Experiments, field notes and essays.
Everything published here, in the order it was written. For the experiment programme on its own, with each question and its result, see Research.
- Agentic Analytics20 min
What It Takes to Trust an AI Analyst
Lessons from six experiments on building reliable AI analytics: trusted data, governed definitions, a bounded verification harness, and evaluation that can see silent failure.
Read
- Agentic Analytics10 min
How an AI Analyst Knows What It Can Answer
A closed-world ontology turns answerability from a model judgement into a system check: governed, computable but undefined, or unsupported by the warehouse.
Read
- Agentic Analytics17 min
When Both Numbers Are Right
Product's governed active-user metric says 886. Platform's says 919. Both are correct. The experiment asks whether an AI analyst will expose the choice, and what it takes to make that reliable.
Read
- Agentic Analytics14 min
Why AI Analysts Pick the Wrong Metric
AI analysts fail before SQL generation when they choose the wrong metric, table, or column. Static checks can catch many of those wrong-choice traps before the agent runs.
Read
- Agentic Analytics25 min
The AI-Readiness Repair Matrix
Simple questions make AI analysts look ready. Real business questions combine several primitives, and small errors compound. From 1,488 graded runs, a repair method for AI-ready data: find the primitive that failed, find where its grounding lives, and move that grounding where the agent cannot skip it.
Read
- Agentic Analytics20 min
The Evidence Graph: Teaching an AI Analyst to Show Its Work
I tried to answer a simple product question: how can a user trust an AI analyst's answer? The answer was not more explanation. It was a typed graph of claims, query results, and support checks.
Read
- Agentic Analytics34 min
Agentic Analytics: Teaching an AI Analyst to Say I Don't Know
My last experiment pushed an AI analyst from 41% to 92% by adding structure. This time I measured what it does when there is no answer to give: structure alone still left it confidently wrong 31% of the time, and nine guardrails took that to 2%.
Read
- Agentic Analytics15 min
Agentic Analytics: How Much Does Grounding Actually Buy You?
Every vendor says their AI analyst works, and the benchmark numbers are high. So I built one in a controlled lab and added grounding one layer at a time (semantic layer, verified examples, knowledge base, metric tree) to get honest numbers on what each layer actually buys.
Read
- Decision Systems9 min
Working Backwards From the Customer Outcome
Amazon's Working Backwards is really a direction of travel: start from the outcome you care about, and let the metric tree tell you what to build.
Read
- Decision Culture9 min
Data Culture Is Really Decision Culture
Most companies chasing a “data culture” are chasing the wrong thing. What they actually want is a decision culture, and you can build one like a product, not a poster.
Read
- Data Platform16 min
Data Modelling in 2026
Kimball isn't dead, the modern stack isn't the answer, and 'AI-ready data' is a battle, not a checkbox. A working take on what to actually do.
Read
- Data Platform9 min
What I Look For in a Data Platform Audit
The fastest way to tell if a data platform is any good has nothing to do with the stack. It's whether two teams can agree on a single number.
Read
- Metric Systems12 min
How to Build a Metric Tree Leadership Actually Uses
A metric tree is not a list of KPIs. It is a model of how your business creates outcomes, and a map for where to look when a number moves.
Read
- Decision Systems9 min
Why Dashboards Don't Fix Decision-Making
A dashboard is the easy part. The layer that turns a number into a decision is the part nobody is assigned to build, which is why it's usually missing.
Read
