AI workflows / how the desk works
Agents do the reading. People make the call.
AgriAI Expert runs AI agents inside the sourcing desk: they structure the brief, put offers on one basis, check documents against the request and keep follow-up from slipping. A person reviews everything before it reaches a buyer or partner.
What the agent does, and what a person checks
- Brief intake — Agent: Reads the request as it arrives — email, PDF, spreadsheet or buyer form — and pulls out product, cut or grade, packing, volume, destination, Incoterm, documents and timing into one structured brief. Person: Checks the brief against the original before any partner is contacted, and goes back to the buyer on anything missing or ambiguous.
- Offer comparison — Agent: Puts partner offers on one basis — same specification, pack, Incoterm and document set — and marks every line where an offer does not match the brief. Person: Confirms the comparison is like for like and decides which routes are worth taking further. Commercial terms are never set by the agent.
- Document review — Agent: Reads specification sheets, COA references, labels and buyer forms against the brief and the destination checklist, and lists the mismatches. Person: Reviews every flagged item. Certificates and approvals are confirmed with the issuer per route; the agent does not certify anything.
- Market and buyer research — Agent: Gathers public information — destination requirements, trade-fair exhibitor lists, company profiles — and keeps the source link next to every finding. Person: Checks each figure against its source before it is used. Contact details that cannot be verified are not used.
- Follow-up — Agent: Tracks open questions, samples, documents and deadlines across email threads, and drafts the next message. Person: Reads, edits and sends every message. Nothing goes to a buyer or partner without a person approving it.
What the agents are not allowed to do
- No prices or terms — Prices, payment terms and commitments are decided by people, per project.
- No unsupervised messages — Agents draft. A person reviews and sends every message to a buyer or partner.
- No compliance verdicts — Agents flag gaps. Certification and regulatory verification stay with accredited bodies and the authorities.
- Confidential files stay in the project — Buyer and partner documents are used for the request they belong to and are not published.
What it runs on
- Claude models by Anthropic — The language models the agents run on.
- Model Context Protocol (MCP) — The open standard that connects the agents to email, documents and research tools.