The Business Problem: Refrigeration Failures Are Found, Not Predicted
Every refrigerated case in a retail store depends on a compressor. When it fails, it takes food safety and revenue with it. Most retailers only learn about a compressor problem when a store associate notices the temperature has already drifted, or when the unit stops entirely. By then the options are all bad: emergency dispatch, missing parts, potential product loss, and a manual scramble to find an available technician.
The underlying telemetry, vibration, pressure, and duty-cycle data, is often sitting in a data platform the maintenance team never looks at, disconnected from the tools that create tickets, dispatch technicians, and track parts. Even when a problem is visible in the data, nothing happens until a person notices it and manually opens a ticket.
This solution and reference architecture, designed for retail corporations, closes that gap. BigQuery scores every compressor's telemetry continuously and predicts failures up to 48 hours in advance. The moment a prediction crosses a threshold, ServiceNow takes over: a fully populated Priority 1 incident is created, an AI Agent triages it, reserves the part, creates a work order, and a technician is dispatched with everything needed for a first-time fix.
Why Google Cloud + ServiceNow
Google Cloud and ServiceNow solve two different halves of this problem. The value comes from connecting them directly rather than requiring manual handoffs and “swivel chairing” across both systems.
- BigQuery is where the signal lives. Its ML runtime continuously scores compressor telemetry for anomalies – a hard problem given the huge volume of data IoT sensors produce at scale -- and can surface a failure prediction well before a human would notice anything wrong.
- ServiceNow is where the response happens. Event Management, Field Service Management (FSM), and AI Agent Studio turn a raw prediction into a structured, assigned, resolved, and audited work order without manual re-entry.
- Workflow Data Fabric's Zero Copy Connection lets ServiceNow's AI Agents read live BigQuery telemetry directly, so the incident is investigated with the same data the model saw, not a stale export.
Solution Overview: Six Stages From Prediction to Verified Repair
The solution is delivered as six connected stages spanning both platforms:
|
Stage |
Platform |
What Happens |
|
1. Telemetry & Detection |
GCP (BigQuery) |
Compressor telemetry (temperature, vibration, pressure, duty cycle) streams into BigQuery. A trained ML autoencoder model scores incoming readings for anomalies. |
|
2. Trigger & Incident Creation |
GCP → ServiceNow |
When reconstruction error crosses threshold, a Cloud Run function publishes to Pub/Sub, and a second function calls the ServiceNow REST API, auto-creating a Priority 1 incident. |
|
3. AI Agent Triage |
ServiceNow |
An AI Agent automatically reviews the incident, creates child tasks (diagnostics, parts check, dispatch), and writes an AI-generated summary into the incident. |
|
4. Human-in-the-Loop Resolution |
ServiceNow |
A Facility Engineer runs a remote diagnostic check. The AI Agent reserves the part, creates the work order, and generates transfer order records; a dispatcher qualifies and assigns a technician. |
|
5. On-Site Repair |
ServiceNow (Mobile) |
The technician accepts the work order on the NOW mobile app, follows a knowledge article repair playbook, replaces the part, and closes the work order. |
|
6. Verified Closure & Learning |
ServiceNow → GCP |
Closing the work order notifies GCP, which re-checks telemetry during a monitoring window. A healthy signal closes the incident, and the outcome feeds back into the model for retraining. |
Reference Architecture
The diagram below shows how data and control flow across both platforms, from raw compressor telemetry through to technician dispatch and closure.
Reference architecture: GCP (telemetry, ML scoring, and prediction) and ServiceNow (incident management, AI agents, and field service), connected via REST API and the Workflow Data Fabric Zero Copy Connection.
Prerequisites
Before configuring this pattern, confirm that the following prerequisites are in place across Google Cloud and ServiceNow. These items provide the required foundation for the solution to connect telemetry, trigger incidents, run automated triage, dispatch technicians, and verify repair closure end to end.
On Google Cloud
- A BigQuery dataset with a telemetry table and a staging table for bulk loads
- A trained anomaly-detection model (this reference architecture uses an autoencoder – item #2 in the diagram above), be it a custom model or BigQuery’s own anomaly detection capabilities, and a query to score incoming telemetry
- A Pub/Sub topic and event trigger to publish detected anomalies
- Cloud Run functions to run the prediction query, call the ServiceNow REST API, and receive the closure signal
- A service account with BigQuery, Pub/Sub, and Cloud Run permissions
On ServiceNow
- Event Management, Alert Management, and Integration Hub, to ingest the Google Cloud anomaly event and turn it into a correlated Priority 1 incident.
- Field Service Management (FSM), for work orders, parts, and technician dispatch
- AI Agent Studio and the Now Assist Skill Kit, for automated triage and incident summaries
- Data Fabric Hub -- part of Workflow Data Fabric -- with a Zero Copy Connection to BigQuery
- CMDB configured for the compressor asset class, plus the Facility Maintenance, FSM Dispatcher, FSM Technician, and Facility Engineer groups
Walkthrough: From Predicted Failure to Verified Repair
The scenario below traces a single detection through the full lifecycle: a refrigeration compressor at a retail store, showing early signs of condenser coil inefficiency. The event was scored, escalated, triaged, resolved on-site, and verified closed end-to-end, illustrated with the actual solution screens.
