vendor figure and outside count thus diverge, and for a selection decision the split matters more than the total anyway: a commitment to EU processing cannot be held for every model. Which ones are available in your own working area, which Langdock calls a workspace, is an administrator's decision.

Assistants are called "Agents" at Langdock: a saved role with instructions and attached files that you build once and then share with a team or the whole organisation. Prompt know-how held by individuals becomes a tool that works without those individuals.

Behind knowledge access sits a procedure that makes your own documents searchable: they are split up and stored, the passages matching a question are retrieved and handed to the model. The technical term for this is RAG. The appeal of a ready-made platform is that you do not have to build this yourself. The price is a lack of control over the details: Langdock's public documentation says nothing about how documents are split, which search model is used, or what the file size and document limits are.

Logging supplies the evidence. Without access logs and a usage and cost overview, every statement about your own AI usage is a guess, and that evidence is precisely the argument against shadow AI.

That leaves billing: one invoice, one data processing agreement, one notice period instead of twenty credit card subscriptions buried in expense reports. The data protection side of this bundling is covered in our article on GDPR-compliant AI.

A rollout is decided by the first task somebody actually completes with it. Three examples from organisations of this size.

In membership administration an email arrives asking about fees, deadlines or evidence. The caseworker gets a draft reply that quotes the relevant passage from the statutes and the fee schedule and supplies the reference, and approves it in the same place. Federations of this size report 30 to 60 such enquiries a week in our initial conversations. That figure is not measured; it is the order of magnitude we encounter there.

The project management of a funded project works from a grant decision with its ancillary provisions, usually 20 to 40 pages. Out of it comes a list of the obligations with deadline, form of evidence and owner, in a form you can keep working on. That comes up once per decision, so ten times a year with ten running projects.

The communications officer puts in a set of minutes or an annual report and gets back three versions of the same text, for the website, the newsletter and the noticeboard, in the organisation's agreed wording.

What all three examples have in common is that a particular document sits at the start and a particular form of result at the end. How much time that saves depends on the organisation, and no platform can promise it. Anyone who cannot name such a task needs an assessment first and a tool second.

The first question is not a technical one: what is the platform supposed to make provable? Faster draft texts need less than an organisation that must be able to explain to a supervisory authority which data went where.

After that, six points matter. The table gives the question to put to the vendor for each, and what marks an answer you can rely on.

The threshold decision-makers are really asking about separates two cost types. A price per seat, meaning per person with access, scales linearly with headcount: every additional person costs the same amount again. A model with a one-off setup and ongoing maintenance is independent of headcount. The break-even sits where the other side's fixed cost meets the seat cost over the same period, sensibly calculated across three years.

Nobody should put that point value in a proposal, because two items move it. On the seat side the rollout effort comes on top, since configuration, assistants, usage rules and training occur once; a fixed-price offer usually buries that item in the setup. It lowers the break-even. On the fixed-price side, infrastructure costs and the token costs billed by volume of text rise with actual usage, which pulls the other way. Your own figure therefore only emerges from two concrete quotes. How arocom shapes a rollout is on the service page introducing Langdock.

Below about ten people, price does not decide anyway. There the only thing that matters is whether the organisation has to be able to evidence its AI usage. If it does, the cheapest route to that evidence still runs through a platform, because a data processing agreement and a log cost less than the first data protection incident. There is no upper limit. From around a hundred seats the task does change: more departments with their own requirements, an IT lead at the table, and a staff directory that user management is wired into. The Bitkom survey of 21 October 2025 shows the same step in supply: 23 percent of companies with 20 to 99 employees provide AI access of their own, and from 500 employees upwards it is 43 percent.

A platform gets confused with workflow automation first. Langdock has workflows in the product, and the temptation to equate that with n8n is strong. The numbers say otherwise: 57 integrations with 754 actions at Langdock against 2,012 in the n8n integration directory, five trigger types, 18 of the 57 integrations with an event-based trigger and the rest available only as actions or through scheduled polling, a limit of 2,000 steps per run, and self-hosting only above the enterprise threshold, while the n8n Community Edition can be self-hosted free of charge indefinitely. The tools solve different problems. Langdock workflows attach AI processing to data already connected in your workspace. n8n connects line-of-business applications to each other, including ones on nobody's integration list. The details are in the article on n8n.

AI inside your own website or product is a different case altogether. A platform is a work tool for staff behind the login. A semantic search in a customer portal, an editorial assistant in the CMS or automatic data extraction from incoming documents belong in the application itself, with model APIs wired in directly, your own vector database and full control over data flow and cost. That is a development project. The two do not exclude each other, and many organisations do both, but the budgets and the people involved differ.