MMustapha El HajjCxO · Builder

Writing

Ideas, written down.

Energy and electrification, buildings and connected products, software, commercial growth, venture building and markets. I write these while I am still doing the work, which is why some of them change their mind halfway through.

Mustapha El Hajj writing

Newsletter

Working notes, straight to your inbox.

Occasional, no-noise notes from inside the work: what shipped, what broke, and what I would do differently.

Articles

All topics Energy, Electrification & Power Systems Buildings, IoT & Connected Products Strategy, Commercial & Growth Software, Product & Platform Engineering Operating Discipline & Turnaround Fintech, Crypto Rails & Trading Systems

#AI governance #governance #operating model #procurement #applied ai #agentic ai #risk management #workflow design #operations #risk controls

A quiet data center aisle beside a workstation showing an empty compute job queue.

31 Aug 20267 min read

Software, Product & Platform Engineering

Sovereign Compute Isn’t a Purchase. It’s an Operating Model.

Sovereign compute fails the same way many “strategic” platforms fail: you buy capacity before you can create demand, govern usage, and exit bad vendor paths. Denmark’s idle supercomputer is a case study in why operating model beats procurement theater.

Sketchnote-style path showing a workflow with handoffs, an owner, exception handling, and a stop button.

10 Jul 20268 min read

Software, Product & Platform Engineering

AI Leadership Is Workflow Leadership: Own the Handoffs

AI creates value when a real team owns the full workflow: inputs, approvals, exceptions, and failure handling. If nobody owns the handoffs, you do not have an AI system. You have a demo.

A single treasury operator follows a winding path from an unsafe AI autopilot conveyor toward a near-miss, hits a kill switch to stop it, then rebuilds the same line with gated steps and a replay log so automation runs safely within a marked safe range.

29 Jun 20267 min read

Software, Product & Platform Engineering

Agentic Treasury Needs a Kill Switch: Design Autopilot Like a Plant

Agentic AI in treasury will only be safe once we treat it like a production line, not a chatbot. The winning pattern is simple: gated autonomy, deterministic replay, and a kill switch that actually works under stress.

A single operator follows one winding workflow line from a flashy AI demo that causes spills and complaints to a bolted-in AI station with owned handoffs, an error budget, fast escalation, and a final board showing improved real outcomes.

20 May 20267 min read

Software, Product & Platform Engineering

AI Adoption Is Mostly Workflow Ownership, Not Models

Most “AI adoption” programs fail because they treat AI like a demo, not like a process. The operator move is to define handoffs, error budgets, and escalation paths, then measure outcomes in cycle time, leakage, and complaint rate.

A single stick-figure manager walks a winding path from celebrating a big “HOURS SAVED” stopwatch, through a clogged approval bottleneck and rising risk, to adopting a control checklist focused on cycle time, errors, leakage, and margin, ending with smoother flow and higher margin.

29 Apr 20267 min read

Software, Product & Platform Engineering

Stop Measuring AI by Hours Saved

Hours saved is a comfort metric. If you want AI to move the P&L without creating new operational risk, measure cycle time, error rate, leakage, and margin per constrained headcount, then change the operating model to match.