The challenge
Adding AI to a company is two problems, not one. The first is engineering: building agents that do useful work reliably against real systems, with guardrails, observability and a way to swap models. The second is adoption: most of the value is lost if only the AI team can use the tools.
The company needed both, on a platform serving more than 14 million users, without disrupting the teams that keep it running.
Approach
On the engineering side, agents were built with AWS Strands Agents and Anthropic's agent tooling, with LiteLLM as the routing layer so models can be changed without rewriting the agents. A multi-agent setup built on Claude splits larger tasks across specialised agents with clear hand-offs rather than one prompt trying to do everything.
On the adoption side, Amazon Quick Suite was introduced and rolled out team by team, with the AI team helping each group find the workflows where AI saved real time. The goal was not a demo but colleagues in ten-plus teams using the tools in their daily work.
Architecture
- Agents on AWS Strands and Claude, with LiteLLM providing a model-agnostic interface and cost tracking.
- A multi-agent orchestration pattern on Claude for tasks that need planning, execution and review as separate roles.
- Amazon Quick Suite for business teams, connected to the company's data.
- All of it on the existing AWS estate, with the same infrastructure-as-code and observability as the rest of the platform.
Results
Agents and the multi-agent setup are in production use, and AI adoption reached more than ten teams and over 250 colleagues.
This work was delivered by Ahsan Sheraz as part of his employer's AI team. It is included here because the same pattern, working agents plus practical adoption, is how ILMOTECH approaches AI integration for clients.
- 10+Teams using AI tools and agents
- 250+Colleagues enabled to work with AI