Operator & Model Harness
What Switchboard's AI can do, and the harness that keeps its answers accurate.
Operator is the AI assistant built into Switchboard. It takes a request in plain English and carries it out: running the analysis, building the document, writing the file, or sending the report.
It differs from a general-purpose chatbot in three ways: it reads your models rather than pasted data, it can execute queries and code, and its ability to act outside Switchboard is gated on your approval.
The model harness
The model harness is the structure around the language model: what it is given, what it is allowed to run, and where it has to stop. The model itself is general-purpose; the harness is what makes its output specific to your data and checkable afterwards.
Context. Operator works from the semantic layer: the properties, aggregations, and relationships that define your metrics. It does not have to infer a metric definition from column names, because the definition is recorded in the model.
It also has access to your documents, and it retains context from earlier conversations — metric definitions you have corrected, preferences about how you want output formatted, and processes you have walked it through.
Planning. For anything beyond a single query, Operator states which models it will use and what it will compute before running anything. The plan is visible, which is usually where an unexpected result gets explained.
Tools. Operator retrieves values rather than recalling them. It has:
- SQL against your models, returning tables and charts that can be downloaded.
- Python — pandas, numpy, scipy, scikit-learn — for statistics, cohort analysis, forecasting, and anomaly detection.
- Build operations for creating and editing models, documents, and blocks.
- Inspection, reading back exact values and rendering charts to check what it produced.
Approval. Analysis inside a thread runs without interruption. Anything that leaves Switchboard — a Slack message, an email, a scheduled job — requires explicit approval first.
What it is used for
Analysis. Querying any model, profiling columns, checking data quality, reconciling two sources that disagree, and running statistical analysis in Python.
Model work. Creating a model from SQL or an uploaded spreadsheet, editing an existing one, adding aggregations, relationships, or AI-enriched properties, tracing lineage, and diagnosing failed builds.
Documents. Building dashboards from a description — charts, scorecards, pivots, filters — and editing existing blocks.
Delivery. Posting to Slack, sending email, exporting multi-tab Excel workbooks, and scheduling any of the above to recur.
Limits
Operator works from what is modeled. If a source is not connected, it reports that rather than estimating. If a metric is not defined, it proposes a definition rather than assuming one.
When it is wrong, the plan, the SQL, and the model definitions are all visible, so the error can be located.
Access from Claude
The same data access is available in Claude through an MCP connector. See Connect Claude (MCP).