Thinking you can hold us to
Essays, technical notes and function deep-dives on decisions, prediction and proof. Every claim carries its number and its source.
An evaluation is not a pilot
A pilot proves software can run. An evaluation proves it changes a number you care about, on your data, against a baseline agreed before anyone saw a result. Eight weeks is enough, if the metric comes first.
One event backbone
A change happens, a KPI moves, a rule fires, an alert reaches an owner, a plan responds, an audit trail records all of it. In most enterprises those are six systems stitched together by exports. Lineage dies in the stitching.
Beyond enterprise search
Search finds documents. The job was always answers a person can act on and defend, computed inside their permissions, with the sources cited and the whole thing running where the data lives. That is a different machine from a search box, and from most of what is currently sold as one.
The day-100 AI bill
The first hundred days of an AI programme are paid for by enthusiasm. Day 100 is when the CFO asks what a decision costs and what it returns, and few have kept the books that answer either question.
Same inputs, same number, every time
A figure that changes on refresh cannot be reconciled, audited or defended. Determinism is a governance property, and it draws the line for where generative AI belongs in a decision stack and where it must not be.
Honesty is a feature
Every vendor tells you what their AI can do. The more valuable statement is what your data cannot support, said early and in writing. Honest limits are not a marketing sacrifice. They are the fastest route to being believed.
Trust is a record you publish
Executives say they trust AI, then re-check everything it produces. The behaviour is the honest signal, and it is rational. Systems that publish no track record have not earned the review step being removed.
The best model is not an opinion. It is a league table.
Which model should make this prediction is an empirical question with a daily answer, not an architectural belief. We run 62 models from 13 families against every question, re-score them continuously, and make every champion beat the naive baseline to keep its seat.
Success should never be a surprise
In most companies, whether the year worked or not is only discovered in December. The information that decided it was visible in March. Closing that nine-month gap is what planning was always supposed to be for.
Half the forecast is a person. Nobody scores that half.
Almost every enterprise forecast is adjusted by hand before it is used. The evidence says those adjustments help only about half the time. Yet organisations that measure model error to two decimal places do not measure the human half at all.
Twelve problems, not thirty modules
Enterprise software is organised around functions, so every function buys intelligence separately and starts from zero. The underlying problems, though, are shared, a dozen recurring computational shapes. Build for the problems once, and the second module costs less than the first.
Compliance should predict risk, not record failure
Two decades of compliance investment built superb systems for recording what went wrong. The record got better. Prevention barely moved. The future of compliance is better decisions, not better reporting.
Follow the thinking
New essays land on LinkedIn first. Follow 3RDi to catch them.