AI Agent Development Company in Leeds

How much human sits in the loop is the real cost driver on a Leeds agent build: a scoped agent proof of concept covering one workflow with three to five tools at Leeds rates, a production single-task agent for a building society or NHS Trust, or a multi-agent supervisor system for ASDA-style operations workflows or full Consumer Duty complaint pipelines. Scope drivers include the tool-use policy, an evaluation harness, human-in-the-loop controls, an initial regulatory bundle (DCB0129 hazard log or SS1/21 materiality assessment), tenant isolation, audit logging, the full regulatory pack, red-team testing, and a 12-month support window. Leeds rates sit below London but the NHS and FCA Consumer Duty expectations are identical, and we price for the human-in-the-loop and audit overhead rather than discovering it during change requests.

Any agent that touches clinical data opens a DCB0129 hazard log on day one, with a registered Clinical Safety Officer engaged before architecture. Tool calls are classified by impact: read-only retrieval is logged, advisory outputs route to a clinician, and any state change to an EHR or care record requires explicit clinician approval through a recorded human-in-the-loop step. We ship a Clinical Safety Case Report aligned with DCB0129 v4.2, a hazard log with severity and likelihood ratings, mitigations including retrieval-grounded outputs only, structured schemas with refusal patterns, confidence thresholds, and post-deployment monitoring. The deploying NHS organisation completes DCB0160 against our DCB0129 bundle. We will not ship a clinical agent without a CSO sign-off.

Only with explicit human approval on any account-changing action, and only after SS1/21 model risk validation and a Consumer Duty outcome assessment. The default agent pattern in retail banking is read-only retrieval and advisory output, with a human agent or customer pressing the final confirmation button. Tool calls that move money, change product terms, or close accounts require step-up authentication and a documented human-in-the-loop step that an FCA supervisor or Financial Ombudsman case-handler can audit. We map every tool to PRIN 2A Consumer Duty outcomes (products and services, price and value, consumer understanding, consumer support) and document vulnerable customer detection triggers under FCA Financial Lives patterns.

Default residency is UK South (Azure London) or eu-west-2 (AWS London) for agent state, tool-call audit logs, and any model inference. NHS deployments use the dedicated NHS England Microsoft commercial tenant with customer-managed keys in Azure Key Vault HSM. Financial services deployments use AWS Bedrock eu-west-2 with KMS customer-managed keys and PrivateLink, or Azure UK South with the equivalent. Tool-call logs are retained for the regulatory minimum (typically six years for FCA records under SYSC 9, eight years for NHS records under the Records Management Code of Practice for Health and Social Care 2021), then deleted on a documented schedule. International transfers run under the UK IDTA with a Transfer Risk Assessment, never on assumption.

The NCSC Cyber Assessment Framework (CAF) is the UK's outcomes-based framework used by Ofgem, the Department of Health and Social Care, the Department for Transport, and other competent authorities to assess regulated operators. For NHS deployments the CAF maps to the Data Security and Protection Toolkit, for financial services it maps to FCA and PRA operational resilience expectations under SS1/21 and SYSC 15A. Every Codazz agent ships with a CAF outcome map covering managing security risk, protecting against cyber attack, detecting cyber security events, and minimising the impact of cyber security incidents. We produce the evidence pack the client's assessor or supervisor will actually ask for, including tool-use boundary controls, prompt-injection defences, and incident response runbooks tested against tabletop exercises.

Defence runs across input filtering (Azure AI Content Safety, NeMo Guardrails, custom validators on tool-call shapes), system-prompt hardening with read-only injection of authoritative content, output validation against schema and policy, tool-call allowlisting with per-tenant scopes and per-tool rate limits, mTLS or OAuth on every tool endpoint, and continuous adversarial red-team testing against OWASP LLM Top 10 and Meta Purple Llama harnesses. Every production push runs the adversarial harness in CI. For First Direct, Virgin Money, and NHS Trust deployments we coordinate with the client's existing red team and feed findings into their bug-bounty programme. We document residual risk in the model card and the CAF assessment, not in a marketing claim.

Both, but we default to single-task agents until the workflow genuinely requires coordination. Most Leeds use cases (a complaint triage, a mortgage document extraction, a clinical referral router) are best solved by one well-tested agent rather than a supervisor-worker tree that adds debugging complexity and audit surface. When we do build multi-agent systems (typical for ASDA-style operations workflows or full Channel 4 commissioning research pipelines) we use LangGraph supervisor-worker patterns with explicit state passing, per-agent tool scopes, and a single supervisor responsible for human-in-the-loop checkpoints. Every agent in the tree has its own model card, evaluation harness, and audit log stream. We will tell you when a multi-agent pattern is genuinely justified versus when one well-built agent solves the problem at a fraction of the operational cost.

Channel 4 HQ at The Majestic and Channel 4 commissioning teams operate under the Ofcom Broadcasting Code including Section 5 due impartiality, Section 2 harm and offence, and Section 7 fairness. Editorial research agents we build surface citations on every claim, refuse to fabricate quotes, and route flagged outputs to human commissioning editors rather than producing publishable content directly. Tool calls cover licensed databases (Factiva, Westlaw, FAME), the Channel 4 internal archive, and the open web with provenance tracking. We map every output to the Channel 4 Producers Handbook and document the impartiality controls. No agent we ship can post directly to a Channel 4 publishing or scheduling system without a named human producer in the approval loop, regardless of demo enthusiasm.