Stage 1: Telemetry & Detection in BigQuery
Compressor telemetry (suction and discharge pressure, evaporator and condenser temperature, superheat, subcooling, compressor current, and oil differential pressure) streams into a BigQuery table. A Cloud Run function runs the scoring query against the trained autoencoder model, which flags a record as an anomaly once its reconstruction error, the mean squared error between predicted and actual sensor values, crosses 6.0. That cutoff lives directly in the scoring query, so a team can tighten or loosen detection sensitivity without retraining the model.
The Cloud Run function that runs the anomaly-scoring query against the trained BigQuery ML model.
Stage 2: Trigger & Incident Creation in ServiceNow
Once the threshold is crossed, a Cloud Run function publishes the anomaly to a Pub/Sub topic, which triggers a second function sends the anomaly payload into ServiceNow through Integration Hub, where the inbound flow normalizes the event and hands it to Event Management. ServiceNow receives a fully populated event, including the alert ID, node, metric, and data source.
The Event record created in ServiceNow from the BigQuery anomaly payload.
An Alert Management Rule immediately promotes that event into a Priority 1 incident, auto-assigned to the Facility Maintenance Group and, in this scenario, to NOC engineer Amelia Bryant. No form was opened manually, and no data was re-typed.
The auto-created Priority 1 incident, fully populated with store, asset, and telemetry context, and assigned to Amelia Bryant.
Stage 3: AI Agent Triage
Before Amelia opens the incident, an AI Agent has already reviewed it. Using Workflow Data Fabric’s Zero Copy Connection to BigQuery, the AI Agent reviews the same live telemetry that drove the prediction, creates five child tasks (telemetry and incident review, remote diagnostic check, nearest-store parts availability check, part reservation, and technician dispatch) and writes the live sensor readings into the incident.
The AI Agent-created incident tasks, along with the actual vs. setpoint sensor readings pulled from BigQuery.
From the incident's Overview tab, Amelia can also trigger a Now Assist summary. It condenses the alert, the affected store, and the likely root cause into a few readable lines, instead of asking her to parse a raw payload.
A Now Assist-generated incident summary, surfacing the issue and the key actions already taken.
Stage 4: Human-in-the-Loop Resolution
A Facility Engineer performs a remote diagnostic check against the checklist attached to the incident task. Once the issue is confirmed as something that needs a site visit, the engineer creates a draft work order. From there, an AI Agent takes over: it identifies the likely failed part from the telemetry signature, checks nearby inventory, and reserves it. Because that telemetry is available through Workflow Data Fabric, the agent can reason over external operational data without first copying it into ServiceNow.
The guided resolution playbook attached to the incident, walking the engineer through each stage of triage.
The work order created from the incident, in Draft state, linked back to the originating incident and asset.
The part requirement record, generated by Now Assist, showing the Emerson Condenser Coil sourced and delivered against the work order.
An FSM dispatcher reviews and qualifies the work order, and the AI Agent assigns and dispatches the technician, closing out the tasks completed so far.
Stage 5: On-Site Repair
The technician receives the work order on the NOW mobile app, with the store address, the specific asset, the part already reserved, and the option to accept or reject the job.
The technician's mobile view of the work order task, with asset, location, and reserved part already populated.
On site, the technician looks up the relevant knowledge article for the repair, follows the documented inspection, cleaning, and repair steps, and replaces the part.
The knowledge article repair steps the technician follows on the mobile app.
Once the repair is done, the technician marks the work order task complete directly from the mobile app.
Stage 6: Verified Closure & Learning
Work order closure is not the end of the story; it is a trigger. A ServiceNow business rule notifies GCP, which re-runs the anomaly-detection query against the freshest telemetry. This health check uses a stricter, more sensitive threshold (mean squared error above 3.2, versus 6.0 for the original detection) than the initial detection, so the incident only closes once the telemetry is convincingly stable, not just marginally below the original alarm line.
The Events list showing both the original Critical detection event and the follow-up OK event confirming the compressor is healthy.
The OK signal automatically closes the alert and the incident with resolution notes. The confirmed outcome, part replaced, resolution path, and final telemetry, is also written back to BigQuery, so the model keeps learning from real repairs instead of only from synthetic training data.
Business Outcomes
- Up to 48 hours of advance warning before a compressor actually fails, instead of finding out after a store loses cooling.
- Incidents arrive fully populated and pre-assigned, with zero manual data entry between detection and dispatch.
- Workflow Data Fabric gives agents and operators live access to BigQuery telemetry in the flow of work, avoiding stale exports while keeping Google Cloud as the system of record for sensor data.
- An AI Agent triages, summarizes, and reserves parts before a human ever opens the incident, so the engineer starts informed, not from a blank record.
- Technicians arrive with the right part already reserved and a step-by-step repair playbook, supporting a first-time fix.
- Closure is verified by a second, stricter telemetry check, not by self-reported completion, and every confirmed outcome retrains the model.
Conclusion
BigQuery predicts. ServiceNow responds. By scoring compressor telemetry continuously and automating the incident-to-repair lifecycle, this solution turns refrigeration maintenance from reactive firefighting into a scheduled, low-drama fix, days before a failure would have taken the case offline.
Interested in learning more? If your team is interested in combining the power of ServiceNow with our partners' solutions, please fill out this form and we will follow up with more information.
