The governance layer for the agentic enterprise.
Every model read and every agent step is checked before it commits, and the proof writes itself. Built for finance. Three planes: know what can act, govern the action, prove it held.
3planes
One governance lifecycle
11surfaces
Each a subproduct
1estate
Data, AI, agents
0blind spots
Nothing acts unseen
The problem, in three parts
AI scaled faster than anyone could govern it. Agents made it autonomous.
01
You cannot see the estate.
No single inventory of the data, models, and agents that can act. You cannot govern what you cannot see.
02
Controls do not reach the action.
Policy lives in binders and pipelines. Agents act outside both, in milliseconds.
03
Findings arrive as surprises.
Evidence is assembled after the fact. The exam finds the problem before you do.
One layer closes all three ↓
Where Fiducia sits
One layer between the estate and everyone it answers to.
Actions flow in from the systems Fiducia reads. Evidence flows out to the people the estate must answer. Nothing commits in between without a check.
Reads from
- Warehouses and lakes
- Model and feature registries
- Agent frameworks and copilots
- SaaS and core systems
- Identity and ticketing
- Third party AI and vendors
The governance layer
Fiducia
The Runtime
Checked before it commits
Answers to
- The board
- The regulator
- Internal and external audit
- Model risk and compliance
- Customers and members
- The public record
Arcade governs the tool call. Redpanda governs the data plane. Fiducia governs the decision, and maps it to the rule it must answer to.
The stance
The controls that made you examinable will not make you governable.
Control has moved through four eras. The first three all assumed a person in the path with time to intervene. Autonomy removes that time, and with it the one assumption every old control was built on.
A control that is not enforced at the moment of action is not a control. It is a record.
The consequence gap · time to intervene
Closing
Every control built for the first three eras assumed the bar was wide enough for a person to step in. At era IV there is no bar.
I
Ledger
Control by record. The books reconciled after the fact.
II
Periodic review
The model risk world. Validation on a calendar.
III
Continuous monitoring
Observability. Watching in real time, still after the act.
IV
Autonomous execution
Agents act and the person is gone. Control must move into the action.
Know what can act. Govern the action. Prove it held.
Governing autonomy is not one feature. It is a stack. Three planes, read left to right. Eleven surfaces sit across them, and every surface keys off the one before it.
Plane 01 · 3 surfaces
Provenance
Know what can act
Plane 02 · 4 surfaces
Runtime
Govern the action
Plane 03 · 4 surfaces
Assurance
Prove it held
The governed estate
Govern one decision and it is a demo. Govern every decision and it is the product.
Fiducia scores the whole estate continuously. Every row is a model or an agent. Every column is a check. Every cell is current.
The governed estate · live
87Assurance Score · rising
| Sens | Qual | Line | Bias | Auth | Evid | |
|---|---|---|---|---|---|---|
| underwriting_v4 | ||||||
| kyc_agent | ||||||
| rebalance_agent | ||||||
| claims_copilot |
1 red · kyc_agent authority out of bounds · blocked, routed to owner
Stochastic capability. Deterministic mandate. Both, at once.
Regulators require reproducibility. Fiducia resolves the tension between stochastic models and deterministic mandates by wrapping every governed action in machinery that behaves the same way every time, and writes down what it did.
01
Structured prompt templates
Versioned, audit tracked templates. The same inputs produce the same evaluation pathway, every time.
02
Multi agent consensus
Multiple agents evaluate the same decision independently. Disagreement above threshold routes to a human reviewer.
03
LLM as judge
A separate model instance scores primary outputs against calibrated rubrics. Generator and evaluator are architecturally separated.
04
Compliance as code
Controls expressed as versioned, testable code. A policy change ships like a software release: reviewed, tested, rolled back if it fails.
05
Tiered decision rights
Reversible actions clear at machine speed. Consequential actions route to a named human before commit. The line is codified.
06
Immutable audit log
An append only record of every reasoning step. Not editable, not deletable.
Built for the institution, not the startup.
Tenant isolation
Single tenant deployment options. Your data does not train foundation models.
