Improve the work first. Then teach it to AI.

Define the task, who runs it, frequency and the metric that will prove value.

Map inputs, sources, decisions, exceptions, checks and outputs.

Remove waste and standardise the best path before automating it.

Build instructions, examples, resources, quality criteria and edge cases with a narrow operating boundary.

Use native OpenAI/Anthropic access first. Add MCP only where a required internal system is not covered.

Test representative and difficult cases before rollout.

Create a Skill Access Matrix so each role receives only the capabilities it needs.

The client remains Owner. We document and transfer control, then stay for maintenance only if requested.