Solutions with Claude

We build solutions and train teams with Claude.

Claude, by Anthropic, is the model we use to build most of our client solutions and to train their teams. Here is what we do with it, with which products and in which real cases, without attributing to Claude what we did not build with it.

Which Claude products we work with

Claude API

Claude Sonnet and Claude Haiku models integrated into solutions running in production: auditable SQL generation, classification and drafting with structured output, tool use, and document extraction.

Claude Code

Our building tool. We used it to develop the solutions in the cases below, this site and its MCP server for agents. It is also the central tool of the academies and of AI Product Development for companies.

Claude for teams

We teach product, PMO, risk and talent teams to work with Claude on their real processes: versioned prompts, output evaluation, data rules and deployment.

Cases built with Claude

Change Tracker Core — auditable advisor for Product Owners (banking)

It turns the history of one or more Jira projects into a local analytical database, learns a versioned profile of the workflow and answers Product Owner questions with reproducible calculations. Claude generates the SQL; the numbers come from the database, never from the model.

ModelClaude Sonnet 5 via the Claude API (conversational agent, issue classification and document ingestion with structured output).
IntegrationsJira Cloud (incremental ingestion with reconciliation), Microsoft Entra ID for B2B identity, DuckDB as the analytical database, deployed on Railway.
DataPseudonymization before any model call: Claude only sees aliases (P-001, B-001); real names never enter the prompt or the history. Snapshots, decisions and audit are append-only.
OutcomeAnswers with visible provenance SQL and citations, a curated project knowledge base with mandatory human review, and verified isolation between projects.

Risk Preparations — risk assurance without committees (Banco Pichincha)

From a guided form it generates the first risk matrix in the bank’s official template and its backlog, reconciles it with existing matrices and publishes the backlog to Jira. It is the "4 days from idea to first version" case we tell in AI Implementation.

ModelClaude Sonnet 4.6 (tool use) and Claude Haiku 4.5 (drafting, backlog, delivery) inside a multi-provider architecture with automatic failover, so no stage depends on a single provider.
IntegrationsJira (one task per story, durable queue with retries), Microsoft Entra ID, Supabase with RLS, the bank’s official Excel template as a read-only source of truth.
DataNo fine-tuning: two years of real matrices enter as governed, minimized examples in the prompt, with deterministic mapping first and Claude only for what remains unresolved. Every classification records provider, model and prompt version.
OutcomeV1 and V2 matrices with live formulas in the official workbook, a backlog synchronized in Jira, and fail-closed behavior: a partial matrix is never published as success.

This site and its MCP server

agilitychanges.com is maintained with Claude Code and exposes an MCP server so Claude and other agents can query courses, prices and contact options and start enrollments. It is the same pattern we apply with clients: the solution exists to be used by people and by agents.

How we work with Claude at a client

Claude and Anthropic are trademarks of Anthropic, PBC. Agility Changes builds with Claude; we do not speak on behalf of Anthropic.

Let’s talk about your case