Connect BigQuery read-only
A project ID, a dataset, and a service account key. Two IAM roles on that service account: BigQuery Data Viewer and Job User. It can read and it can run a query. It cannot write.
Bicycle on BigQuery
Know why your KPI moved before anyone asks.
Connect BigQuery read-only. Bicycle watches your revenue-critical KPIs, names the cause, and recommends the next step. First agent in about 15 minutes.
No credit card·Real product, real data
bicycle-analytics.checkout
cart_value_band
checkout_error_type
checkout_session
checkout_step
customer_segment
device_type
payment_attempt
payment_method
recovery
recovery_channel
WATCHINGCheckout Completion Rate by Funnel Step
On top of BigQuery AI
The KPI is in BigQuery. The cause often is not.
Checkout Completion Rate
DROP
It falls on mobile web. The cause is a payment provider's certificate rotation, and it lives in your gateway logs, not in your warehouse.
With BigQuery AI
It does notice, once you tell it what to watch
Agentic workflows run on a schedule and deliver briefings. You define the measures, choose the metrics, and write the directive that says what to investigate.
It does run the investigation
Deep-dive mode builds its own analytical plan. AI.KEY_DRIVERS names the exact segments behind the move, Apple Pay and mobile web, in one statement, with the SQL visible.
The cause sits outside the warehouse
The segment is where the drop landed, not what caused it. The certificate rotation is in a gateway log, the config push is in a deploy log. Reaching those, correlating them, and acting on what you find are systems you build.
With Bicycle
D
Detect
The drop surfaces on its own, ranked by revenue impact, with the segment and device already named. From ten unmodeled tables, with no measure authored first.
E
Explain
Cause analysis tests drivers inside BigQuery and signals outside it, ties the drop to the certificate rotation, and rules out a code deployment in the same window.
A
Act
Escalate to the payment provider, recommended to the owner who can approve it. Previewed, bounded, reversible, logged.
L
Learn
The signature is kept, so the next drop with this shape is recognized the moment it appears.
Bicycle runs on top of Gemini in BigQuery and your warehouse rather than replacing them. What it adds is the detection you did not have to schedule, cause analysis that reaches past the warehouse boundary, and a governed next step.
Compare in depth
Also compared: Dashboards · Build your own
What to expect
One connection. Four ways to get the answer.
Connect BigQuery read-only and the agent comes to life across the surfaces your team already lives in: alerts, data stories, chat, and dashboards. All from the same KPI and the same approved definitions. About 15 minutes after you connect.
Alerts
KPI movement, surfaced before anyone asks. Checkout completion falls on mobile web, and the alert names the surface, the size of the drop, and what caused it.
Data stories
The week or the quarter, as slides. What changed, why it changed, and what it cost, assembled into a deck you can present and share. Alerts cover today. Data stories cover the stretch.
Chat
Every follow-up, grounded. Ask why, ask where, ask what changed, and the answer comes back with its evidence attached.
Dashboards
Dashboards you did not have to build. Bicycle creates them from the KPIs it onboarded, and you can ask for another in plain language. They sit alongside the dashboards you already have.
Video placeholder
BigQuery connection walkthrough
Recording pending. This slot holds the real 16/9 frame.
A real BigQuery connection, start to finish: connect, pick an agent, watch the model assemble.
No integration project. Connect, and Bicycle does the rest.
Select the tables in scope
Pick only the tables your agent needs. Fewer tables, fewer queries, fewer bytes scanned.
Choose your first agent
Bicycle reads your schema and proposes the agents your data can already support. Checkout completion rate, payment decline analysis, cart recovery by channel, and more.
Bicycle builds the data model
Confirm three screens. Data model, KPIs, patterns, thresholds. About 10 to 15 minutes, and we email you when it's ready.
Tell it when your dataset loads. Set it to 2 AM and it analyzes a complete day rather than a half loaded table. That is the difference between a real alert and a false one.
What it built
Ten BigQuery tables in. A working data model out.
Bicycle classifies your tables into events and dimensions, maps each one to its source, and sets a refresh schedule per object. Every inference is visible, and every one is editable. You keep the guardrails.
Review events3 events
checkout_sessionbicycle-analytics.checkout.checkout_sessiondaily
payment_attemptbicycle-analytics.checkout.payment_attemptdaily
recoverybicycle-analytics.checkout.recoverydaily
Events: the business activities it tracks.
Review dimensions7 dimensions
cart_value_bandbicycle-analytics.checkout.cart_value_banddaily
checkout_error_typebicycle-analytics.checkout.checkout_error_typedaily
checkout_stepbicycle-analytics.checkout.checkout_stepdaily
customer_segmentbicycle-analytics.checkout.customer_segmentdaily
device_typebicycle-analytics.checkout.device_typedaily
payment_methodbicycle-analytics.checkout.payment_methoddaily
recovery_channelbicycle-analytics.checkout.recovery_channeldaily
Dimensions: how it narrows a problem down to the segment where it actually is.
This is the part that normally takes an analytics engineer weeks. It happens once, from a read-only connection.
Your warehouse, your rules.
Your data team sets the guardrails and the agent runs inside them, which keeps the recurring investigations off the analyst queue without giving up control of what gets trusted.
01
Read-only
Bicycle never writes to your project.
02
Queried in place
Your data is not copied out of BigQuery.
03
Scoped by you
You pick the tables. Nothing outside that scope is visible.
04
Fully editable
Every inferred mapping, KPI, and threshold is yours to change.
What you need before you connect.
What do I need to connect?
A project ID, a dataset, and a service account key with BigQuery Data Viewer and Job User.
Does Bicycle write to my project?
No.
What will this cost me in BigQuery spend?
Bicycle runs read queries on the schedule you set. On on-demand pricing you pay for bytes scanned, so a narrower table scope and a lower refresh frequency mean less spend. On a reservation it draws from your existing slots.
Can I add sources beyond BigQuery?
Yes. BigQuery for the KPI, and as many other sources as you like for cause analysis. That is usually where the cause turns out to be.
How is this different from Conversational Analytics in BigQuery?
BigQuery AI finds the segment a KPI moved in. Bicycle connects that movement to the thing that caused it, including systems outside BigQuery, and recommends a governed next step. It runs on top of BigQuery, not instead of it.
What if I'm on Snowflake?
Same flow, same four steps. See Bicycle on Snowflake.
How long until the first agent is live?
Usually 10–15 minutes.
Bring one KPI. Connect Bicycle to BigQuery.
Read-only, in place, live in about 15 minutes.
No credit card · Real product, real data
bicycle.ai/bigquery