AI systems with evaluation, governance and a human in the right place.

The problem

AI programmes commonly start from the technology rather than a costed problem, and stall when the underlying data turns out to be unusable.

What the work involves

We run workshops with the people doing the work, quantify the current cost of each candidate process, and assess data availability, quality, ownership and lawful basis for each. Every use case gets a score across value, feasibility, data readiness and risk, and a recommendation — including an explicit "do not do this" where that is the honest answer.

Typical deliverables

• Scored and ranked use-case register
• Data readiness assessment per use case
• Risk and regulatory screening, including DPIA triage
• A sequenced roadmap with a recommended first proof of value

The problem

Demonstrations succeed on curated inputs. Production fails on the long tail, and an ungrounded model will answer confidently rather than admit it does not know.

What the work involves

We ground generation in your own content, constrain outputs to structured schemas where the downstream system needs reliability, and design explicitly for refusal and escalation. Provider access sits behind a gateway abstraction so a model change is a configuration decision, and cost and latency are controlled at the boundary.

Typical deliverables

• Model gateway with provider abstraction and cost controls
• Prompt and schema versioning with structured output validation
• Guardrails, redaction and refusal behaviour
• Human review paths for consequential output

The problem

Naive retrieval returns plausible answers from documents the asker should not be able to read, and cites nothing you can check.

What the work involves

We build retrieval that enforces the source system’s permissions at query time rather than at indexing time, so an answer can never be assembled from content the user cannot access. Every answer carries citations to its sources. Retrieval quality is measured against a labelled evaluation set rather than assessed by impression.

Typical deliverables

• Ingestion pipeline with permission-aware indexing
• Hybrid retrieval with reranking, tuned against a labelled set
• Cited answers with source-document links
• Retrieval and answer-quality evaluation suites

The problem

An agent that can act on your systems is a security and safety design problem long before it is a model problem.

What the work involves

We define the tool surface narrowly and explicitly, run every action through the same authorisation checks a human user would face, and require confirmation for anything consequential or irreversible. Every action is logged with its inputs, the reasoning available, and the identity it acted on behalf of.

Typical deliverables

• Scoped tool definitions with least-privilege authorisation
• Confirmation and approval gates for consequential actions
• Full action audit trail with attribution
• Escalation to a human with the full context preserved

The problem

Chatbots deployed without a handover path trap users in loops and damage trust more than having no bot at all.

What the work involves

We design the handover to a human first and the automation second. Conversations carry their full context across the handover. AI participation is disclosed to the user. We support web chat, WhatsApp, SMS and voice through provider interfaces, and we build against mock adapters where credentials are not yet available rather than pretending an integration exists.

Typical deliverables

• Conversation design with explicit escalation paths
• Channel integrations behind provider interfaces
• AI disclosure and consent handling
• Transcript retention, moderation and reporting

The problem

Extraction accuracy is never one hundred per cent, and a pipeline with no review step quietly writes wrong data into systems of record.

What the work involves

We define per-field confidence thresholds and route anything below them to a human review queue with the source document visible alongside the extracted values. Corrections feed the evaluation set. Accuracy is reported per field and per document type rather than as a single headline number.

Typical deliverables

• Classification and extraction pipeline
• Confidence thresholds with a human review queue
• Per-field accuracy reporting
• Retention and secure deletion for source documents

The problem

Significant staff time goes on moving data between systems that were never integrated, and on chasing approvals through email.

What the work involves

We map the process as it is actually performed, including the workarounds, then automate the mechanical parts and leave judgement with people. Automations are durable: they record trigger, version, inputs, result, error, retries and any human override, and they respect communication consent and quiet hours.

Typical deliverables

• Process map of the as-is and to-be flows
• Durable, observable automation with retry and dead-letter handling
• Approval and exception routing
• Before-and-after measurement against a baseline

The problem

Without evaluation, AI quality is a matter of anecdote, and a prompt change made on a Friday can silently degrade output for months.

What the work involves

We build labelled evaluation sets from your real cases, run them in CI on every prompt or model change, and track quality, cost and latency over time. Every run is recorded with model, prompt version, inputs, outputs, cost, latency, feedback and human overrides — which is also what your risk function needs to approve the system.

Typical deliverables

• Labelled evaluation sets and automated scoring in CI
• Per-run records: model, prompt version, cost, latency, outcome
• Quality, cost and drift dashboards
• AI register, governance documentation and review workflow