How I use AI in finance
I use AI at three levels in the monthly FP&A cycle.
One rule for all three levelsCode calculatesAI explainsA person signs off
LEVEL 1
Single tasks
- Drafting, research, reviewing my own logic
- Comparing data sets
- Excel formulas and modelling, DAX and Python, tested against known numbers
LEVEL 2
Workflow automation
- CRM pipeline to revenue forecast, refreshed daily
- Rules do the maths; AI flags anomalies and drafts commentary
- Automated checks on mappings, duplicates and reconciliation
LEVEL 3
Assistants & Agents
- Finance agent: compares versions, drafts commentary and the management report Live
- Knowledge assistant with cited sources Live
- Multi-agent close-to-report team In build
Tools and models
| AI models | Claude, on Team plan: Opus for day-to-day work and the standard monthly runs, Fable for complex, multi-step analysis. I pick the model that is good enough for each step, not the largest one everywhere. |
|---|---|
| Building | Claude Code for writing and reviewing code. The workflows call the Claude API with a key that IT manages. |
| Orchestration | n8n: triggers, connectors and the visible flow from step to step. |
| Logic | Python (pandas) for validation, reconciliation and the forecast rules. SQL for fixed, approved data queries. |
| Knowledge assistant | Embeddings stored in Postgres with pgvector; answers cite their sources. |
Skills and techniques
What I use today
- Python
- LLM prompting and workflow design
- AI and machine learning fundamentals: how models learn, embeddings, transformers and attention, encoder and decoder, tokens and context windows
- Model Context Protocol (MCP)
- Structured outputs
What I'm still developing
- Retrieval-Augmented Generation (RAG) and vector databases
- Agent orchestration and state management
- LangGraph
→ next · ← back · M switches between "What I built" and "At Medartis"
I wouldn't start with AI. Prioritise first, then four steps that each pay for themselves.
The same sequence works for any finance process.
↺ Then the cycle starts again with the next process. What scaled becomes the new standard.
How I followed these steps at my current company
| 0 · Prioritise | About a year ago, when we started this journey with AI, we prioritised two areas: headcount planning, because personnel is the largest part of our cost base, and the CRM pipeline for new clients, because new business is our key revenue metric. |
|---|---|
| 1 · Standardise | I reviewed our processes, guidelines, financial model and headcount planning with Claude, and rebuilt them in a clear structure that both people and AI can read. With that structure in place, the AI understood the context and its results became noticeably better. |
| 2 · Right tool | Headcount planning runs on rules and checks, linked directly from the HR system to our FTE report. No AI. |
| 3 · AI where needed | The CRM-to-forecast workflow and the finance agent, where AI flags, explains and drafts. |
| 4 · Scale | The next step: we are planning to extend this to more processes. |
AI is the tip of the iceberg. What makes it work is underneath.
The model is 20%. The other 80% is underneath: making it safe and useful in a finance process.
That work is a team effort, with Reporting, IT and the data owners. It's the part I'd want to build with you.
Eight rules for rolling out AI in finance
- Numbers from systems, never from the model. Code calculates; AI explains.
- Standardise before you automate. An agent on messy data only produces faster confusion.
- Fixed rules wherever possible. Consolidation, FX and allocations run in code; AI only does judgement steps.
- A check after every step. Totals reconcile with the source before the next step runs.
- Everything traceable. Every number links to its source; every step is logged for audit.
- Least access. Approved tools only, access to only what's needed, synthetic data for prototypes, IT and legal involved before real data.
- A named person signs off. AI never posts entries, approves payments or sends reports on its own.
- Start small, test and measure. One workflow, a fixed test set before go-live, hours saved and errors caught, then expand.
Data protection
- Approved tools only, under a business agreement with no training on company data
- Read-only access, limited to what each workflow needs
- Pre-close figures of a listed company stay inside approved systems
- Start with use cases that need no sensitive data; demos and tests run on synthetic data
How I test it
- Hard checks, pass or fail: numbers reconcile, drivers exist, format valid
- A test case per rule: missing field, wrong currency, oversized deal
- Past months as a reference set, re-run after every prompt or model change
- Every failure becomes a new rule