Implementing AI in Organisations

Building AI into organisational workflows requires more than selecting a model. The implementation layer — data pipelines, integration, evaluation, and change management — is where most projects succeed or fail.

The make vs. buy decision

Most organisations should start with commercial AI products (Microsoft Copilot, Google Workspace AI, Salesforce Einstein) before building custom solutions. Custom builds require technical capability, maintenance overhead, and ongoing model management that is not always justified. The exception: where proprietary data, regulatory requirements, or competitive differentiation make a custom build necessary.

The data readiness question

AI quality is bounded by data quality. Before implementing AI that depends on internal data, honestly assess: is the data clean, complete, current, and accessible? Projects often stall at data preparation rather than AI capability. Legacy systems, inconsistent formats, and access controls create more friction than the AI itself.

Change management is the hardest part

Research consistently shows that the biggest AI implementation challenge is not technical — it is adoption. People need to trust the system, understand its limitations, and have workflows that support appropriate oversight. Implementations that skimp on training and change management typically see low utilisation and poor outcomes.

Measuring outcomes, not outputs

Define success metrics before implementation and measure outcomes (time saved, error rate reduction, customer satisfaction improvement) rather than outputs (volume of AI-generated content). AI implementations frequently produce impressive output volume while failing to move the metrics that matter.

A common pattern: the AI demo works brilliantly. The production implementation underperforms. The gap is almost always data quality, integration complexity, and change management — not the AI itself.