Skip to main content

Run a 10-minute demo

A presenter's script for a live walkthrough. It hits three beats in order: the wow (ask your data a question), the trust (rows never leave the org), and the automation (flows + pipelines). Timings are a guide — slow down on whichever beat your audience leans into.

Before you start (2 min, off-screen)
  • Open the hosted app and sign in; have the Dashboard showing.
  • Make sure a source is connected and its entities are confirmed (use the bundled Northwind demo, or a tenant with the SAP/warehouse model already loaded).
  • Have one API key ready (or be ready to create one live).
  • Optional: an MCP client (Claude desktop / web) already pointed at your server, to show the agent path.

0:00 — Frame it (1 min)​

[Show] the Dashboard.

AgentData dashboard

[Say] "This is AgentData — a semantic layer over your databases. It connected read-only, discovered your business entities — Customer, Order, Product — and now anyone, or any AI, can ask questions in plain language. Nothing was copied; it reads your data where it lives."

Point at the pipeline (Connect → Profile → Classify → Cluster → Emit → Review) and the counts: connected sources, entities pending, entities confirmed.

1:00 — The wow: ask a question (2 min)​

[Do] Go to Query → NL. Type:

top 5 products by revenue this quarter

[Show] the answer — rows and the generated SQL.

Query result

[Say] "No dashboard was built for this. It planned the query against the approved model, ran the SQL locally, and returned the answer. Let me ask a follow-up…"

[Do] Ask a follow-up ("…and just the ones in Europe") to show conversational memory. If your audience is non-English, ask the same question in their language.

3:00 — The model behind it (1.5 min)​

[Do] Open Entities, then open one entity, then the Model graph.

Entities

[Say] "The answer is governed because it comes from this — a reviewed model. Measures like revenue are defined once, here. You rename to business language, merge duplicates, and draw relationships to join across systems — SAP customers to warehouse events, for example."

Model graph

4:30 — The trust beat (2 min) — the one that closes regulated buyers​

[Say] "The obvious question: if an AI is querying our data, do our rows end up in a cloud model? No."

Only the model's names and the question reach the LLM. Result rows never do — the LLM plans, the SQL runs locally, rows return to the caller.

[Do] Open Admin → Security. Point at the toggles.

[Say] "Per tenant, you can block external LLM egress entirely and force an on-prem model, set an allow-list of permitted hosts, and give each API key least-privilege scopes. And Admin → MCP traffic shows every call and every failed-auth attempt."

(See Security for the detail if asked.)

6:30 — The agent path: MCP (1.5 min)​

[Do] Go to Query → Connect. Create or show an API key.

Connect over MCP

[Say] "Same model, opened to AI clients over MCP." If you have a client ready, switch to Claude/ChatGPT and ask it the same business question — it calls query_nl and answers from the governed model.

8:00 — The automation (1.5 min)​

[Say] "Getting data in is visual too — this replaces hand-coded ETL."

[Do] Open Flows → Pipelines and show a saved pipeline (or drag two flows and draw an arrow).

Pipeline DAG editor

[Say] "Flows extract and transform; a pipeline runs them as a dependency graph — parallel where it can, in order where it must. Schedule it, and watch every run here."

[Do] Open Monitoring to show live runs.

Monitoring

9:30 — Close (30 sec)​

[Say] "So: questions in plain language for people and AI, a governed model defined once, a hard privacy guarantee, and the loading behind it as visual flows. Where would you want to point this first — a database you have, or a CRM you want enriched?"

Then hand them a starting path:

If something fails​

  • A query returns nothing / errors — fall back to a known-good question on the demo model, or show a saved metric query (pick entity + measure + dimension) which doesn't depend on the LLM.
  • No internet for the LLM — demo the metric query path and the flows/monitoring beats; both work without an LLM call.
  • MCP client not set up — skip the live agent step; the Connect screen + the query_nl description make the point.
  • Keep it to 5 minutes — do beats 1:00 (wow), 4:30 (trust), and 8:00 (automation) only.