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.
Enterprise software has about twelve problems
We claim that a dozen problem classes do almost all the real work in enterprise software. A count that provocative earns the obvious question. Which ones? Here is the list, what each class answers, and the test we invite you to run against it.
Nervousness is a property you can measure
Continuous planning cured stale plans and bought a new disease, churn. A schedule changes every time it can, and imposes a cost every time it does, and no planning vendor publishes an instability metric. Define it, measure it, manage it.
98% accurate deserves a methods section
A forecast accuracy claim without a method is a marketing sentence. Revenue technology's boldest number, 98% accuracy by week two, is missing the four things every accuracy claim owes its reader.
Permission-aware is the hard part
Enterprise search vendors publish hours saved. None publishes how often an answer leaks content its reader was never allowed to see. That second number decides whether the category is safe to buy.
Sampling was a concession, not a method
Audit sampled because looking at everything was impossible. The constraint died and the habit survived. Once the whole population can be tested continuously, the sample stops being rigour and becomes a choice that needs justifying.
Supplier failure has a lead time
Suppliers rarely fail suddenly. They fail observably, then officially. The product worth buying is not a risk score; it is lead time. How many weeks of warning does the deterioration give you, and what is each week worth?
The accuracy gains that never reach the P&L
Forecast accuracy genuinely improved, and the P&L did not notice. The gain leaks downstream of the model, in safety stocks nobody retuned, plans nobody trusted and schedules frozen against a forecast nobody used.
The agent question is an exception question
The industry is busy restraining agents. The harder question is older. Of the million actions machines will take on your behalf, which exceptions deserve a human minute, and who decides?
The AI Act deferred the hard part
Brussels moved the high-risk deadlines to late 2027 and 2028. It did not move the burden of proof. A model that will face scrutiny then needs a track record that starts now, because you cannot backfill evidence.
The backtest includes the misses
Any vendor can show you a backtest. Ask to see the misses; whether they appear is what separates evidence from marketing.
The ECCP grades your algorithms
Since September 2024, US prosecutors evaluating a compliance programme are directed to assess how the company governs its own AI. The tool you bought to manage compliance risk is itself in scope, and "we use AI for compliance" became a claim you must be able to defend.
The maturity ladder ends at prevention
Most enterprise software climbs three rungs, record, track, report, and calls the ladder complete. The ladder has eight rungs. The upper five, predict, detect, explain, intervene, prevent, are where outcomes change, and they demand a discipline the lower three never needed.
Follow the thinking
New essays land on LinkedIn first. Follow 3RDi to catch them.