The privacy firewall
Entity extraction identifies and redacts personally identifiable information before a prompt reaches an external model. Policy stops being ‘do not paste client names’ and becomes something the system will not allow.
Governance by default
Endeavor enforces AI governance in the platform rather than in a policy document. Guardrails, access control, audit logging, and evaluation are properties of the system, so compliance is something the platform does rather than something people are asked to remember.
Governance as an enabler, not a brake
Traditional assessments measure readiness on paper but supply none of the infrastructure needed to act on it. Endeavor is the scaffolding that builds maturity through deployment: business, IT, and security work inside one environment instead of negotiating across three.
Most organisations believe they must be ready before adopting a platform like this. The reality is the opposite — Endeavor supplies the operating model that makes an organisation ready.
Three guarantees
The flight recorder
Every interaction is logged at user and agent level, including which sources were read. When an agent is wrong you can prove whether the fault was the data, the prompt, or the reasoning.
The IAM wrapper
You already have identity management for humans. Endeavor adds it for agents: an agent inherits the clearance of whoever invoked it and cannot reach a record that person could not open.
Guardrails run on both sides of every model call
A guardrail is itself a small task, applied to the prompt before it is sent and to the text that comes back. That symmetry matters: filtering only the input leaves generated content unchecked, and filtering only the output lets sensitive data leave your environment in the first place.
Content moderation
Screening against safety datasets for categories you choose to block.
PII prevention
Entity extraction plus discard lists and pattern rules to stop leakage.
Supervisor checks
A second pass that assesses whether output satisfies your data-validation rules before anyone sees it.
Prompt-injection resistance
Content fetched from the web is treated as untrusted input. Tools are constrained to declared operations, and every call is logged, so an injection attempt is visible rather than silent.
Coverage across the assessment domains
Governance and oversight
A unified workspace with global rules and guardrails where non-technical stakeholders review, approve, and audit the specific tasks agents perform.
Inventory and ownership
A live agent registry acting as a control tower over the portfolio: every agent, the models it uses, and the data it may touch. Shadow AI stops being invisible.
Data privacy and protection
Automated redaction, encryption in transit and at rest, and source-level logging showing exactly which records grounded a given answer.
Identity and access management
Role-based access separating who builds, who approves, and who consumes, with centralised revocation when someone leaves.
Logging and auditability
Reasoning logs that reconstruct the chain of thought, correlated with an immutable record of every change to an agent's instructions.
Incident response
Forensic traceability plus granular containment - pause or roll back one misbehaving agent without taking down the rest.
Lifecycle and SDLC controls
Version and roll back an agent's instructions, or switch its model, without a developer changing a line of code.
Maturity, measured
| Level | Where most organisations are | What Endeavor supplies |
|---|---|---|
| 1 Ad-hoc | Siloed AI across SaaS tools, no central inventory | One accountability layer over existing services |
| 2 Reactive | Policies on paper, no way to enforce them live | Technical guardrails built into the platform |
| 3 Defined | Formal approval exists but is too slow to use | Decomposition into small, pre-approved tasks |
| 4 Managed | Activity logged, but only IT can see it | Domain experts score and validate outputs directly |
| 5 Optimised | AI managed as a digital workforce with real ROI tracking | The full discovery-to-monitoring cycle |
Deployment and data sovereignty
- Private Kubernetes namespace on AWS, Azure, Google Cloud, or on-premises.
- Bring your own model keys and retain full data jurisdiction.
- Run inside your VPC with egress restricted to endpoints you approve.
- Host models locally with vLLM so no data reaches a third-party provider.
- Self-managed, managed on your infrastructure, or fully hosted by Rotational.
- No GPUs required to get started - standard CPU, RAM, and storage.
Frequently asked questions
How does Endeavor stop sensitive data reaching external models?
Built-in entity extraction using GLiNER identifies and redacts personally identifiable information before a prompt leaves your environment. Guardrails run as mini-tasks on both the input to the model and the text it generates.
If an agent gives a wrong answer, can we tell why?
Yes. Every interaction is logged at user and agent level, including which internal sources were read, so you can establish whether the fault was the source data, the prompt, or the model's reasoning.
How do you prevent prompt injection?
Guardrails are part of the answer, not all of it. Because models accept ambiguous input, content pulled from the web or social media is a genuine attack surface - hidden instructions can attempt to make an agent leak data or take harmful action. Endeavor treats input as untrusted, constrains tools to declared parameterised operations, and logs every call so an attempt is visible.
Can we run Endeavor entirely inside our own environment?
Yes. Deploy into a private Kubernetes namespace on your cloud or on-premises, bring your own model keys, restrict egress to endpoints you approve, and retain full data jurisdiction. Models can be hosted locally with vLLM so no data reaches a third party.
Who can approve an agent for production?
Whoever you decide. Role-based access separates who can build agents, who can approve them, and who can only use them, and approval is evidence-based: domain experts score outputs against test cases before an agent is green-lit.