Direct answers to the questions enterprise procurement, risk and operations teams ask before deployment.
How is a hierarchical agent team different from a chatbot fleet?
A chatbot fleet is a collection of independent bots with no shared accountability, no role assignment and no audit trail. A hierarchical agent team has named roles (manager, specialist, executor, auditor), explicit reporting lines, configurable human approval gates and immutable decision logs. The difference is the same as the difference between a crowd and an organisation.
Does this satisfy EU AI Act Article 14 on human oversight?
Yes. Article 14 requires that high-risk AI systems can be effectively overseen by natural persons during the period in which they are in use. The hierarchical model provides this through configurable approval gates at every hierarchy level. The approval matrix is logged, reviewable and adjustable. The Article 14 documentation is produced automatically.
How are agent decisions logged?
Every decision is logged to an immutable, EU-hosted object store. Each log entry contains timestamp, agent identity, role, inputs, full reasoning chain, model used, outputs, downstream actions taken and any human approval. Logs are linked into a hash chain so tampering is detectable. Retention is configurable from 12 to 60 months. This satisfies EU AI Act Article 12.
Where do the agents actually run?
On EU-hosted infrastructure only. Orchestration runs on AWS Frankfurt or OVHcloud, models execute on EU endpoints (Mistral Paris, self-hosted Llama on Hetzner GPU, Claude or GPT via EU endpoints), and the audit log lives on Scaleway or EU object storage. Nothing leaves the European Economic Area.
Which models do you use?
We select models per agent role based on cost, accuracy and EU hosting availability. Typical mix: Mistral Large 2 for general reasoning, Llama 3.1 70B for high-volume specialist tasks (self-hosted on Hetzner GPU), Claude 3.5 Sonnet via Anthropic EU for nuanced writing, GPT-4o via Azure Sweden Central for vision and complex coordination. The model assignment is configurable per agent.
How long does deployment take?
Six weeks from kick-off to live production. Week 1: process mapping and role design. Weeks 2 to 3: agent build and tool integration. Week 4: oversight wiring and audit instrumentation. Weeks 5 to 6: sandbox pilot, calibration and go-live with a 30-day hyper-care window.
What happens when an agent makes a wrong decision?
Three layers catch it: the manager agent reviews specialist outputs and routes uncertain cases through a human approval gate; the auditor agent continuously samples production outputs and flags anomalies; the immutable decision log enables post-incident review with full reasoning chain. Corrections are made by adjusting the prompt or escalation threshold, the change is sandbox-tested, and the corrected version is promoted to production.
Can the agents integrate with our existing tools?
Yes. Standard integrations exist for HubSpot, Salesforce, Pipedrive, Zendesk, Intercom, Notion, Slack, Microsoft 365, Google Workspace, Stripe and most major SaaS providers. Custom integrations are built via REST API, webhook or direct database connection. Agent tool access is permissioned per role.
How is pricing structured?
A fixed deployment fee covering the six-week build, plus a monthly operational retainer that scales with the agent count and workload. The retainer covers infrastructure, model usage, monitoring, compliance documentation updates and quarterly reviews. Contact us for a tailored quote.