Your analysts, working with AI.
Properly.
I've trained hundreds of analysts to work with AI across fintechs and scale-ups, including the first AI-assisted analytics training at a global fintech, and I build the Claude Code skills analysts run daily to fix charts, repair explores and migrate reports. Not a prompting webinar. Working skills, real workflows, and a team that keeps using them after I leave.
Aggregated across consulting engagements at fintechs and scale-ups. Happy to walk through the detail on a call.
Why AI rollouts fizzle in analytics teams
The licence is not the hard part. These four problems are. I have heard every one of these sentences in real life.
Everyone has a licence, nobody has a workflow
"We rolled out Claude to the whole team. Usage peaked in week two."
Buying the tool is the easy part. Without skills built for your stack and workflows your analysts actually follow, AI stays a fancy autocomplete that people quietly stop opening.
The AI gives confident, wrong answers
"It wrote the SQL in seconds. The number was off by a factor of four."
An AI agent with no context about your dbt models, metric definitions and naming conventions will hallucinate. The fix is not a better prompt. It is giving the agent your semantic layer.
AI knowledge lives in two people
"Ask Tom, he's the one who knows how to do the agent stuff."
One or two enthusiasts get great results while the rest of the team watches. Without structured training and shared skills, the gap widens instead of closing.
The training was a webinar
"There was a lunch-and-learn about prompting. I think there are slides somewhere."
Analysts learn by shipping, not by watching. If they have not fixed a real chart or opened a real PR with AI assistance by the end of the first session, the training did not happen.
What's included
Five areas of work. The mix and depth depend on your team, and we establish that on the first call.
Workflow Audit & Use-Case Mapping
- Shadow your analysts to find where the hours actually go
- Map your stack: dbt, BI tool, repos, review process, support channels
- Pick the use cases where AI assistance pays off fastest
- Agree what stays human: review, sign-off, stakeholder judgement
Custom Claude Code Skills for Your Stack
- Skills built around your dbt models, your YAML conventions, your repos
- Skills I have shipped before: fix broken charts, repair explores, migrate reports, all run by analysts themselves
- Each skill encodes your standards, so the output passes review
- Versioned in your repo and maintained like any other code
Hands-On Analyst Training
- Live sessions where every analyst ships something real, usually within 90 minutes
- Curriculum from first prompt to confident PR workflows
- Includes the first AI-assisted analytics training at a global fintech
- Sessions run across offices and time zones, not just HQ
Train-the-Trainer & Ambassadors
- An Ambassador programme: your strongest analysts become the internal trainers
- Trainer guides and session materials they can run without me
- Office hours model for the long tail of questions
- A named owner for the skills and the curriculum on handover
Documentation & Guardrails
- Skill usage guides and troubleshooting cheat sheets analysts actually open
- Review standards: what AI output needs a second pair of eyes
- Prompt patterns and worked examples from your own codebase
- It all lives in your wiki, not mine. You keep everything
An analyst types one command.
The skill does the boring half.
Take a broken chart. At my last engagement, an analyst would run a fix-chart skill in Claude Code. It reads the chart, traces the failure back through the YAML and the dbt model, proposes the fix, and drafts the PR. The analyst reviews, adjusts and merges. What used to be a support ticket and a two-day wait became a ten-minute job the analyst owns end to end.
That is the pattern for every skill I build: the agent does the tracing and the typing, your analyst does the judgement. The skill encodes your conventions, so what comes out the other end passes review.
Built for your stack, not a demo
Skills are written against your repos, your YAML conventions and your review process. They ship as code in your repo.
Run by analysts, not by me
The goal is not me being productive with AI. It is your whole team being productive without me.
Value in the first session
Training runs on real backlog items. Analysts fix actual charts in the first 90 minutes, not toy examples.
I only do this with Claude Code
Not because other tools are bad. Because skills, sub-agents and the workflows I build on top of them are where I have done the deep work, in production, with hundreds of analysts. You are paying for that depth, and I will not pretend to have it everywhere.
If your team is on Claude Code, or ready to adopt it, we will move fast. If you are committed to a different stack, I will say so on the first call and save us both the time.
This is the right fit if...
- Your team already uses Claude Code, or you are ready to adopt it
- Your stack runs on dbt and a modern BI tool, so there is a semantic layer for the AI to stand on
- Your analysts spend more time fixing charts and chasing definitions than analysing
- One or two people get great results with AI and you want that to be everyone
- You want adoption that survives after the consultant leaves
This is not the right fit if...
- You are hoping AI will replace your analysts rather than upgrade them
- There is no dbt or semantic layer yet, so the AI has nothing reliable to stand on
- You want a strategy deck about AI rather than working skills and trained people
- Your security or compliance posture rules out AI coding tools entirely
Every team starts from a different place. Let's find yours.
Some teams need the skills built first. Some need the training. Most need both, in an order that depends on where the pain is. I work that out with you on the first call, and everything else follows from there.
30-minute call. No obligation. I'll tell you honestly if this isn't the right fit.
Engagements start with a fixed-fee diagnosis: £8,950 + VAT. If you don't see value in the first week, we stop, and you pay half.
Mid-migration or planning one? This pairs naturally with my Looker to Lightdash migration work, where these skills were born.