From pilot to production: the part most AI programmes skip
A pilot that works in a demo is not yet a programme. I deliver AI and RPA work from business case to adoption, with the ownership, controls and cost model that let it survive production.
Five questions I settle before scale
- 1 What number must it move, and what is the baseline?
- 2 Who owns it after go-live?
- 3 What does a wrong answer cost, and who reviews it?
- 4 Is the data clean before the model sees it?
- 5 What does it cost at ten times the volume?
Where I take over
Business case, baseline and funding approval. Use-case selection and pilot design. Governance with risk, compliance and legal: data masking before any model sees it, confidence thresholds and limits that decide when a human approves. Shadow mode first, then low-risk cases, then full volume. Change management through a champion network, training and adoption KPIs reported to executives. Then the hand-off to a business owner. I lead delivery; specialist build partners can be brought in for the engineering.
Evidence
- Agentic AI for billing and collections at a large national telecom: three coordinating agents (classify, validate, respond) on a cloud data platform, answering from live account data rather than model memory, and routing disputes above a set limit to a person. Over 20 use cases, from simple questions to credit requests.
- Results: CSAT up 15 to 25 per cent, manual email handling down 60 per cent, and a seven-figure annual benefit with payback inside the first year of operation.
- RPA inside ERP billing modules: file synchronisation, multi-point entry validation and invoice generation. Cycle time down 40 to 60 per cent, with 100 per cent data integrity proven in parallel runs before any manual step was retired.
The full story, on my personal site: AI/ML and RPA case study ↗ Article: scaling AI pilots into production ↗
NDA first, on request. I take a small number of clients at a time.