DEV Community

Cover image for AI agents in agribusiness: procurement and quoting connected to the ERP
Ronan Oliveira
Ronan Oliveira

Posted on

AI agents in agribusiness: procurement and quoting connected to the ERP

Procurement in agribusiness is short-window work. Inputs have to arrive before planting, freight shifts with the harvest, and the quote that was valid on Monday is gone by Thursday. Buyers spend the day between the ERP, the spreadsheet and the supplier's WhatsApp — and the bottleneck is rarely deciding; it's gathering the information to decide.

This post describes a published Zihin case in that setting and carefully separates what is a fact of the case from what is the solution pattern we recommend for procurement and quoting.

The case: Grupo Conceito and eBarn

In June 2026, Zihin published the numbers of a procurement and quoting operation in agribusiness. The agents work connected to the eBarn ERP and to TOTVS Protheus, used by Grupo Conceito and eBarn. The connection to Protheus goes through the MCP Bridge, in production — the moment when, in our timeline, the "legacy ERP became an AI tool".

What the case records:

  • Process: procurement and quoting.
  • Systems: TOTVS Protheus and the eBarn ERP.
  • Connection: MCP Bridge, in production.

We don't publish volume, success rate, financial savings or time-reduction figures for this case. Numbers only come back with a measurement period, a success definition, a source and the customer's authorization.

Note: Everything we state about the case is in this section: industry, supported processes (procurement and quoting), connected systems (eBarn and Protheus via MCP Bridge) and the fact that it runs in production. The following sections describe the typical architecture pattern for this kind of operation — not the customer's internal details.

Why procurement and quoting are AI's entry point in agribusiness

Not every process is a good candidate for an agent. Procurement and quoting have three characteristics that help:

  • Repetitive questions about data that already exists. "What's the balance of this input?", "how much did we pay on the last purchase?", "which orders are open with this supplier?". The answer is in the ERP; the cost is fetching it.
  • Decisions with clear rules, but frequent exceptions. There are approval limits, approved suppliers, reference prices. An agent can prepare the comparison; the final decision stays with a person.
  • Volume that grows during the harvest. The team doesn't double when demand doubles. An agent absorbs the mechanical part of the peak.

The obstacle is usually the same one the market reports: integration with the systems. A pilot chats well in the demo and gets stuck when it touches the real ERP. That's exactly where the case above stands out — the agents operate connected to the operation's systems, not to a data sample.

What the typical architecture for procurement agents looks like

What follows is the pattern we apply to procurement operations tied to ERPs like Protheus. Use it as a design reference, not as a description of Grupo Conceito's or eBarn's flow.

Reading the ERP with no credential in the agent

The agent doesn't get access to the ERP database. Zihin Replay keeps a read-only copy of the agreed tables in a Postgres database inside the customer's network; on top of that copy, the MCP Bridge publishes business queries with defined contracts — stock by item, open orders, purchase history by supplier. The connection between the customer's network and the platform goes through Zihin Tunnel: outbound only, over port 443, with no inbound port and no VPN. Credentials stay in the customer's network.

Quoting as a process, not a loose message

A typical quote involves building the demand from the purchase request, checking price history, organizing the proposals received and presenting the comparison. In this pattern, the agent does the preparation work and delivers a comparable view — the buyer decides.

Tip: The gain isn't in the agent "choosing the supplier". It's in the buyer opening the quote with history, balance and open orders already on the same screen, instead of assembling that across four systems. Automate the preparation; preserve the decision.

Human approval where money is involved

Reading is just reading. Any action with an effect — approving a quote, creating an order, changing a supplier — goes through a human decision step configured in the process. The deployment defines which actions require approval and who approves; the person reviews the context and records the decision before the process continues. With Zihin Chat, that approval happens right in the conversation, and one team member can approve what another triggered.

Policies the agent can't ignore

Contextual policies (CSP) are enforced on every execution: the hours the agent operates, fields it can't expose, the origin it can be triggered from. They add to the platform's RBAC and the credential vault described at /security. For procurement, the data policy tends to be the most relevant — negotiated prices and commercial terms are information not every channel should see.

Tracking the operation

Every execution is recorded, with its tool calls and policy decisions. That count, not an estimate, is what backs any metric the operation decides to publish. Zihin Watch monitors the operation and produces reports according to its configuration.

What the operations director should ask before starting

If your scenario is similar — consolidated ERP, procurement team under harvest pressure, scattered data — take four questions into the evaluation:

  1. Which queries does the buyer run against the ERP every day? They become the first MCP Bridge contracts.
  2. Which actions require approval, and from whom? That defines the process and the approval limits.
  3. Which data can't leave a given channel? That defines the data policies.
  4. Which ERP, and which access path? The Bridge is documented in production for Protheus; SAP and Oracle EBS are planned. The availability and deployment scope of Replay, Bridge and Tunnel depend on the infrastructure and the data agreement — check the FAQ and the sales team.

Conclusion

AI agents in agribusiness work when they reach the data the operation already has, under rules the company already follows. The Grupo Conceito and eBarn case shows this in production — Protheus and eBarn connected via MCP Bridge — and the architecture pattern described here is the path to repeat the design: reads through a read-only copy, quoting as a process, human approval where there's an effect, policies enforced on every execution.

Want to see what this looks like with your ERP and your procurement flow? Book a diagnostic.

Sources

  • Zihin.ai, home page — published agribusiness procurement and quoting case (June 2026; measurement period not disclosed): /
  • Zihin.ai, About — timeline, MCP Bridge in production with TOTVS Protheus at Grupo Conceito and eBarn: /about
  • Zihin.ai, Integrations — Replay, MCP Bridge and Tunnel: /integrations
  • Zihin.ai, Security — contextual policies, RBAC and credential vault: /security

Originally published in Portuguese on the Zihin blog. I work at Zihin, which connects AI agents to a company's systems and processes, with permissions, human approvals and cost tracking.

Top comments (0)