High Assurance AI — A CommonAI CIC Programme

Opportunity

AI holds enormous promise in sectors where compliance, correctness and safety are of business critical importance.

The deployment of AI solutions (including LLMs) requires systems that operate safely, reliably and at scale. The High Assurance AI Programme will build the technical foundations that make safe, compliant AI deployment achievable in critical and regulated industries.

Barclays is the programme anchor, with initial focus on solving strategic use cases and industry-wide challenges in financial services.

Current Workstreams

These are the programme’s technical building blocks, each delivered by a dedicated working group, consisting of member organisations and CommonAI Engineers. New workstreams will be added as the programme develops, and you can see those currently in development here.

If you’d like to get involved or have ideas for new workstreams, please get in touch via this form.

Model Fabric

Small Language Models

Model Fabric is a shared framework for building small, specialised AI language models (SLMs). A model with a clearly defined job is more reliable than one expected to know everything. Firm boundaries and carefully prepared training data make its answers more consistent and trustworthy.

Panopticon

Monitoring & control

Panopticon is a unified control platform, developed by Cambridge Inference, for overseeing, securing and auditing how AI is used across an organisation. Panopticon checks AI-generated code, stops sensitive data from leaking through prompts, and keeps a clear record of what the AI does.

Auto-formalisation

Formal verification

Autoformalisation turns written policies, regulations and business rules into precise, machine-checkable logic that computers can apply without ambiguity. This allows important decisions to be verified consistently, backed by proof and a clear audit trail, rather than relying on an AI's best guess.

Workstreams in Development

Multi-Agent Systems

Coordination frameworks

Multi-agent systems use teams of specialised AI agents to tackle complex business problems. In regulated sectors, this coordination must be predictable, transparent and verifiable, so each agent works within set limits on its domain, data and actions, communicating through a governed layer that makes every decision traceable.

Cross Section Areas

Cross Section Area

Synthetic data

This area sets up repeatable processes for sourcing, cleaning and versioning training data, and approaches for generating synthetic data to expand datasets, protect privacy and cover rare scenarios, both still subject to further research. The goal is high-quality, reusable, compliant datasets.

Cross Section Area

Domain definition

This area defines business objectives, domain boundaries, success criteria and data ownership. Tightly scoping what a model can do and access makes its outputs more consistent and testing more practical, while fixed output formats act as guardrails.

Cross Section Area

Security

Security underpins the whole programme through four principles: securing the model against evolving attacks, the data through strong controls, the operation through protected architecture and integrations, and the usage through governance that ensures only authorised use.

Technical Roadmap

Each programme has a Technical Roadmap, set under CommonAI’s Governance Framework, which outlines its current work plan, including high-level targets, major milestones and the prioritisation of key features. It is a living document that evolves as the programme develops. Attached is the public version of the roadmap, while a more detailed internal version, setting out the full technical scope and project plan, is available to members.

Technical Roadmap (PDF)

Join the programme

Organisations join CommonAI programmes as members, building foundational AI and ML technologies alongside other members and CommonAI’s engineers. Members commit engineering time to the Technical Roadmap and pay an annual fee ring-fenced for their chosen programmes. In return, they keep full ownership of their work and gain early access to the resulting IP, a role in programme governance, member-only resources, and a pathway to production. For more detail on membership, see our Member Agreement and commonai.org.

If you are interested in joining the High Assurance AI programme or have any queries please contact us via this form.

Current Members

Programme Anchor

Anchor to the High Assurance programme, providing strategic direction and applying the technical foundations within financial services.

AI Control Platform

Develops high-assurance AI and optimisation systems for regulated environments. For the programme, it is building Panopticon, a monitoring and control layer that audits AI interactions and sandboxes executed code to catch risks in real time.

Auto Formalisation

Specialises in formal methods and proof-based verification of complex software and AI systems. It brings expertise in making system behaviour precise, auditable and deployable in regulated environments.

Synthetic Data

A verticalised foundry that turns specialised workflows and fragmented records into representative synthetic training data. Within the programme, it works across data fidelity, synthetic data, evals and training methods.

AI Safety Research

Independent research and incubation hub focused on AI safety, biosecurity, frontier-risk policy and institutional design. It is collaborating with Cambridge Inference on the programme.

Building the technical foundations that make safe, compliant AI deployment achievable in critical industries.

Small Language Models

Monitoring & control

Formal verification

Coordination frameworks

Cross Section Areas

Synthetic data

Domain definition

Security

Current Members

Why the Future of AI Isn’t Built Alone — Venture Café London

High Assurance AI Technical Roadmap published

High Assurance AI members launch event

Cambridge Tech Week fringe event