Work / 2025
Deal pricing agent
An internal AI agent that helps Commercial and Deal Desk teams structure complex deals without engineering access or tribal knowledge.
Context
Airwallex’s Commercial and Deal Desk teams were structuring complex multi-product deals — custom pricing, fee combinations, tiered discounts — for enterprise customers. The problem was that knowing what the platform could actually support required either tribal knowledge, direct access to engineering configs, or escalating to someone technical.
Every deal that needed a non-standard structure became a coordination cost. Sales cycles slowed. Engineering fielded questions that shouldn’t have needed engineering.
Problem
Commercial teams needed to validate deal structures — quickly, independently, and with confidence that what they were proposing was something the platform could actually execute.
The existing workaround was informal: ask someone who knew, check a config, hope the answer was current. It didn’t scale, and it created risk when answers were wrong.
What I shaped
A deal pricing agent built using Claude and an internal coding sandbox. The core design challenge wasn’t the AI — it was the knowledge packaging and interaction model.
Two things had to be true for the agent to be useful: the knowledge it drew from had to be accurate and verified (not approximated), and the agent had to gather enough intent from the user before surfacing what was possible. Giving someone a list of supported pricing models before understanding what they were trying to do was the wrong order of operations.
Key decisions and trade-offs
Intent gathering before constraint matching. The agent asks the user to describe the deal structure they want before it tells them what’s possible. This surfaces what the user actually needs rather than anchoring them to what’s easiest to support.
Verified knowledge only. The agent is grounded in documented, confirmed pricing capabilities. Hallucinating a possible deal structure that the platform can’t execute would be worse than no agent at all — it would erode trust in the tool and in the underlying platform.
Surface adjacent options when the exact match fails. A hard “no” is rarely useful. When the desired structure falls outside supported scope, the agent surfaces what comes closest and explains the difference. That’s more useful for a sales conversation.
Package knowledge for non-technical users. The agent’s audience doesn’t think in terms of platform configs or data models. The interaction has to translate between commercial intent and technical capability without exposing the underlying complexity.
What launched
MVP built and tested with Commercial and Deal Desk teams. Users can describe a deal structure in plain terms, and the agent guides them through intent clarification before matching to supported pricing models and surfacing adjacent options where needed.
Currently in testing — gathering feedback on where the intent-gathering flow breaks down and where the knowledge base has gaps.
What I’d do next
Connect to live contract and pricing configurations so the agent can check real-time availability rather than static documentation. The current version is only as current as the last knowledge update — live config access would eliminate that lag.