What we learned putting AI into operational change

You have watched it happen. A pilot demos beautifully in the boardroom, then meets real data and real handoffs and quietly never reaches production. It was not a weak model. The model was never the hard part. A durable AI build is a system built underneath the workflow: the data foundation, the redesigned process, and the measurement that runs for as long as the workflow does. That is where the advantage sits, and where most vendors spend the least.

Your project managers spend most of the week manufacturing artifacts about the work, not landing engagements. Here is the operating layer we built, and the boundary we hold.

An AI initiative has been parked on your desk somewhere. The CEO decided 'we need to do something about AI' and handed it to IT, the future Chief AI Officer hire, or the Digital Transformation portfolio. The initiative dies the same death every IT-sponsored change program dies. AI strategy and sponsorship belong to the CEO. Execution belongs to a senior executive accountable for delivery. The 12.2% AI adoption rate is not Canada's innovation report card. It is Canada's CEO report card.

Somewhere on your roadmap, an AI initiative has been parked for a quarter behind a six- or seven-figure data-modernization program. The pitch sounded right: fix the data first, then the AI. For a narrow set of workflows that pitch is correct. For most of the AI value an operator can capture this quarter, it is not. The right scope is the workflow, not the enterprise. Ask what data this workflow actually needs, and where it already lives.

Six months ago, "we run rigorous AI evaluations" was a defensible thing for an AI-services firm to say. Today every major enterprise platform ships built-in evaluators by default, and even custom evaluators only measure whether the model is behaving. None of them measures whether the workflow is still moving the KPI you bought it to move. That is the layer above evals, and it is the one most enterprise AI work skips.

AI-services homepages sell verbs because verbs commit to nothing. The test that filters them: name a workflow, a metric, an operator whose job changes.

In February 2025 we did exactly what we tell clients not to do. We automated our own content workflow with nobody left in it who could say no, and we shipped slop. The fix was not a better AI agent. It was a reporting line. The editor has a veto, and the person responsible for shipping cannot override it.