Skip to content
CSA Loom — the Microsoft Fabric experience for Azure tenants where Fabric isn't yet available: lakehouses, warehouses, notebooks, semantic models, Activator rules, Data Agents, across Commercial, GCC, GCC-High, and DoD IL5

Day 4 — BI & AI & Direct Lake & Data Agents (Federal CoE)

Track: 5-Day Federal CoE Workshop · Day 4 of 5 · BI & AI & Direct Lake & Data Agents

Day 4 turns Gold tables into business value: a semantic model, a refreshable report, a Direct-Lake-Shim demonstration with an honest latency discussion, and a Loom Data Agent answering natural-language questions under Gov constraints.

Azure-native by default

Semantic models use the Loom-native tabular layer over the warehouse/ lakehouse; Data Agents run on Azure OpenAI Gov endpoints with identity passthrough. No Power BI / Fabric workspace, no Foundry Agent Service (it is Forecasted in Gov). LOOM_DEFAULT_FABRIC_WORKSPACE unset.

Learning objectives

  1. Author a semantic model over Gold and a report on top of it.
  2. Configure the Direct-Lake-Shim and reason honestly about refresh latency.
  3. Author and test a Loom Data Agent grounded on the lakehouse.
  4. Design an Activator rule that fires into Teams.
  5. Use Loom Copilot to accelerate item creation.

Facilitator guide

Timing (8-hour day)

Time Activity Mode
09:00 Day-3 recap + semantic-layer architecture Lecture
09:30 Author a semantic model over gold.device_hourly Lab
10:30 Break —
10:45 Direct-Lake-Shim config + honest latency demo Lab + Lecture
11:45 Lunch —
12:45 Data Agent authoring + test chat Lab
14:00 Activator rule → Teams Lab
15:00 Break —
15:15 Loom Copilot build-assist tour Lab
16:15 Wrap-up + Day-5 readout prep Plenary

Talking points

  • Semantic layer without Power BI F-SKU: in Gov there is no Fabric/Power BI F-SKU. The Loom-native tabular layer provides measures + relationships over Gold; reports render natively. A P-SKU Power BI is optional, not required.
  • Direct Lake honesty (read this verbatim): Fabric's Direct Lake refresh is framing — a low-cost metadata operation that takes a few seconds and is on by default. The Direct-Lake-Shim reproduces the warm-cache pattern on Azure-native: it programmatically reframes the cache on Event Grid signals. Be honest that the shim's freshness window is seconds to tens of seconds, not a sub-second miracle, and that Direct Lake on OneLake (the Fabric original) is a forward-migration capability, not something Loom claims to out-perform.
  • Sovereign agents: Foundry Agent Service is Forecasted in Gov, so server- side thread persistence and the Agents playground are unavailable. Loom Data Agents run on AOAI Gov with a manual SOC pipeline. See the Sovereign AI Agents use case.

Exercises

  1. Group predicts the shim refresh latency for a Gold update, then measures it live and compares — the point is calibrated honesty, not a number to beat.
  2. Each participant writes one Data Agent grounding instruction that prevents the agent from answering outside its authorized tables.

Common pitfalls

  • AOAI Gov deployment not provisioned → the Data Agent pane names the env var (LOOM_OPENAI_*) and the deployment to create. Provision and retry.
  • Over-promising Direct Lake parity — keep the latency discussion honest.

Participant lab — BI + AI

  1. Semantic model. In Semantic model (/semantic-model), build a model over gold.device_hourly: define a measure (avg reading), a device dimension, and a date relationship. Save.
  2. Report. Create a report with a timechart + a top-N table over the model. Confirm it renders against the live warehouse/lakehouse backend.
  3. Direct-Lake-Shim. Configure the shim for the Gold table, push an update to Gold, and observe the reframe + measured refresh latency. Record the number for the Day-5 readout.
  4. Data Agent. In Data agent (/data-agent), create an agent grounded on the IoT lakehouse tables. Ask: "Which device had the highest average reading in the last 24 hours?" Confirm a grounded answer with a citation.
  5. Activator rule. In Activator (/activator), create a rule that fires a Teams message when a device's hourly average exceeds a threshold. Trigger it with a synthetic spike.
  6. Copilot. Use Copilot (/copilot) build-assist to scaffold a second measure; review and accept the suggestion.

Validation (Day-4 done): semantic model + report render on real data, the shim reframes with a measured latency, the Data Agent returns a grounded answer, and the Activator rule posts to Teams.

Datasets

Homework

  • Prepare your Day-5 readout slides: workload migrated, shim latency observed, agent demo, and one open risk.

Federal-specific emphasis

  • No Foundry Agent Service / Agents playground in Gov — covered honestly; Loom Data Agents are the Azure-native 1:1 on AOAI Gov.
  • Defender for AI is Commercial-only; in Gov use the Sentinel pipeline workaround (PRP-13 ai-defense.bicep).
  • Content Safety at IL4+ may require self-hosted Presidio — note for the SSP.

Slide deck

make loom-decks DECK=docs/fiab/workshops/5-day-federal-coe/day-4-bi-ai.md.