The nine questions
The same nine questions, answered twice
Numbers in brackets point to the numbered sources at the foot of the page. The line under each question says what a complete answer contains.
1 · Where must the data live for questions and agents to work?
A complete answer names the systems the data has to be copied into, and the systems that are read where they stand.
SnowflakeCoWork and Cortex Analyst work on data held inside Snowflake, and Snowflake's own documentation states that answers stay within its governance boundary. Data from other warehouses reaches CoWork once a person ingests it into Snowflake first, per the vendor pages reviewed [1][2].
Lima JunoYour rows stay in the systems that hold them. Juno registers Snowflake tables beside SAP, Oracle and planning tables and reads each one in place, joining all of them in one answer. 2 · How far does it reach into SAP, Oracle and planning systems?
A complete answer names the interface used for each system and says which of them are live today.
SnowflakeSnowflake's pages describe MCP connections reaching outside applications such as Gmail, Slack, Jira and Salesforce for agent actions, alongside a Business Data Cloud share arrangement with SAP described elsewhere in this research. A named, direct reach into SAP or Oracle transaction tables stays unstated on the Snowflake pages reviewed [3][4].
Lima JunoMore than 40 connectors cover SAP ECC and S/4HANA, Oracle Fusion and E-Business Suite, planning suites, warehouse systems, lakes, databases, files and REST services. One question can join several of them, and the one agent in production, the PO cancellation agent, runs on federated ERP data the same way. 3 · How is the shared meaning, the semantic model or ontology, built, and by whom?
A complete answer says what arrives ready, what your team builds, and how long that modelling work takes.
SnowflakeA semantic view is the required foundation, built by data teams through SQL, a YAML file, an AI-assisted wizard in Snowsight, or Semantic Studio, with import from Tableau and Power BI files. One partner warns that skipping this step still returns a confident answer, just the wrong one [5][6].
Lima JunoLima’s supply chain ontology arrives filled in: materials, plants, suppliers, orders and stock. Juno proposes a mapping for each registered column with a confidence and a reason; your team confirms, adjusts and extends it. 4 · How do business users ask questions, and is the working shown?
A complete answer shows the question, the query that ran, the sources read and the access rules applied.
SnowflakePeople ask CoWork a question in plain language; behind it, an agent calls Cortex Analyst for a database query over the semantic view, Cortex Search for documents, or a custom tool. Answers trace back to the source data and the query, and Deep Research adds a cited written report [1][7].
Lima JunoPeople ask ChatLima in plain language. Each answer shows its working: the ontology terms used, the systems read, the seconds taken and the query itself, under the asking person’s row and column access. 5 · How are agents and rules defined, and by whom?
A complete answer says who writes a rule, in which language, and how it is reviewed before it runs.
SnowflakeCortex Analyst is grounded entirely in the semantic view a data team builds and maintains, with verified query pairs and custom instructions steering its answers; practitioners keep the underlying YAML file in version control and re-test it after every change [5][8].
Lima JunoA planner writes the policy in plain language, and the Planner Agent turns it into a reviewable planning table of tasks, data and rules. The PO cancellation agent, built this way, runs in production; Lima’s agent blueprint designs 960 further roles across five chains, the other supply chain task agents among them, yet to be built. 6 · How is autonomy set, and how do approvals work?
A complete answer gives the levels available, who approves what at each one, and how a level is raised and lowered.
SnowflakeSnowflake's pages describe role-based access, row-access policies and masking carried automatically into every answer; a graded, numbered autonomy setting for agent actions stays unstated on the pages reviewed, and Skills and Memory were heading to public preview in June 2026 [2][9].
Lima JunoEach agent carries a level you set, A0 to A4. At A2 a named person approves every action; at A3 the agent acts inside value, budget and confidence limits and sends exceptions to people. The level is one of eight settings on an agent, each set by a named person. 7 · How do actions reach the ERP, and is each write confirmed?
A complete answer names the interface for each ERP and says where the acknowledgement of a write is kept.
SnowflakeCustom tools run stored procedures inside Snowflake, and MCP connectors carry actions out to Gmail, Slack, Jira and Salesforce; whether an outside ERP such as SAP or Oracle receives a write, with an acknowledgement recorded, stays unstated on the pages reviewed [3][9].
Lima JunoWrite-back uses each ERP’s own supported interfaces: released APIs and BAPIs for SAP, REST resources and interface tables for Oracle. It starts per application after a confirmed test write, and every write keeps the acknowledgement. The PO cancellation agent in production writes back this way, under human approval. 8 · How does it learn from people’s decisions, and what is audited?
A complete answer says what is recorded for each decision and how that record shapes the next recommendation.
SnowflakeVerified query pairs and custom instructions in the semantic view are Snowflake's documented feedback path, refined by data teams over time; a record tying an individual person's decision to a later outcome, and an audit trail for agent actions, stays unstated on the pages reviewed [5][8].
Lima JunoEvery accept, edit and override is stored with the rule version, the facts at that moment and the outcome, so the next recommendation improves. The audit trail holds signal, reasoning, limits, approval, action and acknowledgement for each run; in production, 100% of the cancellation agent’s actions trace to a rule. 9 · What is the pricing unit, is there a trial, and where do analysts place it?
A complete answer names the unit of charge, the route to try it, and the analyst reports that cover the category.
SnowflakeCortex Analyst costs 6.7 credits per 100 messages plus warehouse compute, and CoWork has billed in AI Credits at $2.00 or $2.20 each since April 2026, with a 30-day trial worth $400 of credits. Snowflake's place in the 2026 Gartner Analytics and Business Intelligence quadrant stays unstated in the sources checked [10][11].
Lima JunoJuno is licensed with the Lima platform and priced per engagement. The token manager budgets and meters every model call, per agent and per team. Twenty use cases are worked through on this site, one page each, and the analyst guide sets out the markets these products sit in. 