Back to blog

AI agents: measure cost per accepted case before Odoo

Created on November 5, 2026 · Publication analysed: November 4, 2026 · Source: Kong

Kong’s announcement invites Belgian and French companies to examine an agent workflow’s budget as a whole. Our proposal: evaluate the cost of each case accepted by business users and define stopping conditions before connecting automation to Odoo.

1. The official starting point

On November 4, 2026, Kong announced an expansion of AI Management in Konnect, including cost tracking and observability for AI interactions. The release describes attributing consumption to teams, users, applications, agents and business activities. It provides no economic results for an Odoo deployment.

2. What this changes in Belgium and France

We recommend setting a budget tied to a specific task: preparing a support response, matching documents or proposing a quotation. For an SME, start with one process and a budget owner. For mid-market companies and large groups, plan allocations across departments and a method for shared costs. For a public administration, retain the calculation and acceptance criteria so that the operating decision can be explained.

3. Underside analysis: count accepted cases

We propose adding inference, document retrieval, tool calls, infrastructure, supervision and human correction time, then dividing by the number of accepted cases. Keep failure and abandonment costs visible. Avoid using price per call as the sole decision criterion: measure retries too. Compare a simple baseline, potentially a business rule, with the agent workflow on the same set of cases.

4. Set limits that stop the workflow

Before testing, define a budget per case, a maximum step count and an execution deadline. Replay an unavailable tool, insufficient documentary evidence and a sequence of errors. Check that the agent routes work to a manual queue instead of retrying indefinitely. Test simultaneous cases: the owner should be able to suspend one queue without interrupting the entire operation. Document the expected behavior at each threshold, then measure the result.

5. Connect technical evidence to business outcomes

Assign a correlation identifier to each case and link spending to its validated outcome, without unnecessarily copying sensitive data into logs. For RAG, distinguish retrieval, generation and verification; for Odoo Enterprise, initially propose drafts in a test database. Replay a request after a timeout to check that it does not create duplicate writes. These steps are Underside architecture recommendations and do not presume an Odoo connector in the announcement.

6. Check sovereignty and reversibility

To compare local AI with a cloud offering, request telemetry flows, storage locations, support access and export terms. Test recovery with another model on the same cases and check that budget limits remain enforceable. Document data location, tool security and spending control separately. Select a solution only after business users validate its results, costs and return to manual processing.

Operational priority: choose one workflow, set its maximum cost per accepted case and test stopping on errors before allowing business writes.

Scope a business AI workflow

Read Kong’s official November 4 announcement