Ansvar AI docs
    ansvar.eu →

    Tools vs workflows

    The gateway exposes two shapes of interaction. Tools are single calls that return cited data; workflows are multi-stage assessments that produce a structured report. Both run through the same OAuth session and the same MCP fleet — the difference is whether you want a lookup or an analysis.

    Tools — single calls

    Tools are stateless: send a query, get a cited result. The agent picks one and calls it; the gateway routes it across the relevant MCPs in the fleet. Examples:

    • search(query, jurisdictions, sectors) — full-text search across whichever corpus matches the routing axes.
    • get_provision(jurisdiction, law, article) — pull a specific article at full fidelity (text, source URL, publisher, license). Also addressable by canonical_ref from an earlier search hit.
    • validate_citation(jurisdiction, law, article) — check a citation resolves against the served corpus and read its current text back for comparison.
    • list_coverage(jurisdiction) — what's available in the fleet for a given country; list_coverage(domain) answers "Using Ansvar: which jurisdictions for NIS2?"-style questions (e.g. domain="cybersecurity").

    Use tools when the answerable unit is one citation: "Using Ansvar: what does Article 32 of GDPR require?" "Using Ansvar: has Sweden transposed NIS2 yet?" "Using Ansvar: show me the case-law on selection criteria under Sweden's LOU" (reading case law is Premium and above; Solo shows that case law exists).

    Workflows — multi-stage analyses

    A workflow runs the same kind of lookups, but threaded through a scripted sequence of stages with a quality gate at each step. The agent calls four tools in turn:

    start_workflow(workflow_type: "dpia" | "gap_analysis" | "threat_model" | ...)
get_current_step(workflow_id)            // what to ask the user
submit_response(workflow_id, response)   // hand the user's answer back
generate_report(workflow_id)             // finalise once all stages pass

    Between start_workflow and generate_report, the agent drives a conversation: scoping questions, document collection, control or risk assessment, review, then the final report. Each stage's required fields are declared in YAML inside the workflow MCP — the gateway enforces them, so the agent can't skip a step.

    Free includes 1 workflow run a month and Solo 2 — seven types (threat model, gap analysis incl. NIS2, DORA, CRA and EU AI Act, DPIA) on a system you describe, watermarked render or JSON. Premium includes 5 across the full interview-grounded catalog; workflows grounded in your own documents — DPIA, gap analysis, tender review — are Team and Company. The same analysis can also be assembled by hand — search's automatic premium fan-out covers everything the workflow queries; what the workflow adds is the stage validation. The gap analysis guide walks both paths.

    When to reach for each

    • Tools — quick lookups, ad-hoc questions, embedding cited data into another conversation, programmatic integration where you control the flow.
    • Workflows (seven types from Free up — threat model, gap analysis and its NIS2, DORA, CRA and EU AI Act variants, DPIA; the rest of the interview-grounded catalog on Premium; document-grounded runs on Team and Company) — when you need a defensible report: DPIA, regulatory gap analysis, threat model, tender review. The output is structured so reviewers and auditors can rerun any stage independently.