How an AI Compass run works
Describe the project
Twelve questions: what the AI should do, how demanding the tasks are, what one call must process, what must hold for the model, which operating mode, which data, which regulation, which AI Act class, where it is processed, which commitments, what scale, how important independence is. The compass does not ask for the context window or the model class – it derives them, the model class as a recommendation, not a filter. You can adjust the weights of the five dimensions.
A model scores against the catalogue
A language model reads, per provider, the documented AI catalogue, the data-protection and residency rows and the evidence validated against its source, and assigns five dimension scores. It sees only one provider's data at a time – no prior knowledge, no marketing.
The engine weights and ranks
Your answers produce weights and knock-out criteria. The engine accepts only exclusions you set, discards statements with numbers or names that are not in the data, and computes the fit transparently.
Result by email
As soon as the run is finished, the link to your result arrives by email: top 3 with fit, coverage of your use cases, strengths, limitations and regulatory notes, plus the sources.
You see the three best-fitting providers in full, all others together with the free expert check.
Methodology in short
Five dimensions, 1–100
Models & capabilities (does the catalogue cover your use cases?), operating modes & scale (serverless, batch, dedicated, fine-tuning, GPU), data protection & residency (training, retention, processing location, operator access), regulation & evidence (certificates in scope, support for AI Act duties), cost & independence (price indication, open weights, OpenAI-compatible API). The fit is the weighted mean.
Documented data only
Per provider the model sees exactly the rows from comparison.cloud: the category “AI & Machine Learning” (models, embeddings, inference modes, fine-tuning, service tiers, guardrails, AI data policy), the sovereignty rows on AI processing, residency, operator access and keys, the listed certificates and the validated evidence. No prior knowledge about the provider – what is not documented does not count.
Reasoning and context: from the model list
Every provider has a structured list of its LLMs: context window as served, status, processing location. A model registry adds vendor, reasoning capability, licence, official model card and the vendor-published benchmarks with source. The required context window is derived from the scope per call – pages, hours, files, lines of code instead of a token count. From this the engine decides deterministically whether a provider meets your reasoning and context requirement – e.g. gpt-oss-120b with 131k or Qwen3.8 with 262k tokens.
Model class: a recommendation, not a filter
Whether an open-weight model or a proprietary frontier model fits is not asked. The compass derives it from your model requirements – top quality, latest capabilities, vendor commitments, pinned version, inspectability, second source, unit cost – and from data class, commitments, exit expectations and regulation, and states the reasons. For classified information or operation in your own tenant an open model is necessary. The recommendation excludes no provider; all are always scored. No model class counts as safer or more compliant.
Knock-out only when you set it
Processing location, training exclusion, self-hosting, production readiness and European jurisdiction become exclusions only when your answers demand it. The model may propose an exclusion; the engine accepts it only when it is armed in your profile – and providers without a managed AI offering are always excluded.
Grounding instead of hallucination
Every model statement – strength, gap, regulatory note, cited service – must match the provider data: numbers, identifiers, model and product names must appear there verbatim, otherwise the statement is discarded. The result shows how many statements were discarded and which model did the scoring.
EU AI Act: support, not decide
The AI Compass asks for your risk class and assesses how well a platform carries the duties of that class – guardrails, logging, documentation, certificates. Whether a project is permissible and which duties apply is decided by your own classification. No provider is called “AI Act compliant”.
Frequently asked questions about the AI Compass
What distinguishes the AI Compass from the Cloud Compass and the Sovereign Compass?
The Cloud Compass assesses the workload fit — cost, security, availability, sovereignty and compliance — across the whole portfolio. The Sovereign Compass measures the evidence level of sovereignty on 23 criteria. The AI Compass looks at AI only: which models and services a provider documents, which operating modes exist, what it commits to on training, retention and processing location, which certificates cover that — and how all of this fits what your AI is supposed to do.
Who does the scoring — a human, a formula or an AI?
A language model scores each provider, but only on the documented data of comparison.cloud: the AI catalogue, the data-protection and residency rows, the listed certificates and the evidence validated against its source. It sees one provider at a time and no prior knowledge. The engine then weights by your profile, accepts only knock-out criteria you set, discards statements with numbers or names that are not in the data, and ranks. The result shows which model scored and how many statements were discarded.
Does the AI Compass consider reasoning models and context sizes?
Yes. One question asks how demanding the tasks are – routine, partly tricky, or multi-step reasoning as the core task; from this follows whether a reasoning model is a must. The context window is not asked as a token count but via the scope per call: short inputs, search hits, one long document or recording, many similar items, several documents, whole code projects, long agent runs or whole archives without retrieval. The largest selected input sets the tier (32k, 128k, 256k or 1M tokens). Every provider carries a structured list of its models with the context window it actually serves; a model registry adds reasoning capability, licence, official model card and the vendor-published benchmarks with source. From this the engine deterministically checks which models meet your requirement – for example gpt-oss-120b (131k) or Qwen3.8 27B (262k) at European providers; million-token context is offered by only a few models today, mostly via hyperscalers. A provider without a fitting model is knocked out unless you want to host models yourself.
Does the AI Compass recommend open or proprietary models?
It does not ask – it derives. From your model requirements – top quality, latest capabilities, contractual vendor commitments, pinned model version, inspectability, second source, low unit cost – and from data class, data-protection commitments, exit expectations and regulation, a reasoned recommendation emerges: open-weight model, proprietary frontier model or both classes. For classified information or operation in your own tenant an open model is necessary. The recommendation excludes no provider: all are always scored; a provider without a model of the recommended class loses at most a few points in model fit. No model class counts as safer or more compliant.
Does the AI Compass tell me whether my project is AI Act compliant?
No. The classification under the EU AI Act and the resulting duties are the deployer's responsibility. The AI Compass asks for your risk class and assesses how well a platform supports the duties of that class — guardrails, logging, documentation, certificates in scope. No provider is called “AI Act compliant”, and the fit is not a legal assessment.
Why is a provider knocked out?
Only because a criterion you set is not met according to the documented data: processing location (Germany, EU/EEA, European provider), training exclusion, self-hosting of own models, production readiness. Providers without a managed AI offering are always excluded. A knock-out does not mean “unsuitable for AI” — an individual commitment by the provider can change the picture.
How current is the catalogue data?
Model catalogues change monthly; the rows carry announcements, end-of-life dates and experimental states, which the assessment takes into account. The result freezes exactly the rows the model saw, together with the data version, so a report opened later still shows what you were told. Validated evidence additionally carries a check date.

