Bolt Food · Enterprise KAM Copilot

An agentic workflow on Claude: pull a chain's signals → diagnose vs central targets → grounded play lookup → JBP/QBR prep pack.
Interactive demo · prep packs were produced by the agent and are replayed here (no API key). The real version runs a Claude tool-using loop. Synthetic data.

1 · Pick an enterprise chain

How it works

1
Claude calls get_account_signals() to pull the chain's commercial signals.
2
It diagnoses the top issues vs the central targets (margin 22%, on-time 92%, share 50%), then calls lookup_play() — a grounded retrieval returning only the stored play + its PLAY id.
3
It writes a prep pack whose actions come only from retrieved plays (cited) — never invented — and a QBR opener.

The human sets the goal and owns the conversation; the agent does the prep.

Payoff: ~3–4 hrs of prep → ~2 min — the KAM spends the hour on the chain, toward the 80% commercial-time target.

2 · Agent run

Pick a chain and press Run — you'll see Claude choose its tools in sequence.

Chain signals (what the tool reads)

Select a chain to view its raw signals.

Commercial playbook (what the grounded lookup retrieves from)

Built by Soham Saha as a learn-by-doing prototype. Honest scope: this page replays real agent outputs without an API key; the live version is a Python Claude tool-using loop (grounded retrieval, escalate-don't-invent on a missing play, no automated eval yet) on synthetic data. It works the merchant/commercial side — it makes delivery reliability visible, it doesn't own logistics. Next: live tool-use against real commercial systems, a reviewer sub-agent, and a small eval set.