Source analysis
Inspect files, schemas, samples, and upstream documentation to surface fields, formats, gaps, and assumptions.
AI proposes. Approved controls decide. Deterministic workflows execute.
Fontana agents assist with analysis, drafting, explanation, and setup inside a governed operating layer. They do not execute production workflows without approved controls. AI proposes, people approve, and approved workflows execute deterministically, with policy, logging, lineage, and replay built in.
LLM Catalogue
300+ models via 12+ approved gateways
Fontana is AI agnostic: route any LLM through any approved gateway that fits your compliance, residency, and data-handling requirements, without locking workflows to one vendor or model family.
Agents reduce analysis and setup effort before controlled workflows run.
The agents interpret complex inputs, draft operational logic, explain breaks, and assemble evidence, before approved logic is converted into deterministic workflows.
Mapping drafts
Draft candidate mappings and transformations for review against approved target schemas and controls.
Exception explanation
Summarise likely causes, supporting evidence, prior similar breaks, and proposed next steps for human review.
Operating specifications
Turn source analysis, rules, decisions, and approval requirements into review-ready operating specs.
Evidence packs
Draft lineage, assumptions, reviewer notes, and run evidence so teams do not assemble packs manually.
Workflow setup
Help configure a governed workflow faster while keeping approvals, tests, and deterministic execution explicit.
AI keeps production authority via human-in-the-loop approval.
Useful AI should make work easier to review, not harder to govern. Production authority stays with approved controls and named owners. AI tasks are separated by risk: analysis and drafting run automatically; production-impacting actions stay behind human-in-the-loop approval. Every AI action - automatic or human-approved - is included in the full audit trail.
Tasks that AI agents can perform automatically:
Analyse files, schemas, and source outputs without approval
Infer fields, mappings, and validation checks for review
Draft operating specifications and mapping proposals
Explain breaks, exceptions, and supporting evidence
Propose workflow configuration before it is applied
Summarise evidence packs for reviewers
Tasks that require human-in-the-loop approval:
Change production rules or workflow boundaries
Approve exceptions, outputs, or control changes
Execute operational steps in production
Bypass prompt, output, routing, version, and decision logs
Override named owners, approval gates, or control policies
Act outside defined permissions and data-class boundaries
AI assistance can be disabled without disabling the workflow.
Fontana's execution boundary remains deterministic. If a workflow, data class, client, jurisdiction, or policy requires no AI assistance, the approved rules, validations, approvals, lineage, replay, and audit evidence continue to operate without model involvement.
AI path
Assistive analysis and drafting where policy allows.
Approval gate
Human-owned control decides what is accepted.
No-AI path
Rules execute deterministically with the same evidence requirements.
Prompts, outputs, routes, and reviewer decisions are logged.
Technology and control teams need to know what was asked, which data was referenced, which model or agent was used, what came back, and what humans decided afterwards.
Prompt and task intent
Input references and data class
Model, agent, route, and policy applied
Output, citations, and confidence signals
Reviewer, decision, comments, and approval state
Run context needed for audit and replay
Routing is governed by task, data, policy, cost, confidence, and approval requirement.
Fontana is model and agent agnostic, but not model indifferent. Routing choices are part of the governed operating model, not hidden implementation detail.
Task
Different work can route to different models or agents depending on purpose and risk.
Data class
Sensitive data can be restricted to approved models, environments, or no-AI paths.
Policy
Governance rules define which assistance is allowed and which actions require approval.
Cost and usage
Model usage is visible so teams can monitor spend and tune routing over time.
Confidence
Low-confidence outputs can be blocked, routed to review, or handled without AI assistance.
Approval need
Tasks that affect production remain behind human-owned gates and deterministic execution.
Operations stay in control. Technology can audit and replay activity.
Fontana gives operations teams full control with complete visibility over workflow activity, approvals, changes, and outcomes. Every process can be audited, traced, and replayed when needed.
Operations teams
Own approvals, exception handling, rule changes, and production decisions. AI helps prepare the work; operators decide what becomes accepted process.
Technology teams
Audit prompts, outputs, routes, model usage, versions, permissions, lineage, and replay context without treating the AI layer as a black box.
Fontana AI - Feature Highlights
Choose from 300+ public models, or use your firm's custom LLM.
Route frontier, open, and domain-specific models through OpenRouter, direct APIs, Bedrock, and OpenAI-compatible gateways, or connect a private deployment your architecture team hosts inside the perimeter.
Connect gateways with your firm's own API keys and credentials, stored separately from model routing and never passed through the thread.
Fine-grained configuration of all agents, from system prompts and reasoning level to context pool, temperature, cost guards, and more.
300+ Models
Frontier, open, and domain-specific models from major providers
12+ Gateways
OpenRouter, direct APIs, Bedrock, and OpenAI-compatible routes
Fontana Platform
Governed routing, policy, logging, and workflow boundaries
Orchestrate long-running workflows with explicit agent boundaries.
Create agents for analysis, workflow setup, and review. Use skills, sub-agent delegation, and handoff to keep threads focused and token use efficient.
Use SubAgent Delegation to isolate deep work in child threads while the parent keeps context lean.
Use Agent Handoff to transfer the root thread to a specialist agent when the task surface changes.
Use Agent Interop with A2A, ACP, OFP, and other standards to communicate with external agents.
Build out your firm's Knowledge Graph in permissioned namespaces.
Organise operational knowledge (standards, firm rules, mappings, prior decisions, and workflow logic) into namespaces your teams control. Agents assemble thread context from what you have validated, not ad-hoc uploads.
Build a knowledge graph to model your firm's standards, rules, and decisions.
Use Agent Skills to add tools and context briefly for specific tasks without bloating every turn.
Allow agents to self-improve using auditable, permanent memory that persists across threads.
Enrich your AI threads with accurate live data queries, directly from your data sources.
Allow AI to run complex queries on your data to search, group, filter, aggregate, and more, without pushing bulk datasets through the model.
Agents can query data directly using SQL, TypeScript, Python, and more; results are returned to the thread without bulk data entering the model context.
Use MCP to access external data sources, perform financial data lookups, or communicate directly with your firm's tooling.
Build analysis reports with graph citations and live query proof.
Combine Knowledge Graph context with live queries to produce findings reviewers can trust. Attach query definitions, cite approved sources, and export review-ready evidence, not unsupported narrative.
Render charts and visualisations inline in the thread so reviewers see findings alongside the query proof that produced them.
Generate Markdown, Excel, Word, and PDF documents, complete with citations, and export-ready structure on the audit trail.
Embed questionnaires, clarifications, and approvals in the flow.
Generate structured questionnaires, clarification checks, and approval forms mid-analysis or mid-workflow build. Collect human input as steps in the thread, not side-channel email or ad-hoc chat, so decisions stay on the audit trail.
Surface workflow suggestions in-thread so reviewers can accept or reject proposed changes without leaving the audit trail.
Collect questionnaires, clarifications, and approvals as structured forms embedded directly in the agent thread.