The design document we never typed: how AI automated the monthly financial review of a major capital programme

A new era of consultancy has arrived, powered by AI. On our latest engagement, the best design document we produced was never typed, it was spoken.

We have just delivered a proof of value for a client operating in a complex, regulated construction environment: an agent that automates the monthly financial review of a major capital programme. Hundreds of live projects, a manual review taking days, and the projects that genuinely needed attention buried deep in spreadsheets. 

Here is the part worth sharing. Every discovery session with the finance team was recorded. Those recordings and transcripts became the raw material for the entire solution. 

Feeding them into AI, we were able to: 

• Extract the real requirements, including the unwritten rules experienced reviewers apply instinctively 

• Codify the domain language into a glossary the agent could reason with 

• Turn "here is what I look for each month" into a behavioural instruction set with 14 early warning rules and explicit thresholds 

• Profile the actual ERP extracts against those rules before writing a line of production code 

• 
  

From there we built the full solution: parsers, a rules engine, an orchestrator, regression tests with CI, and an implementation guide covering architecture, phased rollout and governance. 

Two lessons stood out. 

1. Real data is undefeated. Our first live run against actual monthly extracts exposed two bugs within minutes, including a subtle contamination issue where a file from a different programme slipped into the history. Synthetic test data would never have caught it. File provenance matters as much as file format.

   
   

2. Scope is a feature. We deliberately chose one programme, the one causing the most operational pain, rather than trying to boil the ocean. That focus is why we went from discovery conversations to a working agent in weeks, not months. 

   
   

The end result: an AI agent that receives formatted JSON from pre-processed Excel extracts and produces a RAG report (red, amber, green) flagging early warnings across the entire programme. Instead of trawling every project, finance leadership open a prioritised exception report that tells them exactly where to look and why, anchored to verified baseline figures they can trust. Catching margin erosion, missing forecasts and provision risks months earlier has the potential to save millions. 

Business first, technology second. The code was the easy part. Listening properly, then letting AI translate that listening into working software, was where the value came from. 

If your organisation has a high value review process running on institutional knowledge and spreadsheets, this pattern is repeatable.

Talk to Neologik about what this pattern could look like in your organisation. One agent at a time.