Skip to content
Back to home

Case studies

Systems in production. Not slideware.

We don't publish client names by default, our work is often commercially sensitive, but every system below is live and every number is one we can stand behind.

A financial services firm

Financial services · Mid-size, regulated

Turned a compliance knowledge bottleneck into instant, cited answers

The problem. Compliance and front-office staff lost hours a day hunting through policy libraries, regulatory correspondence, and product terms scattered across systems that didn't talk to each other. A wrong or out-of-date answer wasn't a bad experience, it was a supervisory risk.

The baseline. Staff spent a large share of each day locating current policy and regulatory text by hand, with no single source of truth and real risk of citing a superseded version.

What we built. A retrieval-augmented knowledge system grounded in the firm's own current, permissioned documents, answering with the exact clause it drew from. Access control was built into retrieval, and every answer is logged with the source version for audit.

  • 72% reduction in time spent locating current policy and regulatory text
  • 85 staff across compliance and front office using it daily
  • 4 weeks from discovery to a working proof

Architecture. RAG over the firm's document store, with access-controlled retrieval, source-version citations, and full audit logging of every answer.

A professional services firm

Professional services · High document volume

Automated a high-volume document workflow end to end

The problem. A core process depended on people manually reading, classifying, and extracting information from a high volume of documents. It was slow, expensive to scale, and error rates climbed under load, capping how much work the team could take on.

The baseline. The workflow was fully manual, capped by how many documents people could process, with error rates that rose under load.

What we built. A document-intelligence pipeline that extracts, classifies, and structures the information automatically, with an evaluation harness measuring accuracy against a graded set and human review reserved for the genuinely ambiguous cases.

  • 68% of the workload now handled automatically
  • 18,000 documents processed per month
  • 94% measured accuracy on the evaluation set

Architecture. An extraction-and-classification pipeline with an automated evaluation harness and a human-in-the-loop path for ambiguous cases.

A B2B technology company

B2B technology · Scaling support volume

Deployed an AI agent that resolves, not just deflects

The problem. Support volume was outgrowing the team, and a generic chatbot had eroded trust by confidently giving wrong answers. Leadership wanted automation that customers actually trusted, not a deflection metric that hid unresolved problems.

The baseline. Support volume was outpacing the team, and an earlier generic chatbot deflected rather than resolved, leaving customers and staff frustrated.

What we built. A grounded support agent that answers from the company's own knowledge base with citations, escalates honestly when it isn't confident, and integrates with the existing ticketing stack, with cost and latency benchmarked before go-live.

  • 64% of enquiries resolved without a human
  • 41% reduction in average handling time
  • 6 weeks from proof to production

Architecture. A retrieval-grounded support agent with citations, confidence-based escalation, and a live integration into the existing ticketing stack.

Our edge

Engineering depth. Commercial discipline.

Most AI consultancies deliver slide decks. We deliver running systems. Here is what that means in practice.

Outcome first, proof before production

We define what success means in business terms before choosing any technology, then test the critical assumptions on real data before you commit to a major build.

Evaluate, don't guess

Systems are measured against agreed performance criteria rather than judged by how impressive the demo looks.

One team, end to end

The people who design the system build it, deploy it, and can stay on to operate it.

Built around your constraints

Data, security, compliance, latency, reliability, and cost are design inputs, not problems to solve at the end.

No surprise economics

We model expected model and infrastructure costs before production and design around agreed cost requirements.

The right AI, not the biggest

We don't start by picking a model. Hosted frontier model, smaller specialist model, or private deployment, we choose what fits your accuracy, cost, speed, privacy, and control requirements.

Have a use case worth solving?

Tell us what you're trying to automate, improve, or build. We'll tell you whether AI is a good fit, what we'd recommend, what it would take to build, and what it's likely to cost.