Track record

Six engagements, with the numbers attached.

Client names are under NDA. Sector, approach, outcome and the technical stack are disclosed in full — because that is the part you can actually assess us on.

Case studies

Regulated work, held to a review gate.

Every system below has an explicit human review path. Where a figure cannot be verified from source, it is annotated rather than resolved — the model does not get to decide which number was right.

Financial Services

— Client name NDA —

Regional Brokerage · Research Analyst Agent

Challenge

Analysts spent 3–4 hours daily assembling market briefs from fragmented sources.

Approach

Multi-agent system on Claude Agent SDK: a coordinator delegates to search, document-analysis and synthesis subagents over MCP interfaces to internal market data. Conflicting figures are annotated, never silently resolved.

Outcome

340 analyst hours returned per month

Brief preparation reduced to 25 minutes.

Claude Agent SDK · MCP · AWS Bedrock · PostgreSQL

Insurance

— Client name NDA —

Insurance Group · Claims Document Extraction

Challenge

12,000 monthly claim documents in inconsistent formats, manually keyed.

Approach

Structured extraction via tool use with strict JSON schemas. Nullable fields prevent fabricated values; field-level confidence scores route ambiguous cases to human review.

Outcome

94% straight-through processing

Manual review load down 78%. Error rate below the prior manual baseline.

Claude API · Message Batches API · Pydantic validation

Payments

— Client name NDA —

Payments Provider · Compliance Query Assistant

Challenge

Compliance team answering repeated internal policy questions, 6-hour average turnaround.

Approach

Retrieval-backed assistant with strict citation requirements and explicit escalation criteria. Programmatic gates block any response on regulated topics without a cited source.

Outcome

71% of queries resolved without human involvement

Median response under 2 minutes.

Claude API · MCP · internal document store

Legal & Professional

— Client name NDA —

Accounting Firm · RFP Response

Challenge

A high volume of incoming RFPs required manual review and drafting, stretching a limited team thin.

Approach

Claude reads each RFP, extracts the requirements, and drafts a response grounded in the firm's own past submissions. Every draft is flagged for human review before sending.

Outcome

100% of drafts reviewed by a human before sending

More RFPs answered, and answers grounded in prior, proven language.

Claude API · MCP to the submissions library · retrieval with source citation · human review gate

FMCG

— Client name NDA —

Change Management Intranet

Challenge

A global change rollout was being slowed by the need to update content across six languages for every region.

Approach

Extraction from the legacy sites, then automatic generation of editable, multilingual SharePoint pages per region — structured output, so every page stays on template.

Outcome

6 languages generated per region, automatically

Faster regional adoption and a more consistent migration across markets.

Claude API · JSON Schema extraction · Microsoft Graph / SharePoint REST · Azure Functions

Arts & Culture

NGO VOS Gala

Gen AI Animation

Challenge

VOS budgeted for a simple one-minute animated loop of a dragon to accompany their choir performance.

Approach

An AI-driven production run with minimal human direction — generation, refinement and finish handled end to end by the pipeline.

Outcome

7 min delivered against a one-minute brief

Closed the performance to a standing ovation.

Claude API shot direction · text-to-video · upscale & interpolation · compositing

On prior work. Some of the capability evidenced here was delivered by our principals and their teams before Claude Labs was formed, under their own firms. It is presented as evidence of team capability — not as Claude Labs engagements.

Delivery method

Five phases, two-week increments.

Quality assurance is not a phase at the end. Independent code review, automated evaluation of agent behaviour and client-run UAT run against the criteria agreed at design — nothing is done until it passes them.

01

Discover

Scope, stakeholders and success criteria agreed.

02

Design

Architecture, interfaces and written acceptance criteria.

03

Build

Two-week increments, demonstrated each cycle.

04

Validate

Independent review, automated evaluation, client UAT.

05

Handover

Documentation, training and repository access.

Next steps

References available on request.

We can put you in touch with referees from delivered engagements, and share the full capability pack including policy documents.

Book a discovery call How we protect you
01Discovery callUp to 1 hour, no charge
02Scoped proposalWithin 5 working days
03Full capability packPolicy documents and referees on request