Data residency
Deployed in your region, your cloud, or your own perimeter.
Certifications
SOC 2 Type II. Zero data retention options. DPA ready.
Enterprise identity
SSO, SCIM, and role based access mapped to your control functions.
Read only by default
Connectors observe. Enforcement scopes are explicit, granted, and logged.
Examiner ready records
Every governed action leaves an immutable, attributable record built for supervisory review.
Coverage
Every rule you answer to, already mapped to a control.
From SR 11-7 to the EU AI Act, from NYC Local Law 144 to ISO 42001. One control test answers every framework that asks for it.
Explore the Rulebook · 29 frameworks and countingMap once, prove everywhere
29 in the Rulebook
One test result answers every framework that asks for the same thing. Coverage stops being a per framework project.
Why Fiducia, not the alternatives
Everyone owns a slice. Fiducia owns the seam.
One estate, not three tools
Data, AI, and agents governed as one. Everyone else owns a slice. Fiducia owns the seam.
Enforced at the action
Policy decides before the action commits, not in a review that happens after.
The language of the exam
Written in the regulatory vocabulary finance is examined in, from SR 11-7 to DORA.
Evidence that writes itself
Living Evidence as a byproduct of execution, not a binder you assemble each quarter.
On the roadmap · future tense
Govern what defends. Consume what proves.
Agentic security is coming to Fiducia. Oversight of the agents that run your security operations, and proof from offensive testing turned into evidence the board can read. One estate, one record. This stays future tense until it ships.
Questions, answered.
What is Fiducia?
Fiducia is the governance layer for the agentic enterprise. Its production offering, Assurance, is an AI-native IT risk management and governance platform built for the second and third lines of defense at regulated financial institutions. It ingests your frameworks, policies, and live operational data, maps controls to risks continuously, tests controls across the full population, and generates audit-ready evidence, replacing manual spreadsheets and point-in-time sampling.
Do we have to replace our GRC platform to use Fiducia?
No. Fiducia reads from your GRC platform, ITSM, CMDB, and the rest of the estate through read-only connectors, and produces independent challenge evidence alongside them. Institutions typically run both, with Fiducia as the evidence and assurance layer, and revisit the system-of-record question at their own pace.
How is this different from an AI security tool that blocks agent actions?
A security tool blocks on anomaly, when behavior looks wrong. Fiducia evaluates against institutional authority, so an outcome cites the rule, the control, the appetite clause, and the obligation behind it. Both are useful and neither replaces the other, and security telemetry can be a source into the governance graph. Assurance evaluates and evidences actions today; enforcing at the moment of action is the Runtime plane, which is in development.
Does Fiducia write anything back into our systems?
No. Connections are read-only, and write-capable credentials are rejected at setup. Where an outbound integration creates a ticket or a message in another system, that action runs under credentials you supply for that integration and is subject to that system's permissions.
Is our data used to train models?
No, in any deployment model. The platform documentation sets out the full position, including the terms that apply to model serving.
Can the AI close a finding or change a risk rating on its own?
No. Agents propose, people decide. Every finding arrives with the rule it maps to, its pass or fail, and the evidence behind it, and the decision you record in the review queue is the decision of record.
Will the same question always get the same answer?
The governance outcome is designed to be reproducible: the same control assessed against the same evidence and methodology produces the same result, and the methodology itself is versioned. Authority evaluation is deterministic and involves no model at all.
How do we explain a finding to an examiner?
Open the finding's decision trace and export it. It carries the context, the reasoning, the authority derivation from control to appetite clause to obligation, the evaluation, and the human decision. It is the chain most examiners ask for.
What is the difference between shadow AI autonomy and shadow authority?
Shadow autonomy is the risk that an agent acts beyond its authorized scope. Shadow authority is the measurable instance: an action for which no valid delegation path existed at that moment. The first is a category of concern; the second is a finding you can act on.
The convergence
The agentic enterprise runs on governed data and AI.
Know what can act. Govern the action. Prove it held. Proactive, independent, ready for the regulator, always on.