For people making technology decisions

Better software.
Better business?

I write about the decisions behind technology investment: what a company should build, where AI might change the work, and what evidence should reach the people accountable for the result. My starting point is the combination of running technology organizations, writing software, and researching language models.

This page connects those questions to the work already published here. The technical notes are evidence of methods and their limits, not proof of a return on investment in another industry. The business questions are the bridge: what would need to be true for the method to help your organization?

Discuss an engagement Professional record →

The editorial thread

The operating advantage

When software becomes easier to build, the strategic question shifts: which combination of customer access, workflow knowledge, useful data, and organizational capability can turn it into an advantage?

David Teece’s Profiting from technological innovation (1986) asks why an innovator does not necessarily capture the value of an innovation. One answer is complementary assets: the capabilities and routes to market needed to make the invention useful and commercially viable.

That is a useful lens for AI and more accessible software development. A company may be able to buy the same model or generate similar code as its competitors. The harder question is whether it can combine those tools with something customers value and competitors cannot readily obtain.

A working editorial thesis, informed by David Teece’s research on profiting from innovation and complementary assets. It is a question to test, not a claim that AI creates an automatic moat. The argument weakens where the complements are easy to copy, do not affect customer choice, or cost more to maintain than the value they create. Sometimes buying a standard solution is the right decision.

What to own →What to test →What to govern →

Three decision lenses

Start with what needs to change.

01 / Software as a source of differentiation

What should the business own?

If more software becomes practical to build, which customer experience or operating capability is worth making distinctive? Start with the workflow, its data, and who will maintain it, not a target for lines of code.

A useful test: Name the advantage, its owner, the buy-or-build alternative, and the ongoing cost of keeping it useful.

Three Emulator Bugs, Three Different Testsengineering note · Aug 2026
Test the State Space, Not the Noiseengineering note · Sep 2026

02 / AI economics and adoption

What would justify the investment?

A model benchmark is not a business case. Connect task quality, human review, integration, failure costs, and the ability to stop an experiment before deciding whether a smaller model or a new workflow helps.

A useful test: Specify the baseline, total workflow cost, acceptance test, and a reason to stop before the pilot starts.

Merging LoRA Removes a Branch, Not Parametersresearch note · Oct 2026
Perplexity Has a Tokenizer in Its Denominatorresearch note · Oct 2026

03 / Risk, governance, and accountability

What evidence reaches the decision-maker?

A persuasive report and a high evaluation score can conceal weak evidence. Make sources, uncertainty, decision rights, and escalation visible before an AI-assisted recommendation becomes an operating decision.

A useful test: Trace a material claim to its source and identify who can challenge, approve, or reverse the decision.

Deep Research Needs an Evidence Budgetapplied research note · Oct 2026
A Test Set Cannot Choose Its Own Thresholdresearch note · Oct 2026

A proposed decision loop: frame the decision, test the alternatives, review the evidence, then revisit the decision. Ownership, uncertainty, and a stopping rule remain explicit. — Working model A proposed working model, not a measured business result.

The industry changes the question.

These are questions to investigate, not claims of sector-specific results or client engagements.

Professional services: Which parts of analysis can be accelerated without losing the evidence trail or expert accountability?
Retail and distribution: Where would a distinctive planning or service workflow matter more than another generic interface?
Financial and regulated services: Which decisions need human approval, an inspectable record, and a reliable way to challenge the model?
Industrial operations: What happens when a recommendation meets incomplete data, physical constraints, or an operator's exception?