DoveSureAI Governance
Govern the AI You Build and Buy, From Discovery to Operation
DoveSure governs your organisation’s own AI systems and agents: it finds the AI in use, keeps one inventory with the reason for every risk class, ties each approval to the version that runs, tests your systems at endpoints you authorise, decides what your agents may do before they act, and watches what they do in operation. Dove, DoveSure’s intelligence, reads and proposes; rules enforce; your people decide.
How it works
From Record to Proof
Find it
Discovery reads repositories, agent configurations, the vendor register and agent requests, and reconciles what it finds with the inventory. A person decides what each find is.
Classify and approve
Each system’s requirements follow from its own facts. A prohibited practice cannot be approved, and an approval belongs to the version it was given to.
Test and enforce
The assurance lab tests a system at an endpoint you authorised, and agents ask the control plane before they act, within the grants delegated to them.
Watch and fix
Runtime monitoring judges what a system did with the lab’s rules. A breach opens a finding and proposes a bounded remediation that only a person can approve.
What’s included
- DiscoveryAI found in repositories, agent configurations, the vendor register and agent requests, reconciled with the inventory.
- Inventory and lifecycleEach system’s owner, purpose, risk class with its reason, use cases and the assessments its facts call for.
- Agent control planeAgents ask before they act; each request is allowed, limited, held for a person or denied by its grants.
- Assurance labVersioned test cases sent only to an endpoint you authorise, each answer judged by its rule.
- Runtime monitoring and remediationWhat it did in operation, judged by the same rules; a fix is proposed, and a person approves it.
What It Proves
- An approval covers only the version it was given to; a material change makes it stale and production is refused
- Every agent request is allowed, limited, held for a person or denied by its delegated grants, and is on the ledger
- Every test result names its judge: rules decide, and a model-judged case waits for a person
- Counts that open their records, never a score
Frameworks it serves
What it governs
Each Claim, Proven
Find the AI in use
Discovery reconciles what it finds with your inventory, and a person decides what each find is.
- An AI library declared in a repository and not in the inventory is one observation, naming every place it is declared, and raises one finding; scanning again refreshes it rather than adding a second.
- An MCP server in an agent configuration is found the same way, and no secret value in what discovery reads is stored or shown.
- An agent that asks for a tool it was never delegated, or holds a live credential while not approved for use, is reported, naming the agent and the tool or credential.
- Registering a find, or linking it to a registered system, closes its finding once DoveSure’s own check confirms it.
One inventory, with the reason for every risk class
Every AI system and agent you build or buy, with its owner, purpose, data and what it may do.
- Each requirement a system must meet is worked out from its own facts, and names why it applies, its control and the EU AI Act articles that control maps.
- Use cases, AI literacy, transparency and content marking, post-market monitoring, drift and AI incidents are append-only records, and each brings its requirement only when it applies.
- A serious AI incident gets its reporting deadline from its kind, and shows its clock.
Approval tied to the version that runs
An approval is a person’s decision about one version, and it says when it no longer covers what runs.
- A high-risk system is not approved while any requirement is missing for the version running, and the refusal names each one.
- A material change, such as a new model, makes a new version: the approval goes stale, production is refused, and an agent is treated as not approved until a person approves the new version.
- A move the gate would refuse, such as approving a prohibited practice, is shown disabled with its reason.
- Versions and assessments are never edited or removed; an earlier version’s assessments stay as history.
Agents ask before they act
An agent’s own token asks the control plane, through the API or DoveSure’s MCP server, before it acts.
- Each request is allowed, allowed within limits, held for a person or denied, decided from the grants delegated to the agent.
- A person approves or refuses exactly the request that was held, once, and an approval never includes a field the grant withholds.
- Every decision and review is on the ledger and the audit log.
Test your AI at an endpoint you authorise
The assurance lab sends versioned cases only where you allowed it to, and judges each answer by its rule.
- Nothing is sent to an endpoint without an authorisation; production needs a time window, and private addresses are refused.
- Every result names its judge. A rule-judged case is decided by its rule; a model-judged case stays a labelled suggestion until a person confirms or overturns it.
- The cases cover prompt extraction, injection, jailbreaks and tool misuse, fairness pairs, data minimisation, robustness and the declared output contract.
- A failed case opens a finding and a later passing run closes it; an endpoint that cannot answer is never reported as passing.
Watch what it does in operation
A system reports what it did, and the lab’s own rules judge it.
- Only a token bound to a registered AI system reports its traces, and DoveSure keeps hashes and redacted excerpts, never the whole text.
- Each trace is checked for a planted canary, personal data where none is declared, a secret-looking value, a tool outside its grants and a blocked topic answered.
- A recent window is compared with its baseline; too few traces is not judged, and never taken as a pass.
- A breach opens a finding, and a serious one proposes a suspension for a person to approve.
Fix it within bounds, with a person’s approval
Detect, propose, approve, carry out, verify, and roll back where it can be.
- A proposal changes nothing; it shows the approver its bounds and whether it can be undone.
- Approved, it goes through the same path as a person’s change and is verified by reading the effect back; one that cannot be verified is rolled back where possible, and says so where not.
- DoveSure’s own detections propose and never execute.
See what depends on what
Vendors, models, data, tools and credentials, as links that name their records.
- An AI system’s vendor, model, data, delegated tools and live credentials appear as links, each naming the record that states it.
- A change to a vendor, dataset or service shows every AI system it reaches, with the path, and proposes a reassessment to each owner without changing the systems.
One overview, and the next step
Counts read from the records, and a guided journey from the first system to proof.
- Six questions are answered by counts read from the records, each opening exactly the rows it counted.
- Last observed, last evaluated and last approved are kept apart.
- A guided journey shows the next step, from registering a system to keeping nothing waiting, each step read from your records and never ticked by hand.
Evidence that verifies offline
What AI governance produces is evidence, on the same ledger as the rest of your programme.
- An assurance run is kept as system-observed evidence of the version it tested.
- Agent decisions and their reviews are on the hash-chained ledger, verifiable offline in a signed audit export.
Not claimed yet. Discovery of AI applications in single sign-on, cloud machine-learning services and browser extensions, and in source code beyond package manifests, is coming.
How it differs
AI interprets. Rules enforce. People decide.
Dove, DoveSure’s intelligence, reads, explains and proposes. Deterministic rules decide what an agent may do and what a system may be approved for, and only an authorised person approves a change.
No bare scores
Counts that open the records behind them, never one percentage.
Every result says what produced it
A test result names its judge and a run the version it tested; what is unknown is shown as unknown, never as a pass.
One product, not another silo
AI governance shares the controls, findings, evidence and ledger of your security and privacy programme, with ISO/IEC 42001, the NIST AI RMF and the EU AI Act mapped like any other framework.
Questions
What People Ask
Does DoveSure govern our own AI, or only use AI itself?
Your own. AI governance covers the AI systems and agents your organisation builds or buys: discovery, the inventory, approval of the version that runs, an agent control plane, testing, runtime monitoring and remediation. Dove, DoveSure’s intelligence, is governed by the same rules: your permissions, your confirmation, on the record.
Does this cover AI we buy as well as AI we build?
Yes. Register any AI system or agent your organisation uses, with its owner and purpose, whether you built it or bought it. Discovery also reads your vendor register for AI suppliers the inventory does not account for.
Does DoveSure decide whether our AI may be used?
No. DoveSure works out what each system needs and refuses an approval its rules forbid; an authorised person in your organisation approves each version.
Related
The Rest of the Platform
Ready to prove it?
Know what you are responsible for, control whether it works, and prove it to anyone who asks.