AI experience

How I work with AI

I use AI, mainly Claude (Anthropic), for whole tasks rather than one-off questions: research and regulatory checks, financial analysis, data processing, drafting correspondence, and producing finished documents ready to send or use.

• Set up for the job. I work in dedicated AI workspaces with persistent context and access to local files. Where a task needs it, I direct the AI to write and run Python, for example to build Excel and Word outputs or generate formal letters as PDFs.
• Structured handovers. I write structured prompts and handover briefs so work carries cleanly across sessions and between tools.
• In my own voice. I've built a reusable prompt library and a structured AI working profile, including a style guide drawn from my own sent emails, so drafts keep to my tone and professional standards.
• Always checked. I treat AI output the way I'd treat a junior colleague's work: useful and fast, but checked against source data before it goes anywhere. I challenge its reasoning, correct its errors, and revise positions explicitly when new evidence changes the analysis.

Building a live AI product: Regarding

I designed and built Regarding on my own, from concept to launch in September 2026. It helps people write a personalised, sourced letter to their MP in a few minutes. Users enter a postcode, the platform finds their MP through the Parliament Members API, and Claude writes a first draft through an N8N automation backend.

Each issue has a system prompt I wrote and refined, setting the evidence base, tone, parliamentary ask and limits. Getting this right took significant iteration: drafts had to be credible enough to send unedited, but personal enough that people would engage with them rather than dismiss them as generated boilerplate.

Running it as a governed project

I ran the build as separate AI-supported workstreams for project coordination, technical build, website copy, and legal and compliance, each with its own custom Claude Skill. Every issue passed a formal gate review before going live. That let one person cover work that would normally need a small team:

• Code. All of the HTML, CSS and front-end code, running as a flat HTML site on Cloudflare Pages.
• Policy content. Drafting and stress-testing each issue, verifying sources, and writing editorial and neutrality standards to keep the platform non-partisan.
• Legal and compliance. The privacy notice, a legitimate interests assessment, ICO registration and a data minimisation approach.
• Operations. The ongoing content pipeline, including monthly issue monitoring and updates.

I also rebuilt this website with Claude, moving it from a lapsed Squarespace subscription to a fast static site on Cloudflare Pages.

AI in finance work

I use AI for the slow, high-volume parts of finance work so more of my time goes on judgement. Over the past year I've embedded Claude into live fractional finance engagements. AI handles the volume; the positioning, assumptions and sign-off stay with me.

Sale preparation for a multi-entity group

For a founder-owned, multi-territory European construction group heading into an institutional sale process, I worked mainly with Claude and Claude for Excel to build a fully integrated three-year forecasting model covering P&L, balance sheet and cash flow, across nine legal entities in four currencies, on a compressed timetable alongside M&A advisors.

• Analysed twelve months of general ledger exports entity by entity, producing a bottom-up overhead baseline, identifying quality of earnings adjustments and surfacing classification questions for the finance team.
• Reconciled fixed asset registers across sources and diagnosed balance sheet imbalances in the model.
• Researched employer cost and corporate tax rules across seven European jurisdictions, feeding an entity-level tax module that replaced a single blended provision.

VAT compliance

For a growing UK service business, I used Claude to support a first VAT return including a six-month input VAT lookback: extracting and validating data from over 100 supplier invoices against set rules, deduplicating, reconciling to the bookkeeping records and building the return workpapers for MTD submission through bridging software.

On later quarters I used it to cross-check a team member's prepared returns against full transaction listings, identifying missed reverse charges on overseas software suppliers, incorrect VAT codes and input VAT that wasn't recoverable on the evidence held, before anything reached the client.

Advisory and modelling

I've used AI to pressure-test technical positions, including VAT liability under principal and introduction agency models, connected-party structuring, outside-scope income and director remuneration, and to model the pricing impact of different structures. I also built a monthly cash flow forecast that reconciled to historic data, so a non-financial client could judge whether a new commercial lease was affordable.

Systems and everyday finance

• Designed an automated flow between a booking platform, Stripe and the accounting software via Zapier, cutting manual invoicing and reconciliation.
• Turned a year of PDF statements into a structured spreadsheet for review.
• Supported year-end accounts and corporation tax preparation, and worked through Xero, VAT and pension questions.
• Developed a finance health check framework, a financial foundations assessment, client and partner agreements and recruitment briefs.

Automation and productivity

I design and run automations in N8N and Zapier, including a daily news briefing that uses the Claude API to filter and summarise news for founders, then reads it out through a text-to-speech API.

An AI grocery shop

I built an app in Lovable that analysed my historic grocery receipts to predict future purchasing patterns and draft a weekly shopping list. An agent I trained in ChatGPT then took that list, navigated the supermarket's website, looked for the best deal on each item (as defined by me) and filled the basket.

It worked, but I don't use it regularly. It's a little slow, and on balance the token usage and environmental impact aren't worth it for a weekly shop. Knowing when not to use AI matters as much as knowing how.

Beyond work

As a volunteer, I used AI to plan and run a primary school summer fair. That covered licensing research, a risk assessment, an accessibility guide for families who need a quieter experience, a donor and sponsor tracker, and costing and logistics tools.

Tools and skills

Skills I've developed

• Prompt engineering and context design, including writing and testing prompts for live use
• Designing AI workflows with human review built in
• AI project governance and quality control across multiple workstreams
• Workflow automation and API integration
• Building and deploying websites
• Data protection compliance for AI products
• Knowing where AI helps, and where professional judgement has to take over

Certifications

• Xero Advisor Certified (2025)
• The Ultimate MySQL Bootcamp (2021)