From model calls to governed work: the first DataInbox capability review
What our updated Runtime and AI cost explanations mean for teams: reusable capabilities, accountable actions and costs that can be tested in a pilot.
DataInbox
Architecture & Research
This first review covers improvements to existing DataInbox explanations since September 29, 2026. A common business question runs through them: when more AI tools can reach company systems, who owns the current facts, the right to act and the evidence of what happened? The answer matters more than the number of model calls an agent can make.
We clarified the proposed Companion boundary in the Agent Runtime and existing analysis of reusable business logic. We also rebuilt the AI Cost Benchmark Lab around questions, workflows and agent tasks, and made the AI Economics Dashboard distinguish recent provider prices from older stored DataInbox scenarios. These are website and calculation improvements; they do not announce a new DataInbox Runtime release.
Read the SAP Companion analysis
Explore the AI Economics Dashboard
Reuse the capability, keep the decision owned
Google made its Data Agent Kit generally available on September 30. It gives coding agents MCP tools and skills for working with Google Cloud data products. Microsoft says its Dynamics 365 ERP MCP server and Copilot Cowork ERP plugin are generally available, while its proposed reusable ERP business skills and extra agent restrictions have different future availability. SAP's September CAP release made MCP adapters for custom application services generally available; its CAP agents have less mature Gamma and Alpha labels. These releases show increasingly practical ways for agents to discover and invoke existing work. They do not give DataInbox those integrations automatically.
Google Cloud's Data Agent Kit announcement
Microsoft's Dynamics 365 ERP announcement
SAP CAP September release notes
Our architectural inference: a DataInbox Companion Registry could describe which provider capability is eligible for a particular business request. The business operation would still need current state, a rule owner, scoped authority, approval where required and a recorded result. An MCP tool listing describes what a provider offers; it cannot itself decide what an employee or agent should be allowed to do.
Consider an order-status request. Today a support colleague may ask operations to look up an order, copy the answer into a customer message and later search several systems to explain it. In a proposed Companion pilot, DataInbox would resolve the permitted current order context, select an eligible read capability, and keep the answer, source and decision together. If the next step changes an order, an authorised person would review the exact proposed action and the system would recheck state before execution. This could reduce repeated integration work, handoffs and later reconstruction. It depends on reliable source data, connector maintenance and a real approval path; it is not a measured saving.
Telnyx's October 1 Agent Memory API beta gives another useful boundary. Remembered conversation facts may help an agent continue a discussion, but they are not the live order state or permission to change it. Our Runtime explanation now makes that separation explicit.
Telnyx Agent Memory beta release note
Count the complete cost of a useful outcome
The cost lab now starts with volumes a team can estimate: answered questions, workflow requests and agent tasks. It shows an illustrative model-token bill and historical price points, then separates possible answer reuse from a prompt-cache discount. An OpenAI or Anthropic subscription is a personal workspace; DataInbox's published plans cover shared business operations with provider consumption charged separately. The comparison keeps retained accounts and implementation costs visible. Replacing seats is a contractual choice, not an automatic saving.
Our October 7 check found no change in the eleven model rates used by the calculator. The EUR conversion now uses the European Central Bank's October 6 reference rate of EUR 1 = USD 1.1269. That changes displayed euro estimates, not the providers' USD tariffs. The dashboard's stored DataInbox Runtime team projections were last calculated on August 24 and have not been rerun; a current price check is not a new Runtime measurement. Tool calls, cache writes and storage, grounding, regional or faster processing, tax, human review and errors can add costs outside the token estimate. Equal token counts also say nothing about equal model quality.
Test one complete operation
For a first pilot, choose one owner and one bounded request, such as confirming an order status and proposing an approved correction. Compare the old and proposed paths on time to a complete answer, manual handoffs, corrections, approval wait, decision-history retrieval and total cost per successful outcome. Keep quality and permission failures in the measure. That evidence can show whether reuse and governance improve the operation for that team; provider releases and lower token prices alone cannot.