01
The prompt is a specification
It must be written precisely enough to cover exactly what is required. A vague prompt produces a vague deliverable — and costs more to correct than to redo.
Applied AI
AI does not rescue data that does not hold
We use it daily, and we know precisely what we use it for. Here is what works, what failed, and the order in which it must be approached.
Sequence matters more than the tool
On a five-level data maturity scale, AI sits at level 4. Most organisations we meet are at level 1 or 2. Deploying AI on ungoverned data does not produce intelligence: it produces confidently wrong answers. The sequence never changes — diagnose, connect, make reliable, then augment.
Level 1
Level 2
Level 3
Level 4
Level 5
Most often here
Level 1 – 2
AI becomes useful from here
Level 4
What we tried, and what we kept
01
2025 — Assistants and agents
Striking demonstrations, impossible to reproduce inside a corporate environment where every access is siloed. Kept: nothing, beyond an understanding of the underlying architectures.
02
Early 2026 — “Write me the formula”
A clear, documented failure: more time spent fixing what the AI produced than doing the work. Abandoned.
03
Mid-2026 — “Generate the artefact”
The turning point. AI produces the component, the mock-up, the data quality report — not a formula to copy. Measured on one engagement: one day for work estimated at a week.
04
Today
Dashboard mock-ups, data health reports written in plain language and ranked by priority, document summarisation. Always reviewed by a human before it reaches a client.
The two rules on which we do not compromise
02
Output is verified, always
The calculation remains a black box: nobody knows its inner workings. Anything touching a number is therefore reconciled against the source before it is presented.
What we do not do with AI
- Produce consulting deliverables with no substance: “anyone can generate content by the kilometre”.
- Have financial calculations written without reconciling them line by line against the source system.
- Promise AI to an organisation that cannot yet state reliably who its suppliers are or what its projects are.
- Conceal the use of AI from the client. We declare it, and we state what it is for: thinking faster, not padding out slide decks.
Contact