What can our firm actually do with AI once this is in place?
That depends on your sector, but the most common use cases we support are: drafting client-facing documents (suitability letters, reports, correspondence) with an AI assistant; summarising lengthy documents or research; running AI-powered workflows in case management, practice management, or CRM tools; and building internal AI tools on top of the OpenAI or Anthropic API. The gateway enforces your rules on all of it — PII blocked, policy guardrails applied, and a complete audit trail generated automatically. The goal is to say yes to more AI use, not less.
Does it work with the tools we already use?
Yes, if your AI tool calls an OpenAI-compatible API or the Anthropic API — which covers ChatGPT API, Claude, Azure OpenAI, and most internally built agents and workflows. The change is one base URL. Browser-based tools like ChatGPT web or Microsoft 365 Copilot work differently and are not within scope.
What happens if your proxy goes down?
Our infrastructure runs on Cloudflare's global edge network — one of the most reliable platforms on earth. In the event of an outage, requests can be configured to fail-safe (block all AI activity until restored) or fail-open (pass through directly to the model). You choose the behaviour that matches your firm's risk appetite.
Where is our data stored? Is it UK-based?
Audit logs are stored on Cloudflare's infrastructure. EU and UK data residency options are available. We do not store the content of AI responses beyond the audit record — and we never use your firm's data to train AI models or improve third-party systems.
How long does integration actually take?
Most firms are live within an hour. There is no new software to install on your systems. You change one API endpoint URL per tool, configure your compliance policy rules in the dashboard, and you're logging. Your IT team will not need to be heavily involved.
What does "tamper-proof" actually mean?
Each log entry is HMAC-SHA256 hashed and cryptographically chained to the previous entry. Altering any historical record — even a single character — breaks the chain and is immediately detectable. This is the same principle used in financial ledger systems and is what regulators mean when they specify "immutable" audit logs under EU AI Act Article 12.
Does this replace our existing AI tools?
No. Inference Agents sits transparently between your API-connected tools and the AI provider. Your team keeps using the same interfaces, the same workflows, the same models. The only difference is that every API call is now governed, PII-scanned, and logged. Nothing is removed — governance is added.
We already have a firewall and DLP — aren't we covered?
No, and this is one of the most common misconceptions we encounter. Firewalls and DLP operate at the network layer — they scan for known data patterns like credit card numbers or NHS identifiers leaving your perimeter. They cannot evaluate whether an AI response constituted regulated advice, whether a recommendation was suitable for a specific client, or whether your Consumer Duty or SRA obligations were met. They also write standard event logs — not cryptographically chained records. If a regulator asks for immutable evidence that your AI audit trail has not been altered, standard DLP logs cannot satisfy that requirement. DLP and Inference Agents address different layers. Most firms with mature IT security still have no AI-specific governance in place.
How is this different from the audit logs in OpenAI or our AI provider's platform?
AI provider audit logs — such as OpenAI's organisation audit log API — record administrative events: who accessed the platform, which API keys were used, configuration changes. They do not capture the content of conversations in a form suitable for regulatory review, they are owned and controlled by the provider rather than your firm, they cover only that provider's models, and they are not cryptographically chained in a way that proves tampering has not occurred. If you use multiple API-connected AI tools — ChatGPT API, Claude, Azure OpenAI, an internal agent — each has its own proprietary log format in a different system. Inference Agents gives you one consistent, tamper-evident audit trail across every API-connected AI tool your firm uses, in a format you control and can export for regulatory review.
We have an AI policy document — isn't that enough?
A policy document demonstrates intent. Regulators require evidence of practice. When the FCA, SRA, or ICAEW conducts a review, they will ask to see technical records showing how your AI policy was enforced in practice — not just that it existed on paper. A policy that says "staff must not input client data into unapproved AI tools" with no technical control or audit log to support it offers no protection when something goes wrong. The firms that face enforcement action are rarely those without policies — they are those whose policies existed but whose compliance could not be evidenced.
We use Microsoft Copilot with Purview — does that cover our obligations?
Partially, but not completely. Microsoft Purview captures some Copilot activity within the Microsoft 365 ecosystem and provides audit events in Microsoft's proprietary format. However, it covers only Microsoft tools — if your firm also uses ChatGPT, a custom AI agent, Claude, or any non-Microsoft model, those interactions are ungoverned. Purview audit logs are also not cryptographically chained, which means they do not meet the tamper-evident standard required under EU AI Act Article 12 or expected by financial regulators requiring immutable records. If your firm is exclusively Microsoft and Copilot today, that position will not hold as AI adoption expands. Inference Agents works alongside Purview and fills the gaps it leaves.
We only use AI occasionally — do we really need this?
Frequency does not reduce liability. One unsuitable AI-assisted recommendation to a client, one AI-generated document containing a hallucinated legal citation, or one instance of client data being processed through an unapproved tool is sufficient for an enforcement action or professional disciplinary proceeding. Regulators do not apply a volume threshold — they apply a standard of care. If an AI tool was used in the delivery of regulated services, your firm must be able to evidence how that use was governed. Occasional use also tends to become regular use faster than compliance frameworks can keep pace with — the time to put governance in place is before a problem occurs, not after.
What about staff using personal ChatGPT or Claude subscriptions for work?
Browser-based consumer AI tools are a policy and culture problem, not one a proxy can solve without routing all company HTTPS traffic through a network inspection layer — a significant IT project that introduces bottlenecks, requires endpoint agents, firewall changes, and SSL certificate deployment, and that most regulated firms rightly will not undertake. Our product governs the AI tools and workflows your firm has deliberately deployed via the API. That is where your regulatory accountability is highest: systematic, firm-sanctioned processes that produce client-facing work or inform regulated decisions. We recommend pairing Inference Agents with a clear acceptable use policy that prohibits personal AI subscriptions for client work. The combination of technical governance for your deployed tools and a written policy framework for everything else is a defensible compliance position — and the one regulators expect to see.
We run our AI through Azure — doesn't Azure AI Foundry cover this?
Azure AI Foundry governs what you deploy — model management, access control, content filtering, and resource administration. It does not govern what your AI says. Its audit logs record administrative events, not a per-interaction record of every prompt, response, and compliance determination. Azure's content safety filters are designed to block harmful content categories; they are not configured around FCA Consumer Duty obligations, SRA client confidentiality rules, or ICAEW ethical standards. Foundry also only covers models deployed through Azure OpenAI Service — if any tool in your firm calls Claude, a direct OpenAI endpoint, or any third-party AI, those interactions are invisible to Foundry. Finally, Azure Monitor logs are not cryptographically chained in a way that satisfies regulators requiring tamper-evident records. The simplest way to frame it: Azure Foundry governs what you deploy. Inference Agents governs what it says.