Agentic AI after the pilot: running AI workflows day to day
Client: A B2B SaaS company · Service operations
The problem
I built the company's service operations function from zero to 500+ members. AI-led automation worked in demos, but pilots kept stalling once they met real queues, handoffs and exceptions. The model was rarely the problem. What was missing was clear ownership of each output, a way to catch errors before customers saw them, and a way to measure whether anything had really improved.
The approach
- I started from the process, not the model: Lean Six Sigma baselines and explicit defect definitions decided which workflows to automate first.
- I let the language model interpret unstructured input, and let deterministic workflow tooling (n8n, Zapier) own state, transactions and retries.
- I built on the Claude API and Asana AI Studio, with retrieval-augmented generation (RAG) and MCP-enabled integrations connecting to the tools the team already used.
- Every automated workflow got an evaluation gate and a named human owner for exceptions before it went into daily use.
- I measured end-to-end cycle time and quality against the baseline, not the speed of individual steps.
The outcome
AI-led workflow automation delivered 300%+ productivity gains across a service operation that grew from zero to 500+ members, and the workflows ran in everyday use instead of staying stuck as pilots.